Context has its own sequence length

This reverts commit 1cd09cccf3.

.github/ISSUE_TEMPLATE/config.yml

@@ -1,7 +1,4 @@
 contact_links:
-  - name: Forum
-    url: https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63
-    about: General usage questions and community discussions
   - name: Blank issue
     url: https://github.com/huggingface/diffusers/issues/new
-    about: Please note that the Forum is in most places the right place for discussions
+    about: General usage questions and community discussions

.github/actions/setup-miniconda/action.yml

@@ -0,0 +1,146 @@
+name: Set up conda environment for testing
+
+description: Sets up miniconda in your ${RUNNER_TEMP} environment and gives you the ${CONDA_RUN} environment variable so you don't have to worry about polluting non-empeheral runners anymore
+
+inputs:
+  python-version:
+    description: If set to any value, dont use sudo to clean the workspace
+    required: false
+    type: string
+    default: "3.9"
+  miniconda-version:
+    description: Miniconda version to install
+    required: false
+    type: string
+    default: "4.12.0"
+  environment-file:
+    description: Environment file to install dependencies from
+    required: false
+    type: string
+    default: ""
+
+runs:
+  using: composite
+  steps:
+      # Use the same trick from https://github.com/marketplace/actions/setup-miniconda
+      # to refresh the cache daily. This is kind of optional though
+      - name: Get date
+        id: get-date
+        shell: bash
+        run: echo "::set-output name=today::$(/bin/date -u '+%Y%m%d')d"
+      - name: Setup miniconda cache
+        id: miniconda-cache
+        uses: actions/cache@v2
+        with:
+          path: ${{ runner.temp }}/miniconda
+          key: miniconda-${{ runner.os }}-${{ runner.arch }}-${{ inputs.python-version }}-${{ steps.get-date.outputs.today }}
+      - name: Install miniconda (${{ inputs.miniconda-version }})
+        if: steps.miniconda-cache.outputs.cache-hit != 'true'
+        env:
+          MINICONDA_VERSION: ${{ inputs.miniconda-version }}
+        shell: bash -l {0}
+        run: |
+          MINICONDA_INSTALL_PATH="${RUNNER_TEMP}/miniconda"
+          mkdir -p "${MINICONDA_INSTALL_PATH}"
+          case ${RUNNER_OS}-${RUNNER_ARCH} in
+            Linux-X64)
+              MINICONDA_ARCH="Linux-x86_64"
+              ;;
+            macOS-ARM64)
+              MINICONDA_ARCH="MacOSX-arm64"
+              ;;
+            macOS-X64)
+              MINICONDA_ARCH="MacOSX-x86_64"
+              ;;
+            *)
+            echo "::error::Platform ${RUNNER_OS}-${RUNNER_ARCH} currently unsupported using this action"
+              exit 1
+              ;;
+          esac
+          MINICONDA_URL="https://repo.anaconda.com/miniconda/Miniconda3-py39_${MINICONDA_VERSION}-${MINICONDA_ARCH}.sh"
+          curl -fsSL "${MINICONDA_URL}" -o "${MINICONDA_INSTALL_PATH}/miniconda.sh"
+          bash "${MINICONDA_INSTALL_PATH}/miniconda.sh" -b -u -p "${MINICONDA_INSTALL_PATH}"
+          rm -rf "${MINICONDA_INSTALL_PATH}/miniconda.sh"
+      - name: Update GitHub path to include miniconda install
+        shell: bash
+        run: |
+          MINICONDA_INSTALL_PATH="${RUNNER_TEMP}/miniconda"
+          echo "${MINICONDA_INSTALL_PATH}/bin" >> $GITHUB_PATH
+      - name: Setup miniconda env cache (with env file)
+        id: miniconda-env-cache-env-file
+        if: ${{ runner.os }} == 'macOS' && ${{ inputs.environment-file }} != ''
+        uses: actions/cache@v2
+        with:
+          path: ${{ runner.temp }}/conda-python-${{ inputs.python-version }}
+          key: miniconda-env-${{ runner.os }}-${{ runner.arch }}-${{ inputs.python-version }}-${{ steps.get-date.outputs.today }}-${{ hashFiles(inputs.environment-file) }}
+      - name: Setup miniconda env cache (without env file)
+        id: miniconda-env-cache
+        if: ${{ runner.os }} == 'macOS' && ${{ inputs.environment-file }} == ''
+        uses: actions/cache@v2
+        with:
+          path: ${{ runner.temp }}/conda-python-${{ inputs.python-version }}
+          key: miniconda-env-${{ runner.os }}-${{ runner.arch }}-${{ inputs.python-version }}-${{ steps.get-date.outputs.today }}
+      - name: Setup conda environment with python (v${{ inputs.python-version }})
+        if: steps.miniconda-env-cache-env-file.outputs.cache-hit != 'true' && steps.miniconda-env-cache.outputs.cache-hit != 'true'
+        shell: bash
+        env:
+          PYTHON_VERSION: ${{ inputs.python-version }}
+          ENV_FILE: ${{ inputs.environment-file }}
+        run: |
+          CONDA_BASE_ENV="${RUNNER_TEMP}/conda-python-${PYTHON_VERSION}"
+          ENV_FILE_FLAG=""
+          if [[ -f "${ENV_FILE}" ]]; then
+            ENV_FILE_FLAG="--file ${ENV_FILE}"
+          elif [[ -n "${ENV_FILE}" ]]; then
+            echo "::warning::Specified env file (${ENV_FILE}) not found, not going to include it"
+          fi
+          conda create \
+            --yes \
+            --prefix "${CONDA_BASE_ENV}" \
+            "python=${PYTHON_VERSION}" \
+            ${ENV_FILE_FLAG} \
+            cmake=3.22 \
+            conda-build=3.21 \
+            ninja=1.10 \
+            pkg-config=0.29 \
+            wheel=0.37
+      - name: Clone the base conda environment and update GitHub env
+        shell: bash
+        env:
+          PYTHON_VERSION: ${{ inputs.python-version }}
+          CONDA_BASE_ENV: ${{ runner.temp }}/conda-python-${{ inputs.python-version }}
+        run: |
+          CONDA_ENV="${RUNNER_TEMP}/conda_environment_${GITHUB_RUN_ID}"
+          conda create \
+            --yes \
+            --prefix "${CONDA_ENV}" \
+            --clone "${CONDA_BASE_ENV}"
+          # TODO: conda-build could not be cloned because it hardcodes the path, so it
+          # could not be cached
+          conda install --yes -p ${CONDA_ENV} conda-build=3.21
+          echo "CONDA_ENV=${CONDA_ENV}" >> "${GITHUB_ENV}"
+          echo "CONDA_RUN=conda run -p ${CONDA_ENV} --no-capture-output" >> "${GITHUB_ENV}"
+          echo "CONDA_BUILD=conda run -p ${CONDA_ENV} conda-build" >> "${GITHUB_ENV}"
+          echo "CONDA_INSTALL=conda install -p ${CONDA_ENV}" >> "${GITHUB_ENV}"
+      - name: Get disk space usage and throw an error for low disk space
+        shell: bash
+        run: |
+          echo "Print the available disk space for manual inspection"
+          df -h
+          # Set the minimum requirement space to 4GB
+          MINIMUM_AVAILABLE_SPACE_IN_GB=4
+          MINIMUM_AVAILABLE_SPACE_IN_KB=$(($MINIMUM_AVAILABLE_SPACE_IN_GB * 1024 * 1024))
+          # Use KB to avoid floating point warning like 3.1GB
+          df -k | tr -s ' ' | cut -d' ' -f 4,9 | while read -r LINE;
+          do
+            AVAIL=$(echo $LINE | cut -f1 -d' ')
+            MOUNT=$(echo $LINE | cut -f2 -d' ')
+            if [ "$MOUNT" = "/" ]; then
+              if [ "$AVAIL" -lt "$MINIMUM_AVAILABLE_SPACE_IN_KB" ]; then
+                echo "There is only ${AVAIL}KB free space left in $MOUNT, which is less than the minimum requirement of ${MINIMUM_AVAILABLE_SPACE_IN_KB}KB. Please help create an issue to PyTorch Release Engineering via https://github.com/pytorch/test-infra/issues and provide the link to the workflow run."
+                exit 1;
+              else
+                echo "There is ${AVAIL}KB free space left in $MOUNT, continue"
+              fi
+            fi
+          done

.github/workflows/build_docker_images.yml

@@ -0,0 +1,50 @@
+name: Build Docker images (nightly)
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 0 * * *" # every day at midnight
+
+concurrency:
+  group: docker-image-builds
+  cancel-in-progress: false
+
+env:
+  REGISTRY: diffusers
+
+jobs:
+  build-docker-images:
+    runs-on: ubuntu-latest
+
+    permissions:
+      contents: read
+      packages: write
+
+    strategy:
+      fail-fast: false
+      matrix:
+        image-name:
+          - diffusers-pytorch-cpu
+          - diffusers-pytorch-cuda
+          - diffusers-flax-cpu
+          - diffusers-flax-tpu
+          - diffusers-onnxruntime-cpu
+          - diffusers-onnxruntime-cuda
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v3
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ env.REGISTRY }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Build and push
+        uses: docker/build-push-action@v3
+        with:
+          no-cache: true
+          context: ./docker/${{ matrix.image-name }}
+          push: true
+          tags: ${{ env.REGISTRY }}/${{ matrix.image-name }}:latest

.github/workflows/pr_quality.yml

@@ -31,3 +31,20 @@ jobs:
           isort --check-only examples tests src utils scripts
           flake8 examples tests src utils scripts
           doc-builder style src/diffusers docs/source --max_len 119 --check_only --path_to_docs docs/source
+
+  check_repository_consistency:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.7"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: |
+          python utils/check_copies.py
+          python utils/check_dummies.py

.github/workflows/pr_tests.yml

@@ -1,4 +1,4 @@
-name: Run non-slow tests
+name: Run fast tests
 
 on:
   pull_request:
@@ -10,19 +10,46 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  HF_HOME: /mnt/cache
-  OMP_NUM_THREADS: 8
-  MKL_NUM_THREADS: 8
+  DIFFUSERS_IS_CI: yes
+  OMP_NUM_THREADS: 4
+  MKL_NUM_THREADS: 4
   PYTEST_TIMEOUT: 60
+  MPS_TORCH_VERSION: 1.13.0
 
 jobs:
-  run_tests_cpu:
-    name: Diffusers tests
-    runs-on: [ self-hosted, docker-gpu ]
+  run_fast_tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        config:
+          - name: Fast PyTorch CPU tests on Ubuntu
+            framework: pytorch
+            runner: docker-cpu
+            image: diffusers/diffusers-pytorch-cpu
+            report: torch_cpu
+          - name: Fast Flax CPU tests on Ubuntu
+            framework: flax
+            runner: docker-cpu
+            image: diffusers/diffusers-flax-cpu
+            report: flax_cpu
+          - name: Fast ONNXRuntime CPU tests on Ubuntu
+            framework: onnxruntime
+            runner: docker-cpu
+            image: diffusers/diffusers-onnxruntime-cpu
+            report: onnx_cpu
+
+    name: ${{ matrix.config.name }}
+
+    runs-on: ${{ matrix.config.runner }}
+
     container:
-      image: python:3.7
+      image: ${{ matrix.config.image }}
       options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
 
+    defaults:
+      run:
+        shell: bash
+
     steps:
     - name: Checkout diffusers
       uses: actions/checkout@v3
@@ -31,25 +58,93 @@ jobs:
 
     - name: Install dependencies
       run: |
-        python -m pip install --upgrade pip
-        python -m pip install torch --extra-index-url https://download.pytorch.org/whl/cpu
         python -m pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate
 
     - name: Environment
       run: |
         python utils/print_env.py
 
-    - name: Run all non-slow selected tests on CPU
+    - name: Run fast PyTorch CPU tests
+      if: ${{ matrix.config.framework == 'pytorch' }}
       run: |
-        python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=tests_torch_cpu tests/
+        python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "not Flax and not Onnx" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
+
+    - name: Run fast Flax TPU tests
+      if: ${{ matrix.config.framework == 'flax' }}
+      run: |
+        python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "Flax" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
+
+    - name: Run fast ONNXRuntime CPU tests
+      if: ${{ matrix.config.framework == 'onnxruntime' }}
+      run: |
+        python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "Onnx" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
 
     - name: Failure short reports
       if: ${{ failure() }}
-      run: cat reports/tests_torch_cpu_failures_short.txt
+      run: cat reports/tests_${{ matrix.config.report }}_failures_short.txt
 
     - name: Test suite reports artifacts
       if: ${{ always() }}
       uses: actions/upload-artifact@v2
       with:
-        name: pr_torch_test_reports
+        name: pr_${{ matrix.config.report }}_test_reports
+        path: reports
+
+  run_fast_tests_apple_m1:
+    name: Fast PyTorch MPS tests on MacOS
+    runs-on: [ self-hosted, apple-m1 ]
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Clean checkout
+      shell: arch -arch arm64 bash {0}
+      run: |
+        git clean -fxd
+
+    - name: Setup miniconda
+      uses: ./.github/actions/setup-miniconda
+      with:
+        python-version: 3.9
+
+    - name: Install dependencies
+      shell: arch -arch arm64 bash {0}
+      run: |
+        ${CONDA_RUN} python -m pip install --upgrade pip
+        ${CONDA_RUN} python -m pip install -e .[quality,test]
+        ${CONDA_RUN} python -m pip install --pre torch==${MPS_TORCH_VERSION} --extra-index-url https://download.pytorch.org/whl/test/cpu
+        ${CONDA_RUN} python -m pip install git+https://github.com/huggingface/accelerate
+
+    - name: Environment
+      shell: arch -arch arm64 bash {0}
+      run: |
+        ${CONDA_RUN} python utils/print_env.py
+
+    - name: Run fast PyTorch tests on M1 (MPS)
+      shell: arch -arch arm64 bash {0}
+      run: |
+        ${CONDA_RUN} python -m pytest -n 0 -s -v --make-reports=tests_torch_mps tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: cat reports/tests_torch_mps_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: pr_torch_mps_test_reports
         path: reports

.github/workflows/push_tests.yml

@@ -6,6 +6,7 @@ on:
       - main
 
 env:
+  DIFFUSERS_IS_CI: yes
   HF_HOME: /mnt/cache
   OMP_NUM_THREADS: 8
   MKL_NUM_THREADS: 8
@@ -13,12 +14,38 @@ env:
   RUN_SLOW: yes
 
 jobs:
-  run_tests_single_gpu:
-    name: Diffusers tests
-    runs-on: [ self-hosted, docker-gpu, single-gpu ]
+  run_slow_tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        config:
+          - name: Slow PyTorch CUDA tests on Ubuntu
+            framework: pytorch
+            runner: docker-gpu
+            image: diffusers/diffusers-pytorch-cuda
+            report: torch_cuda
+          - name: Slow Flax TPU tests on Ubuntu
+            framework: flax
+            runner: docker-tpu
+            image: diffusers/diffusers-flax-tpu
+            report: flax_tpu
+          - name: Slow ONNXRuntime CUDA tests on Ubuntu
+            framework: onnxruntime
+            runner: docker-gpu
+            image: diffusers/diffusers-onnxruntime-cuda
+            report: onnx_cuda
+
+    name: ${{ matrix.config.name }}
+
+    runs-on: ${{ matrix.config.runner }}
+
     container:
-      image: nvcr.io/nvidia/pytorch:22.07-py3
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache
+      image: ${{ matrix.config.image }}
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ ${{ matrix.config.runner == 'docker-tpu' && '--privileged' || '--gpus 0'}}
+
+    defaults:
+      run:
+        shell: bash
 
     steps:
     - name: Checkout diffusers
@@ -27,45 +54,68 @@ jobs:
         fetch-depth: 2
 
     - name: NVIDIA-SMI
+      if : ${{ matrix.config.runner == 'docker-gpu' }}
       run: |
         nvidia-smi
 
     - name: Install dependencies
       run: |
-        python -m pip install --upgrade pip
-        python -m pip uninstall -y torch torchvision torchtext
-        python -m pip install torch --extra-index-url https://download.pytorch.org/whl/cu116
         python -m pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate
 
     - name: Environment
       run: |
         python utils/print_env.py
 
-    - name: Run all (incl. slow) tests on GPU
+    - name: Run slow PyTorch CUDA tests
+      if: ${{ matrix.config.framework == 'pytorch' }}
       env:
         HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
       run: |
-        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=tests_torch_gpu tests/
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "not Flax and not Onnx" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
+
+    - name: Run slow Flax TPU tests
+      if: ${{ matrix.config.framework == 'flax' }}
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      run: |
+        python -m pytest -n 0 \
+          -s -v -k "Flax" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
+
+    - name: Run slow ONNXRuntime CUDA tests
+      if: ${{ matrix.config.framework == 'onnxruntime' }}
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "Onnx" \
+          --make-reports=tests_${{ matrix.config.report }} \
+          tests/
 
     - name: Failure short reports
       if: ${{ failure() }}
-      run: cat reports/tests_torch_gpu_failures_short.txt
+      run: cat reports/tests_${{ matrix.config.report }}_failures_short.txt
 
     - name: Test suite reports artifacts
       if: ${{ always() }}
       uses: actions/upload-artifact@v2
       with:
-        name: torch_test_reports
+        name: ${{ matrix.config.report }}_test_reports
         path: reports
 
+  run_examples_tests:
+    name: Examples PyTorch CUDA tests on Ubuntu
 
+    runs-on: docker-gpu
 
-  run_examples_single_gpu:
-    name: Examples tests
-    runs-on: [ self-hosted, docker-gpu, single-gpu ]
     container:
-      image: nvcr.io/nvidia/pytorch:22.07-py3
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache
+      image: diffusers/diffusers-pytorch-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
 
     steps:
     - name: Checkout diffusers
@@ -79,10 +129,8 @@ jobs:
 
     - name: Install dependencies
       run: |
-        python -m pip install --upgrade pip
-        python -m pip uninstall -y torch torchvision torchtext
-        python -m pip install torch --extra-index-url https://download.pytorch.org/whl/cu116
         python -m pip install -e .[quality,test,training]
+        python -m pip install git+https://github.com/huggingface/accelerate
 
     - name: Environment
       run: |
@@ -92,11 +140,11 @@ jobs:
       env:
         HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
       run: |
-        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_gpu examples/
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
 
     - name: Failure short reports
       if: ${{ failure() }}
-      run: cat reports/examples_torch_gpu_failures_short.txt
+      run: cat reports/examples_torch_cuda_failures_short.txt
 
     - name: Test suite reports artifacts
       if: ${{ always() }}

.gitignore

@@ -163,4 +163,6 @@ tags
 *.lock
 
 # DS_Store (MacOS)
-.DS_Store
+.DS_Store
+# RL pipelines may produce mp4 outputs
+*.mp4

MANIFEST.in

@@ -1 +1,2 @@
+include LICENSE
 include src/diffusers/utils/model_card_template.md

Makefile

@@ -67,6 +67,7 @@ fixup: modified_only_fixup extra_style_checks autogenerate_code repo-consistency
 # Make marked copies of snippets of codes conform to the original
 
 fix-copies:
+	python utils/check_copies.py --fix_and_overwrite
 	python utils/check_dummies.py --fix_and_overwrite
 
 # Run tests for the library

README.md

@@ -1,6 +1,6 @@
 <p align="center">
     <br>
-    <img src="docs/source/imgs/diffusers_library.jpg" width="400"/>
+    <img src="https://github.com/huggingface/diffusers/raw/main/docs/source/imgs/diffusers_library.jpg" width="400"/>
     <br>
 <p>
 <p align="center">
@@ -27,10 +27,12 @@ More precisely, 🤗 Diffusers offers:
 
 ## Installation
 
+### For PyTorch
+
 **With `pip`**
     
 ```bash
-pip install --upgrade diffusers
+pip install --upgrade diffusers[torch]
 ```
 
 **With `conda`**
@@ -39,6 +41,14 @@ pip install --upgrade diffusers
 conda install -c conda-forge diffusers
 ```
 
+### For Flax
+
+**With `pip`**
+
+```bash
+pip install --upgrade diffusers[flax]
+```
+
 **Apple Silicon (M1/M2) support**
 
 Please, refer to [the documentation](https://huggingface.co/docs/diffusers/optimization/mps).
@@ -64,44 +74,55 @@ In order to get started, we recommend taking a look at two notebooks:
 - The [Training a diffusers model](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) notebook summarizes diffusion models training methods. This notebook takes a step-by-step approach to training your
   diffusion models on an image dataset, with explanatory graphics. 
   
-## **New** Stable Diffusion is now fully compatible with `diffusers`!  
+## Stable Diffusion is fully compatible with `diffusers`!  
 
-Stable Diffusion is a text-to-image latent diffusion model created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/) and [LAION](https://laion.ai/). It's trained on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database. This model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts. With its 860M UNet and 123M text encoder, the model is relatively lightweight and runs on a GPU with at least 10GB VRAM.
+Stable Diffusion is a text-to-image latent diffusion model created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/), [LAION](https://laion.ai/) and [RunwayML](https://runwayml.com/). It's trained on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database. This model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts. With its 860M UNet and 123M text encoder, the model is relatively lightweight and runs on a GPU with at least 4GB VRAM.
 See the [model card](https://huggingface.co/CompVis/stable-diffusion) for more information.
 
-You need to accept the model license before downloading or using the Stable Diffusion weights. Please, visit the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section](https://huggingface.co/docs/hub/security-tokens) of the documentation.
+You need to accept the model license before downloading or using the Stable Diffusion weights. Please, visit the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license carefully and tick the checkbox if you agree. You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section](https://huggingface.co/docs/hub/security-tokens) of the documentation.
 
 
 ### Text-to-Image generation with Stable Diffusion
 
+First let's install
+```bash
+pip install --upgrade diffusers transformers scipy
+```
+
+Run this command to log in with your HF Hub token if you haven't before (you can skip this step if you prefer to run the model locally, follow [this](#running-the-model-locally) instead)
+```bash
+huggingface-cli login
+```
+
 We recommend using the model in [half-precision (`fp16`)](https://pytorch.org/blog/accelerating-training-on-nvidia-gpus-with-pytorch-automatic-mixed-precision/) as it gives almost always the same results as full
 precision while being roughly twice as fast and requiring half the amount of GPU RAM.
 
 ```python
-# make sure you're logged in with `huggingface-cli login`
+import torch
 from diffusers import StableDiffusionPipeline
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_type=torch.float16, revision="fp16")
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, revision="fp16")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
 image = pipe(prompt).images[0]  
 ```
 
-**Note**: If you don't want to use the token, you can also simply download the model weights
-(after having [accepted the license](https://huggingface.co/CompVis/stable-diffusion-v1-4)) and pass
+#### Running the model locally
+If you don't want to login to Hugging Face, you can also simply download the model folder
+(after having [accepted the license](https://huggingface.co/runwayml/stable-diffusion-v1-5)) and pass
 the path to the local folder to the `StableDiffusionPipeline`.
 
 ```
 git lfs install
-git clone https://huggingface.co/CompVis/stable-diffusion-v1-4
+git clone https://huggingface.co/runwayml/stable-diffusion-v1-5
 ```
 
-Assuming the folder is stored locally under `./stable-diffusion-v1-4`, you can also run stable diffusion
+Assuming the folder is stored locally under `./stable-diffusion-v1-5`, you can also run stable diffusion
 without requiring an authentication token:
 
 ```python
-pipe = StableDiffusionPipeline.from_pretrained("./stable-diffusion-v1-4")
+pipe = StableDiffusionPipeline.from_pretrained("./stable-diffusion-v1-5")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
@@ -114,7 +135,7 @@ The following snippet should result in less than 4GB VRAM.
 
 ```python
 pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", 
+    "runwayml/stable-diffusion-v1-5", 
     revision="fp16", 
     torch_dtype=torch.float16,
 )
@@ -125,25 +146,13 @@ pipe.enable_attention_slicing()
 image = pipe(prompt).images[0]  
 ```
 
-If you wish to use a different scheduler, you can simply instantiate
+If you wish to use a different scheduler (e.g.: DDIM, LMS, PNDM/PLMS), you can instantiate
 it before the pipeline and pass it to `from_pretrained`.
     
 ```python
 from diffusers import LMSDiscreteScheduler
 
-lms = LMSDiscreteScheduler(
-    beta_start=0.00085, 
-    beta_end=0.012, 
-    beta_schedule="scaled_linear"
-)
-
-pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", 
-    revision="fp16", 
-    torch_dtype=torch.float16,
-    scheduler=lms,
-)
-pipe = pipe.to("cuda")
+pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
 
 prompt = "a photo of an astronaut riding a horse on mars"
 image = pipe(prompt).images[0]  
@@ -158,7 +167,7 @@ please run the model in the default *full-precision* setting:
 # make sure you're logged in with `huggingface-cli login`
 from diffusers import StableDiffusionPipeline
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 
 # disable the following line if you run on CPU
 pipe = pipe.to("cuda")
@@ -169,6 +178,75 @@ image = pipe(prompt).images[0]
 image.save("astronaut_rides_horse.png")
 ```
 
+### JAX/Flax
+
+Diffusers offers a JAX / Flax implementation of Stable Diffusion for very fast inference. JAX shines specially on TPU hardware because each TPU server has 8 accelerators working in parallel, but it runs great on GPUs too.
+
+Running the pipeline with the default PNDMScheduler:
+
+```python
+import jax
+import numpy as np
+from flax.jax_utils import replicate
+from flax.training.common_utils import shard
+
+from diffusers import FlaxStableDiffusionPipeline
+
+pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5", revision="flax", dtype=jax.numpy.bfloat16
+)
+
+prompt = "a photo of an astronaut riding a horse on mars"
+
+prng_seed = jax.random.PRNGKey(0)
+num_inference_steps = 50
+
+num_samples = jax.device_count()
+prompt = num_samples * [prompt]
+prompt_ids = pipeline.prepare_inputs(prompt)
+
+# shard inputs and rng
+params = replicate(params)
+prng_seed = jax.random.split(prng_seed, jax.device_count())
+prompt_ids = shard(prompt_ids)
+
+images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
+images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
+```
+
+**Note**:
+If you are limited by TPU memory, please make sure to load the `FlaxStableDiffusionPipeline` in `bfloat16` precision instead of the default `float32` precision as done above. You can do so by telling diffusers to load the weights from "bf16" branch.
+
+```python
+import jax
+import numpy as np
+from flax.jax_utils import replicate
+from flax.training.common_utils import shard
+
+from diffusers import FlaxStableDiffusionPipeline
+
+pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5", revision="bf16", dtype=jax.numpy.bfloat16
+)
+
+prompt = "a photo of an astronaut riding a horse on mars"
+
+prng_seed = jax.random.PRNGKey(0)
+num_inference_steps = 50
+
+num_samples = jax.device_count()
+prompt = num_samples * [prompt]
+prompt_ids = pipeline.prepare_inputs(prompt)
+
+# shard inputs and rng
+params = replicate(params)
+prng_seed = jax.random.split(prng_seed, jax.device_count())
+prompt_ids = shard(prompt_ids)
+
+images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
+images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
+```
+
 ### Image-to-Image text-guided generation with Stable Diffusion
 
 The `StableDiffusionImg2ImgPipeline` lets you pass a text prompt and an initial image to condition the generation of new images.
@@ -183,14 +261,14 @@ from diffusers import StableDiffusionImg2ImgPipeline
 
 # load the pipeline
 device = "cuda"
-model_id_or_path = "CompVis/stable-diffusion-v1-4"
+model_id_or_path = "runwayml/stable-diffusion-v1-5"
 pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
     model_id_or_path,
     revision="fp16", 
     torch_dtype=torch.float16,
 )
-# or download via git clone https://huggingface.co/CompVis/stable-diffusion-v1-4
-# and pass `model_id_or_path="./stable-diffusion-v1-4"`.
+# or download via git clone https://huggingface.co/runwayml/stable-diffusion-v1-5
+# and pass `model_id_or_path="./stable-diffusion-v1-5"`.
 pipe = pipe.to(device)
 
 # let's download an initial image
@@ -210,14 +288,16 @@ You can also run this example on colab [![Open In Colab](https://colab.research.
 
 ### In-painting using Stable Diffusion
 
-The `StableDiffusionInpaintPipeline` lets you edit specific parts of an image by providing a mask and text prompt.
+The `StableDiffusionInpaintPipeline` lets you edit specific parts of an image by providing a mask and a text prompt. It uses a model optimized for this particular task, whose license you need to accept before use.
+
+Please, visit the [model card](https://huggingface.co/runwayml/stable-diffusion-inpainting), read the license carefully and tick the checkbox if you agree. Note that this is an additional license, you need to accept it even if you accepted the text-to-image Stable Diffusion license in the past. You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section](https://huggingface.co/docs/hub/security-tokens) of the documentation.
+
 
 ```python
-from io import BytesIO
-
-import torch
-import requests
 import PIL
+import requests
+import torch
+from io import BytesIO
 
 from diffusers import StableDiffusionInpaintPipeline
 
@@ -231,21 +311,15 @@ mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data
 init_image = download_image(img_url).resize((512, 512))
 mask_image = download_image(mask_url).resize((512, 512))
 
-device = "cuda"
-model_id_or_path = "CompVis/stable-diffusion-v1-4"
 pipe = StableDiffusionInpaintPipeline.from_pretrained(
-    model_id_or_path,
-    revision="fp16", 
+    "runwayml/stable-diffusion-inpainting",
+    revision="fp16",
     torch_dtype=torch.float16,
 )
-# or download via git clone https://huggingface.co/CompVis/stable-diffusion-v1-4
-# and pass `model_id_or_path="./stable-diffusion-v1-4"`.
-pipe = pipe.to(device)
+pipe = pipe.to("cuda")
 
-prompt = "a cat sitting on a bench"
-images = pipe(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images
-
-images[0].save("cat_on_bench.png")
+prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
 ```
 
 ### Tweak prompts reusing seeds and latents
@@ -255,8 +329,26 @@ You can generate your own latents to reproduce results, or tweak your prompt on
 
 For more details, check out [the Stable Diffusion notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_diffusion.ipynb) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_diffusion.ipynb)
 and have a look into the [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.2.0).
-  
-## Examples
+
+## Fine-Tuning Stable Diffusion
+
+Fine-tuning techniques make it possible to adapt Stable Diffusion to your own dataset, or add new subjects to it. These are some of the techniques supported in `diffusers`:
+
+Textual Inversion is a technique for capturing novel concepts from a small number of example images in a way that can later be used to control text-to-image pipelines. It does so by learning new 'words' in the embedding space of the pipeline's text encoder. These special words can then be used within text prompts to achieve very fine-grained control of the resulting images. 
+
+- Textual Inversion. Capture novel concepts from a small set of sample images, and associate them with new "words" in the embedding space of the text encoder. Please, refer to [our training examples](https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion) or [documentation](https://huggingface.co/docs/diffusers/training/text_inversion) to try for yourself.
+
+- Dreambooth. Another technique to capture new concepts in Stable Diffusion. This method fine-tunes the UNet (and, optionally, also the text encoder) of the pipeline to achieve impressive results. Please, refer to [our training example](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth) and [training report](https://huggingface.co/blog/dreambooth) for additional details and training recommendations.
+
+- Full Stable Diffusion fine-tuning. If you have a more sizable dataset with a specific look or style, you can fine-tune Stable Diffusion so that it outputs images following those examples. This was the approach taken to create [a Pokémon Stable Diffusion model](https://huggingface.co/justinpinkney/pokemon-stable-diffusion) (by Justing Pinkney / Lambda Labs), [a Japanese specific version of Stable Diffusion](https://huggingface.co/spaces/rinna/japanese-stable-diffusion) (by [Rinna Co.](https://github.com/rinnakk/japanese-stable-diffusion/) and others. You can start at [our text-to-image fine-tuning example](https://github.com/huggingface/diffusers/tree/main/examples/text_to_image) and go from there.
+
+
+## Stable Diffusion Community Pipelines
+
+The release of Stable Diffusion as an open source model has fostered a lot of interesting ideas and experimentation. 
+Our [Community Examples folder](https://github.com/huggingface/diffusers/tree/main/examples/community) contains many ideas worth exploring, like interpolating to create animated videos, using CLIP Guidance for additional prompt fidelity, term weighting, and much more! [Take a look](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview) and [contribute your own](https://huggingface.co/docs/diffusers/using-diffusers/contribute_pipeline).
+
+## Other Examples
 
 There are many ways to try running Diffusers! Here we outline code-focused tools (primarily using `DiffusionPipeline`s and Google Colab) and interactive web-tools.
 
@@ -265,7 +357,7 @@ There are many ways to try running Diffusers! Here we outline code-focused tools
 If you want to run the code yourself 💻, you can try out:
 - [Text-to-Image Latent Diffusion](https://huggingface.co/CompVis/ldm-text2im-large-256)
 ```python
-# !pip install diffusers transformers
+# !pip install diffusers["torch"] transformers
 from diffusers import DiffusionPipeline
 
 device = "cuda"
@@ -284,7 +376,7 @@ image.save("squirrel.png")
 ```
 - [Unconditional Diffusion with discrete scheduler](https://huggingface.co/google/ddpm-celebahq-256)
 ```python
-# !pip install diffusers
+# !pip install diffusers["torch"]
 from diffusers import DDPMPipeline, DDIMPipeline, PNDMPipeline
 
 model_id = "google/ddpm-celebahq-256"
@@ -303,10 +395,14 @@ image.save("ddpm_generated_image.png")
 - [Unconditional Latent Diffusion](https://huggingface.co/CompVis/ldm-celebahq-256)
 - [Unconditional Diffusion with continuous scheduler](https://huggingface.co/google/ncsnpp-ffhq-1024)
 
-**Other Notebooks**:
+**Other Image Notebooks**:
 * [image-to-image generation with Stable Diffusion](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb) ![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg),
 * [tweak images via repeated Stable Diffusion seeds](https://colab.research.google.com/github/pcuenca/diffusers-examples/blob/main/notebooks/stable-diffusion-seeds.ipynb) ![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg),
 
+**Diffusers for Other Modalities**:
+* [Molecule conformation generation](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/geodiff_molecule_conformation.ipynb) ![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg),
+* [Model-based reinforcement learning](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/reinforcement_learning_with_diffusers.ipynb) ![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg),
+
 ### Web Demos
 If you just want to play around with some web demos, you can try out the following 🚀 Spaces:
 | Model                          	| Hugging Face Spaces                                                                                                                                               	|
@@ -329,7 +425,7 @@ If you just want to play around with some web demos, you can try out the followi
 <p>
     
 **Schedulers**: Algorithm class for both **inference** and **training**.
-The class provides functionality to compute previous image according to alpha, beta schedule as well as predict noise for training.
+The class provides functionality to compute previous image according to alpha, beta schedule as well as predict noise for training. Also known as **Samplers**.
 *Examples*: [DDPM](https://arxiv.org/abs/2006.11239), [DDIM](https://arxiv.org/abs/2010.02502), [PNDM](https://arxiv.org/abs/2202.09778), [DEIS](https://arxiv.org/abs/2204.13902)
 
 <p align="center">

docker/diffusers-flax-cpu/Dockerfile

@@ -0,0 +1,42 @@
+FROM ubuntu:20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# follow the instructions here: https://cloud.google.com/tpu/docs/run-in-container#train_a_jax_model_in_a_docker_container
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --upgrade --no-cache-dir \
+        clu \
+        "jax[cpu]>=0.2.16,!=0.3.2" \
+        "flax>=0.4.1" \
+        "jaxlib>=0.1.65" && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docker/diffusers-flax-tpu/Dockerfile

@@ -0,0 +1,44 @@
+FROM ubuntu:20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# follow the instructions here: https://cloud.google.com/tpu/docs/run-in-container#train_a_jax_model_in_a_docker_container
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        "jax[tpu]>=0.2.16,!=0.3.2" \
+        -f https://storage.googleapis.com/jax-releases/libtpu_releases.html && \
+    python3 -m pip install --upgrade --no-cache-dir \
+        clu \
+        "flax>=0.4.1" \
+        "jaxlib>=0.1.65" && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docker/diffusers-onnxruntime-cpu/Dockerfile

@@ -0,0 +1,42 @@
+FROM ubuntu:20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        torch \
+        torchvision \
+        torchaudio \
+        onnxruntime \
+        --extra-index-url https://download.pytorch.org/whl/cpu && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docker/diffusers-onnxruntime-cuda/Dockerfile

@@ -0,0 +1,42 @@
+FROM nvidia/cuda:11.6.2-cudnn8-devel-ubuntu20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        torch \
+        torchvision \
+        torchaudio \
+        "onnxruntime-gpu>=1.13.1" \
+        --extra-index-url https://download.pytorch.org/whl/cu117 && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docker/diffusers-pytorch-cpu/Dockerfile

@@ -0,0 +1,41 @@
+FROM ubuntu:20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        torch \
+        torchvision \
+        torchaudio \
+        --extra-index-url https://download.pytorch.org/whl/cpu && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docker/diffusers-pytorch-cuda/Dockerfile

@@ -0,0 +1,41 @@
+FROM nvidia/cuda:11.7.1-cudnn8-runtime-ubuntu20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y bash \
+                   build-essential \
+                   git \
+                   git-lfs \
+                   curl \
+                   ca-certificates \
+                   python3.8 \
+                   python3-pip \
+                   python3.8-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        torch \
+        torchvision \
+        torchaudio \
+        --extra-index-url https://download.pytorch.org/whl/cu117 && \
+    python3 -m pip install --no-cache-dir \
+        accelerate \
+        datasets \
+        hf-doc-builder \
+        huggingface-hub \
+        modelcards \
+        numpy \
+        scipy \
+        tensorboard \
+        transformers
+
+CMD ["/bin/bash"]

docs/source/_toctree.yml

@@ -10,11 +10,13 @@
   - sections:
     - local: using-diffusers/loading
       title: "Loading Pipelines, Models, and Schedulers"
+    - local: using-diffusers/schedulers
+      title: "Using different Schedulers"
     - local: using-diffusers/configuration
       title: "Configuring Pipelines, Models, and Schedulers"
-    - local: using-diffusers/custom_pipelines
-      title: "Loading and Creating Custom Pipelines"
-    title: "Loading"
+    - local: using-diffusers/custom_pipeline_overview
+      title: "Loading and Adding Custom Pipelines"
+    title: "Loading & Hub"
   - sections:
     - local: using-diffusers/unconditional_image_generation
       title: "Unconditional Image Generation"
@@ -24,9 +26,19 @@
       title: "Text-Guided Image-to-Image"
     - local: using-diffusers/inpaint
       title: "Text-Guided Image-Inpainting"
-    - local: using-diffusers/custom
-      title: "Create a custom pipeline"
+    - local: using-diffusers/custom_pipeline_examples
+      title: "Community Pipelines"
+    - local: using-diffusers/contribute_pipeline
+      title: "How to contribute a Pipeline"
     title: "Pipelines for Inference"
+  - sections:
+    - local: using-diffusers/rl
+      title: "Reinforcement Learning"
+    - local: using-diffusers/audio
+      title: "Audio"
+    - local: using-diffusers/other-modalities
+      title: "Other Modalities"
+    title: "Taking Diffusers Beyond Images"
   title: "Using Diffusers"
 - sections:
   - local: optimization/fp16
@@ -34,7 +46,7 @@
   - local: optimization/onnx
     title: "ONNX"
   - local: optimization/open_vino
-    title: "Open Vino"
+    title: "OpenVINO"
   - local: optimization/mps
     title: "MPS"
   title: "Optimization/Special Hardware"
@@ -44,9 +56,11 @@
   - local: training/unconditional_training
     title: "Unconditional Image Generation"
   - local: training/text_inversion
-    title: "Text Inversion"
+    title: "Textual Inversion"
+  - local: training/dreambooth
+    title: "Dreambooth"
   - local: training/text2image
-    title: "Text-to-image"
+    title: "Text-to-image fine-tuning"
   title: "Training"
 - sections:
   - local: conceptual/stable_diffusion
@@ -74,6 +88,10 @@
   - sections:
     - local: api/pipelines/overview
       title: "Overview"
+    - local: api/pipelines/alt_diffusion
+      title: "AltDiffusion"
+    - local: api/pipelines/cycle_diffusion
+      title: "Cycle Diffusion"
     - local: api/pipelines/ddim
       title: "DDIM"
     - local: api/pipelines/ddpm
@@ -90,5 +108,15 @@
       title: "Stable Diffusion"
     - local: api/pipelines/stochastic_karras_ve
       title: "Stochastic Karras VE"
+    - local: api/pipelines/dance_diffusion
+      title: "Dance Diffusion"
+    - local: api/pipelines/vq_diffusion
+      title: "VQ Diffusion"
+    - local: api/pipelines/repaint
+      title: "RePaint"
     title: "Pipelines"
+  - sections:
+    - local: api/experimental/rl
+      title: "RL Planning"
+    title: "Experimental Features"
   title: "API"

docs/source/api/configuration.mdx

@@ -15,9 +15,9 @@ specific language governing permissions and limitations under the License.
 In Diffusers, schedulers of type [`schedulers.scheduling_utils.SchedulerMixin`], and models of type [`ModelMixin`] inherit from [`ConfigMixin`] which conveniently takes care of storing all parameters that are 
 passed to the respective `__init__` methods in a JSON-configuration file.
 
-TODO(PVP) - add example and better info here
-
 ## ConfigMixin
+
 [[autodoc]] ConfigMixin
+	- load_config
 	- from_config
 	- save_config

docs/source/api/diffusion_pipeline.mdx

@@ -32,6 +32,9 @@ Any pipeline object can be saved locally with [`~DiffusionPipeline.save_pretrain
 [[autodoc]] DiffusionPipeline
 	- from_pretrained
 	- save_pretrained
+	- to
+	- device
+	- components
 
 ## ImagePipelineOutput
 By default diffusion pipelines return an object of class

docs/source/api/experimental/rl.mdx

@@ -10,6 +10,6 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Custom Pipeline
+# TODO
 
-Under construction 🚧
+Coming soon!

docs/source/api/models.mdx

@@ -25,6 +25,12 @@ The models are built on the base class ['ModelMixin'] that is a `torch.nn.module
 ## UNet2DModel
 [[autodoc]] UNet2DModel
 
+## UNet1DOutput
+[[autodoc]] models.unet_1d.UNet1DOutput
+
+## UNet1DModel
+[[autodoc]] UNet1DModel
+
 ## UNet2DConditionOutput
 [[autodoc]] models.unet_2d_condition.UNet2DConditionOutput
 
@@ -46,6 +52,12 @@ The models are built on the base class ['ModelMixin'] that is a `torch.nn.module
 ## AutoencoderKL
 [[autodoc]] AutoencoderKL
 
+## Transformer2DModel
+[[autodoc]] Transformer2DModel
+
+## Transformer2DModelOutput
+[[autodoc]] models.attention.Transformer2DModelOutput
+
 ## FlaxModelMixin
 [[autodoc]] FlaxModelMixin
 

docs/source/api/pipelines/alt_diffusion.mdx

@@ -0,0 +1,83 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# AltDiffusion
+
+AltDiffusion was proposed in [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Zhongzhi Chen, Guang Liu, Bo-Wen Zhang, Fulong Ye, Qinghong Yang, Ledell Wu
+
+The abstract of the paper is the following:
+
+*In this work, we present a conceptually simple and effective method to train a strong bilingual multimodal representation model. Starting from the pretrained multimodal representation model CLIP released by OpenAI, we switched its text encoder with a pretrained multilingual text encoder XLM-R, and aligned both languages and image representations by a two-stage training schema consisting of teacher learning and contrastive learning. We validate our method through evaluations of a wide range of tasks. We set new state-of-the-art performances on a bunch of tasks including ImageNet-CN, Flicker30k- CN, and COCO-CN. Further, we obtain very close performances with CLIP on almost all tasks, suggesting that one can simply alter the text encoder in CLIP for extended capabilities such as multilingual understanding.*
+
+
+*Overview*:
+
+| Pipeline | Tasks | Colab | Demo
+|---|---|:---:|:---:|
+| [pipeline_alt_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py) | *Text-to-Image Generation* | - | -
+| [pipeline_alt_diffusion_img2img.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py) | *Image-to-Image Text-Guided Generation* | - |-
+
+## Tips
+
+- AltDiffusion is conceptually exaclty the same as [Stable Diffusion](./api/pipelines/stable_diffusion).
+
+- *Run AltDiffusion*
+
+AltDiffusion can be tested very easily with the [`AltDiffusionPipeline`], [`AltDiffusionImg2ImgPipeline`] and the `"BAAI/AltDiffusion-m9"` checkpoint exactly in the same way it is shown in the [Conditional Image Generation Guide](./using-diffusers/conditional_image_generation) and the [Image-to-Image Generation Guide](./using-diffusers/img2img).
+
+- *How to load and use different schedulers.*
+
+The alt diffusion pipeline uses [`DDIMScheduler`] scheduler by default. But `diffusers` provides many other schedulers that can be used with the alt diffusion pipeline such as [`PNDMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`] etc.
+To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`] method or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the [`EulerDiscreteScheduler`], you can do the following:
+
+```python
+>>> from diffusers import AltDiffusionPipeline, EulerDiscreteScheduler
+
+>>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9")
+>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+>>> # or
+>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("BAAI/AltDiffusion-m9", subfolder="scheduler")
+>>> pipeline = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9", scheduler=euler_scheduler)
+```
+
+
+- *How to conver all use cases with multiple or single pipeline*
+
+If you want to use all possible use cases in a single `DiffusionPipeline` we recommend using the `components` functionality to instantiate all components in the most memory-efficient way:
+
+```python
+>>> from diffusers import (
+...     AltDiffusionPipeline,
+...     AltDiffusionImg2ImgPipeline,
+... )
+
+>>> text2img = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9")
+>>> img2img = AltDiffusionImg2ImgPipeline(**text2img.components)
+
+>>> # now you can use text2img(...) and img2img(...) just like the call methods of each respective pipeline
+```
+
+## AltDiffusionPipelineOutput
+[[autodoc]] pipelines.alt_diffusion.AltDiffusionPipelineOutput
+
+## AltDiffusionPipeline
+[[autodoc]] AltDiffusionPipeline
+	- __call__
+	- enable_attention_slicing
+	- disable_attention_slicing
+
+## AltDiffusionImg2ImgPipeline
+[[autodoc]] AltDiffusionImg2ImgPipeline
+	- __call__
+	- enable_attention_slicing
+	- disable_attention_slicing

docs/source/api/pipelines/cycle_diffusion.mdx

@@ -0,0 +1,99 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Cycle Diffusion
+
+## Overview
+
+Cycle Diffusion is a Text-Guided Image-to-Image Generation model proposed in [Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance](https://arxiv.org/abs/2210.05559) by Chen Henry Wu, Fernando De la Torre.
+
+The abstract of the paper is the following:
+
+*Diffusion models have achieved unprecedented performance in generative modeling. The commonly-adopted formulation of the latent code of diffusion models is a sequence of gradually denoised samples, as opposed to the simpler (e.g., Gaussian) latent space of GANs, VAEs, and normalizing flows. This paper provides an alternative, Gaussian formulation of the latent space of various diffusion models, as well as an invertible DPM-Encoder that maps images into the latent space. While our formulation is purely based on the definition of diffusion models, we demonstrate several intriguing consequences. (1) Empirically, we observe that a common latent space emerges from two diffusion models trained independently on related domains. In light of this finding, we propose CycleDiffusion, which uses DPM-Encoder for unpaired image-to-image translation. Furthermore, applying CycleDiffusion to text-to-image diffusion models, we show that large-scale text-to-image diffusion models can be used as zero-shot image-to-image editors. (2) One can guide pre-trained diffusion models and GANs by controlling the latent codes in a unified, plug-and-play formulation based on energy-based models. Using the CLIP model and a face recognition model as guidance, we demonstrate that diffusion models have better coverage of low-density sub-populations and individuals than GANs.*
+
+*Tips*:
+- The Cycle Diffusion pipeline is fully compatible with any [Stable Diffusion](./stable_diffusion) checkpoints
+- Currently Cycle Diffusion only works with the [`DDIMScheduler`].
+
+*Example*:
+
+In the following we should how to best use the [`CycleDiffusionPipeline`]
+
+```python
+import requests
+import torch
+from PIL import Image
+from io import BytesIO
+
+from diffusers import CycleDiffusionPipeline, DDIMScheduler
+
+# load the pipeline
+# make sure you're logged in with `huggingface-cli login`
+model_id_or_path = "CompVis/stable-diffusion-v1-4"
+scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler")
+pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda")
+
+# let's download an initial image
+url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png"
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((512, 512))
+init_image.save("horse.png")
+
+# let's specify a prompt
+source_prompt = "An astronaut riding a horse"
+prompt = "An astronaut riding an elephant"
+
+# call the pipeline
+image = pipe(
+    prompt=prompt,
+    source_prompt=source_prompt,
+    init_image=init_image,
+    num_inference_steps=100,
+    eta=0.1,
+    strength=0.8,
+    guidance_scale=2,
+    source_guidance_scale=1,
+).images[0]
+
+image.save("horse_to_elephant.png")
+
+# let's try another example
+# See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion
+url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png"
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((512, 512))
+init_image.save("black.png")
+
+source_prompt = "A black colored car"
+prompt = "A blue colored car"
+
+# call the pipeline
+torch.manual_seed(0)
+image = pipe(
+    prompt=prompt,
+    source_prompt=source_prompt,
+    init_image=init_image,
+    num_inference_steps=100,
+    eta=0.1,
+    strength=0.85,
+    guidance_scale=3,
+    source_guidance_scale=1,
+).images[0]
+
+image.save("black_to_blue.png")
+```
+
+## CycleDiffusionPipeline
+[[autodoc]] CycleDiffusionPipeline
+	- __call__

docs/source/api/pipelines/dance_diffusion.mdx

@@ -0,0 +1,33 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Dance Diffusion
+
+## Overview
+
+[Dance Diffusion](https://github.com/Harmonai-org/sample-generator) by Zach Evans.
+
+Dance Diffusion is the first in a suite of generative audio tools for producers and musicians to be released by Harmonai.
+For more info or to get involved in the development of these tools, please visit https://harmonai.org and fill out the form on the front page.
+
+The original codebase of this implementation can be found [here](https://github.com/Harmonai-org/sample-generator).
+
+## Available Pipelines:
+
+| Pipeline | Tasks | Colab
+|---|---|:---:|
+| [pipeline_dance_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/dance_diffusion/pipeline_dance_diffusion.py) | *Unconditional Audio Generation* | - |
+
+
+## DanceDiffusionPipeline
+[[autodoc]] DanceDiffusionPipeline
+    - __call__

docs/source/api/pipelines/ddim.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # DDIM
 
 ## Overview
@@ -8,7 +20,8 @@ The abstract of the paper is the following:
 
 Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.
 
-The original codebase of this paper can be found [here](https://github.com/ermongroup/ddim).
+The original codebase of this paper can be found here: [ermongroup/ddim](https://github.com/ermongroup/ddim).
+For questions, feel free to contact the author on [tsong.me](https://tsong.me/).
 
 ## Available Pipelines:
 

docs/source/api/pipelines/ddpm.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # DDPM
 
 ## Overview

docs/source/api/pipelines/latent_diffusion.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # Latent Diffusion
 
 ## Overview
@@ -21,10 +33,15 @@ The original codebase can be found [here](https://github.com/CompVis/latent-diff
 | Pipeline | Tasks | Colab
 |---|---|:---:|
 | [pipeline_latent_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py) | *Text-to-Image Generation* | - |
+| [pipeline_latent_diffusion_superresolution.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py) | *Super Resolution* | - |
 
 ## Examples:
 
 
 ## LDMTextToImagePipeline
-[[autodoc]] pipelines.latent_diffusion.pipeline_latent_diffusion.LDMTextToImagePipeline
+[[autodoc]] LDMTextToImagePipeline
+    - __call__
+
+## LDMSuperResolutionPipeline
+[[autodoc]] LDMSuperResolutionPipeline
     - __call__

docs/source/api/pipelines/latent_diffusion_uncond.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # Unconditional Latent Diffusion
 
 ## Overview

docs/source/api/pipelines/overview.mdx

@@ -28,7 +28,7 @@ or created independently from each other.
 
 To that end, we strive to offer all open-sourced, state-of-the-art diffusion system under a unified API. 
 More specifically, we strive to provide pipelines that
-- 1. can load the officially published weights and yield 1-to-1 the same outputs as the original implementation according to the corresponding paper (*e.g.* [LatentDiffusionPipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/latent_diffusion), uses the officially released weights of [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752)),
+- 1. can load the officially published weights and yield 1-to-1 the same outputs as the original implementation according to the corresponding paper (*e.g.* [LDMTextToImagePipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/latent_diffusion), uses the officially released weights of [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752)),
 - 2. have a simple user interface to run the model in inference (see the [Pipelines API](#pipelines-api) section), 
 - 3. are easy to understand with code that is self-explanatory and can be read along-side the official paper (see [Pipelines summary](#pipelines-summary)),
 - 4. can easily be contributed by the community (see the [Contribution](#contribution) section).
@@ -41,19 +41,26 @@ If you are looking for *official* training examples, please have a look at [exam
 The following table summarizes all officially supported pipelines, their corresponding paper, and if 
 available a colab notebook to directly try them out.
 
+
 | Pipeline | Paper | Tasks | Colab
 |---|---|:---:|:---:|
-| [ddpm](./ddpm) | [**Denoising Diffusion Probabilistic Models**](https://arxiv.org/abs/2006.11239) | Unconditional Image Generation |
-| [ddim](./ddim) | [**Denoising Diffusion Implicit Models**](https://arxiv.org/abs/2010.02502) | Unconditional Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
-| [latent_diffusion](./latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)| Text-to-Image Generation | 
-| [latent_diffusion_uncond](./latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752) | Unconditional Image Generation | 
-| [pndm](./pndm) | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778) | Unconditional Image Generation | 
-| [score_sde_ve](./score_sde_ve) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
-| [score_sde_vp](./score_sde_vp) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
-| [stable_diffusion](./stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
-| [stable_diffusion](./stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
-| [stable_diffusion](./stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
-| [stochastic_karras_ve](./stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | 
+| [alt_diffusion](./api/pipelines/alt_diffusion) | [**AltDiffusion**](https://arxiv.org/abs/2211.06679) | Image-to-Image Text-Guided Generation | -
+| [cycle_diffusion](./api/pipelines/cycle_diffusion) | [**Cycle Diffusion**](https://arxiv.org/abs/2210.05559) | Image-to-Image Text-Guided Generation |
+| [dance_diffusion](./api/pipelines/dance_diffusion) | [**Dance Diffusion**](https://github.com/williamberman/diffusers.git) | Unconditional Audio Generation |
+| [ddpm](./api/pipelines/ddpm) | [**Denoising Diffusion Probabilistic Models**](https://arxiv.org/abs/2006.11239) | Unconditional Image Generation |
+| [ddim](./api/pipelines/ddim) | [**Denoising Diffusion Implicit Models**](https://arxiv.org/abs/2010.02502) | Unconditional Image Generation |
+| [latent_diffusion](./api/pipelines/latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)| Text-to-Image Generation | 
+| [latent_diffusion](./api/pipelines/latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)| Super Resolution Image-to-Image | 
+| [latent_diffusion_uncond](./api/pipelines/latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752) | Unconditional Image Generation | 
+| [pndm](./api/pipelines/pndm) | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778) | Unconditional Image Generation | 
+| [score_sde_ve](./api/pipelines/score_sde_ve) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
+| [score_sde_vp](./api/pipelines/score_sde_vp) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
+| [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
+| [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
+| [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | 
+| [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation | 
+
 
 **Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers. 
 
@@ -67,8 +74,8 @@ Diffusion models often consist of multiple independently-trained models or other
 Each model has been trained independently on a different task and the scheduler can easily be swapped out and replaced with a different one. 
 During inference, we however want to be able to easily load all components and use them in inference - even if one component, *e.g.* CLIP's text encoder, originates from a different library, such as [Transformers](https://github.com/huggingface/transformers). To that end, all pipelines provide the following functionality:
 
-- [`from_pretrained` method](../diffusion_pipeline) that accepts a Hugging Face Hub repository id, *e.g.* [CompVis/stable-diffusion-v1-4](https://huggingface.co/CompVis/stable-diffusion-v1-4) or a path to a local directory, *e.g.*
-"./stable-diffusion". To correctly retrieve which models and components should be loaded, one has to provide a `model_index.json` file, *e.g.* [CompVis/stable-diffusion-v1-4/model_index.json](https://huggingface.co/CompVis/stable-diffusion-v1-4/blob/main/model_index.json), which defines all components that should be 
+- [`from_pretrained` method](../diffusion_pipeline) that accepts a Hugging Face Hub repository id, *e.g.* [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5) or a path to a local directory, *e.g.*
+"./stable-diffusion". To correctly retrieve which models and components should be loaded, one has to provide a `model_index.json` file, *e.g.* [runwayml/stable-diffusion-v1-5/model_index.json](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/model_index.json), which defines all components that should be
 loaded into the pipelines. More specifically, for each model/component one needs to define the format `<name>: ["<library>", "<class name>"]`. `<name>` is the attribute name given to the loaded instance of `<class name>` which can be found in the library or pipeline folder called `"<library>"`.
 - [`save_pretrained`](../diffusion_pipeline) that accepts a local path, *e.g.* `./stable-diffusion` under which all models/components of the pipeline will be saved. For each component/model a folder is created inside the local path that is named after the given attribute name, *e.g.* `./stable_diffusion/unet`. 
 In addition, a `model_index.json` file is created at the root of the local path, *e.g.* `./stable_diffusion/model_index.json` so that the complete pipeline can again be instantiated 
@@ -100,7 +107,7 @@ logic including pre-processing, an unrolled diffusion loop, and post-processing
 # make sure you're logged in with `huggingface-cli login`
 from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
@@ -123,7 +130,7 @@ from diffusers import StableDiffusionImg2ImgPipeline
 # load the pipeline
 device = "cuda"
 pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16
+    "runwayml/stable-diffusion-v1-5", revision="fp16", torch_dtype=torch.float16
 ).to(device)
 
 # let's download an initial image
@@ -151,10 +158,10 @@ You can generate your own latents to reproduce results, or tweak your prompt on
 The `StableDiffusionInpaintPipeline` lets you edit specific parts of an image by providing a mask and text prompt.
 
 ```python
-from io import BytesIO
-
-import requests
 import PIL
+import requests
+import torch
+from io import BytesIO
 
 from diffusers import StableDiffusionInpaintPipeline
 
@@ -170,15 +177,15 @@ mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data
 init_image = download_image(img_url).resize((512, 512))
 mask_image = download_image(mask_url).resize((512, 512))
 
-device = "cuda"
 pipe = StableDiffusionInpaintPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16
-).to(device)
+    "runwayml/stable-diffusion-inpainting",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
 
-prompt = "a cat sitting on a bench"
-images = pipe(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images
-
-images[0].save("cat_on_bench.png")
+prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
 ```
 
 You can also run this example on colab [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)

docs/source/api/pipelines/pndm.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # PNDM
 
 ## Overview

docs/source/api/pipelines/repaint.mdx

@@ -0,0 +1,77 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# RePaint
+
+## Overview
+
+[RePaint: Inpainting using Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2201.09865) (PNDM) by Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, Luc Van Gool.
+
+The abstract of the paper is the following:
+
+Free-form inpainting is the task of adding new content to an image in the regions specified by an arbitrary binary mask. Most existing approaches train for a certain distribution of masks, which limits their generalization capabilities to unseen mask types. Furthermore, training with pixel-wise and perceptual losses often leads to simple textural extensions towards the missing areas instead of semantically meaningful generation. In this work, we propose RePaint: A Denoising Diffusion Probabilistic Model (DDPM) based inpainting approach that is applicable to even extreme masks. We employ a pretrained unconditional DDPM as the generative prior. To condition the generation process, we only alter the reverse diffusion iterations by sampling the unmasked regions using the given image information. Since this technique does not modify or condition the original DDPM network itself, the model produces high-quality and diverse output images for any inpainting form. We validate our method for both faces and general-purpose image inpainting using standard and extreme masks.
+RePaint outperforms state-of-the-art Autoregressive, and GAN approaches for at least five out of six mask distributions.
+
+The original codebase can be found [here](https://github.com/andreas128/RePaint).
+
+## Available Pipelines:
+
+| Pipeline                                                                                                                      | Tasks              | Colab
+|-------------------------------------------------------------------------------------------------------------------------------|--------------------|:---:|
+| [pipeline_repaint.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/repaint/pipeline_repaint.py) | *Image Inpainting* | - |
+
+## Usage example
+
+```python
+from io import BytesIO
+
+import torch
+
+import PIL
+import requests
+from diffusers import RePaintPipeline, RePaintScheduler
+
+
+def download_image(url):
+    response = requests.get(url)
+    return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+img_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/celeba_hq_256.png"
+mask_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/mask_256.png"
+
+# Load the original image and the mask as PIL images
+original_image = download_image(img_url).resize((256, 256))
+mask_image = download_image(mask_url).resize((256, 256))
+
+# Load the RePaint scheduler and pipeline based on a pretrained DDPM model
+scheduler = RePaintScheduler.from_pretrained("google/ddpm-ema-celebahq-256")
+pipe = RePaintPipeline.from_pretrained("google/ddpm-ema-celebahq-256", scheduler=scheduler)
+pipe = pipe.to("cuda")
+
+generator = torch.Generator(device="cuda").manual_seed(0)
+output = pipe(
+    original_image=original_image,
+    mask_image=mask_image,
+    num_inference_steps=250,
+    eta=0.0,
+    jump_length=10,
+    jump_n_sample=10,
+    generator=generator,
+)
+inpainted_image = output.images[0]
+```
+
+## RePaintPipeline
+[[autodoc]] pipelines.repaint.pipeline_repaint.RePaintPipeline
+    - __call__
+

docs/source/api/pipelines/score_sde_ve.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # Score SDE VE
 
 ## Overview

docs/source/api/pipelines/stable_diffusion.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # Stable diffusion pipelines
 
 Stable Diffusion is a text-to-image _latent diffusion_ model created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/) and [LAION](https://laion.ai/). It's trained on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) dataset. This model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts. With its 860M UNet and 123M text encoder, the model is relatively lightweight and can run on consumer GPUs.
@@ -17,6 +29,45 @@ For more details about how Stable Diffusion works and how it differs from the ba
 | [pipeline_stable_diffusion_img2img.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) | *Image-to-Image Text-Guided Generation* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb) | [🤗 Diffuse the Rest](https://huggingface.co/spaces/huggingface/diffuse-the-rest)
 | [pipeline_stable_diffusion_inpaint.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | **Experimental** – *Text-Guided Image Inpainting* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb) | Coming soon
 
+## Tips
+
+### How to load and use different schedulers.
+
+The stable diffusion pipeline uses [`PNDMScheduler`] scheduler by default. But `diffusers` provides many other schedulers that can be used with the stable diffusion pipeline such as [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`] etc.
+To use a different scheduler, you can either change it via the [`ConfigMixin.from_config`] method or pass the `scheduler` argument to the `from_pretrained` method of the pipeline. For example, to use the [`EulerDiscreteScheduler`], you can do the following:
+
+```python
+>>> from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler
+
+>>> pipeline = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+>>> # or
+>>> euler_scheduler = EulerDiscreteScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
+>>> pipeline = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=euler_scheduler)
+```
+
+
+### How to conver all use cases with multiple or single pipeline
+
+If you want to use all possible use cases in a single `DiffusionPipeline` you can either:
+- Make use of the [Stable Diffusion Mega Pipeline](https://github.com/huggingface/diffusers/tree/main/examples/community#stable-diffusion-mega) or 
+- Make use of the `components` functionality to instantiate all components in the most memory-efficient way:
+
+```python
+>>> from diffusers import (
+...     StableDiffusionPipeline,
+...     StableDiffusionImg2ImgPipeline,
+...     StableDiffusionInpaintPipeline,
+... )
+
+>>> text2img = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+>>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components)
+>>> inpaint = StableDiffusionInpaintPipeline(**text2img.components)
+
+>>> # now you can use text2img(...), img2img(...), inpaint(...) just like the call methods of each respective pipeline
+```
+
 ## StableDiffusionPipelineOutput
 [[autodoc]] pipelines.stable_diffusion.StableDiffusionPipelineOutput
 

docs/source/api/pipelines/stochastic_karras_ve.mdx

@@ -1,3 +1,15 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
 # Stochastic Karras VE
 
 ## Overview

docs/source/api/pipelines/vq_diffusion.mdx

@@ -0,0 +1,34 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# VQDiffusion
+
+## Overview
+
+[Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) by Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, Baining Guo
+
+The abstract of the paper is the following:
+
+We present the vector quantized diffusion (VQ-Diffusion) model for text-to-image generation. This method is based on a vector quantized variational autoencoder (VQ-VAE) whose latent space is modeled by a conditional variant of the recently developed Denoising Diffusion Probabilistic Model (DDPM). We find that this latent-space method is well-suited for text-to-image generation tasks because it not only eliminates the unidirectional bias with existing methods but also allows us to incorporate a mask-and-replace diffusion strategy to avoid the accumulation of errors, which is a serious problem with existing methods. Our experiments show that the VQ-Diffusion produces significantly better text-to-image generation results when compared with conventional autoregressive (AR) models with similar numbers of parameters. Compared with previous GAN-based text-to-image methods, our VQ-Diffusion can handle more complex scenes and improve the synthesized image quality by a large margin. Finally, we show that the image generation computation in our method can be made highly efficient by reparameterization. With traditional AR methods, the text-to-image generation time increases linearly with the output image resolution and hence is quite time consuming even for normal size images. The VQ-Diffusion allows us to achieve a better trade-off between quality and speed. Our experiments indicate that the VQ-Diffusion model with the reparameterization is fifteen times faster than traditional AR methods while achieving a better image quality.
+
+The original codebase can be found [here](https://github.com/microsoft/VQ-Diffusion).
+
+## Available Pipelines:
+
+| Pipeline | Tasks | Colab
+|---|---|:---:|
+| [pipeline_vq_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/vq_diffusion/pipeline_vq_diffusion.py) | *Text-to-Image Generation* | - |
+
+
+## VQDiffusionPipeline
+[[autodoc]] pipelines.vq_diffusion.pipeline_vq_diffusion.VQDiffusionPipeline
+    - __call__

docs/source/api/schedulers.mdx

@@ -16,7 +16,7 @@ Diffusers contains multiple pre-built schedule functions for the diffusion proce
 
 ## What is a scheduler?
 
-The schedule functions, denoted *Schedulers* in the library take in the output of a trained model, a sample which the diffusion process is iterating on, and a timestep to return a denoised sample.
+The schedule functions, denoted *Schedulers* in the library take in the output of a trained model, a sample which the diffusion process is iterating on, and a timestep to return a denoised sample. That's why schedulers may also be called *Samplers* in other diffusion models implementations.
 
 - Schedulers define the methodology for iteratively adding noise to an image or for updating a sample based on model outputs.
     - adding noise in different manners represent the algorithmic processes to train a diffusion model by adding noise to images.
@@ -70,6 +70,12 @@ Original paper can be found [here](https://arxiv.org/abs/2010.02502).
 
 [[autodoc]] DDPMScheduler
 
+#### Multistep DPM-Solver
+
+Original paper can be found [here](https://arxiv.org/abs/2206.00927) and the [improved version](https://arxiv.org/abs/2211.01095). The original implementation can be found [here](https://github.com/LuChengTHU/dpm-solver).
+
+[[autodoc]] DPMSolverMultistepScheduler
+
 #### Variance exploding, stochastic sampling from Karras et. al
 
 Original paper can be found [here](https://arxiv.org/abs/2006.11239).
@@ -89,13 +95,19 @@ Original implementation can be found [here](https://github.com/crowsonkb/k-diffu
 
 [[autodoc]] PNDMScheduler
 
-#### variance exploding stochastic differential equation (SDE) scheduler
+#### variance exploding stochastic differential equation (VE-SDE) scheduler
 
 Original paper can be found [here](https://arxiv.org/abs/2011.13456).
 
 [[autodoc]] ScoreSdeVeScheduler
 
-#### variance preserving stochastic differential equation (SDE) scheduler
+#### improved pseudo numerical methods for diffusion models (iPNDM)
+
+Original implementation can be found [here](https://github.com/crowsonkb/v-diffusion-pytorch/blob/987f8985e38208345c1959b0ea767a625831cc9b/diffusion/sampling.py#L296).
+
+[[autodoc]] IPNDMScheduler
+
+#### variance preserving stochastic differential equation (VP-SDE) scheduler
 
 Original paper can be found [here](https://arxiv.org/abs/2011.13456).
 
@@ -106,3 +118,34 @@ Score SDE-VP is under construction.
 </Tip>
 
 [[autodoc]] schedulers.scheduling_sde_vp.ScoreSdeVpScheduler
+
+#### Euler scheduler
+
+Euler scheduler (Algorithm 2) from the paper [Elucidating the Design Space of Diffusion-Based Generative Models](https://arxiv.org/abs/2206.00364) by Karras et al. (2022). Based on the original [k-diffusion](https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L51) implementation by Katherine Crowson.
+Fast scheduler which often times generates good outputs with 20-30 steps.
+
+[[autodoc]] EulerDiscreteScheduler
+
+
+#### Euler Ancestral scheduler
+
+Ancestral sampling with Euler method steps. Based on the original (k-diffusion)[https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L72] implementation by Katherine Crowson.
+Fast scheduler which often times generates good outputs with 20-30 steps.
+
+[[autodoc]] EulerAncestralDiscreteScheduler
+
+
+#### VQDiffusionScheduler
+
+Original paper can be found [here](https://arxiv.org/abs/2111.14822)
+
+[[autodoc]] VQDiffusionScheduler
+
+#### RePaint scheduler
+
+DDPM-based inpainting scheduler for unsupervised inpainting with extreme masks. 
+Intended for use with [`RePaintPipeline`].
+Based on the paper [RePaint: Inpainting using Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2201.09865) 
+and the original implementation by Andreas Lugmayr et al.: https://github.com/andreas128/RePaint
+
+[[autodoc]] RePaintScheduler

docs/source/conceptual/stable_diffusion.mdx

@@ -12,6 +12,4 @@ specific language governing permissions and limitations under the License.
 
 # Stable Diffusion
 
-Under construction 🚧
-
-For now please visit this [very in-detail blog post](https://huggingface.co/blog/stable_diffusion)
+Please visit this [very in-detail blog post](https://huggingface.co/blog/stable_diffusion) on Stable Diffusion!

docs/source/index.mdx

@@ -34,9 +34,13 @@ available a colab notebook to directly try them out.
 
 | Pipeline | Paper | Tasks | Colab
 |---|---|:---:|:---:|
+| [alt_diffusion](./api/pipelines/alt_diffusion) | [**AltDiffusion**](https://arxiv.org/abs/2211.06679) | Image-to-Image Text-Guided Generation |
+| [cycle_diffusion](./api/pipelines/cycle_diffusion) | [**Cycle Diffusion**](https://arxiv.org/abs/2210.05559) | Image-to-Image Text-Guided Generation |
+| [dance_diffusion](./api/pipelines/dance_diffusion) | [**Dance Diffusion**](https://github.com/williamberman/diffusers.git) | Unconditional Audio Generation |
 | [ddpm](./api/pipelines/ddpm) | [**Denoising Diffusion Probabilistic Models**](https://arxiv.org/abs/2006.11239) | Unconditional Image Generation |
 | [ddim](./api/pipelines/ddim) | [**Denoising Diffusion Implicit Models**](https://arxiv.org/abs/2010.02502) | Unconditional Image Generation |
 | [latent_diffusion](./api/pipelines/latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)| Text-to-Image Generation | 
+| [latent_diffusion](./api/pipelines/latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752)| Super Resolution Image-to-Image | 
 | [latent_diffusion_uncond](./api/pipelines/latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752) | Unconditional Image Generation | 
 | [pndm](./api/pipelines/pndm) | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778) | Unconditional Image Generation | 
 | [score_sde_ve](./api/pipelines/score_sde_ve) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
@@ -45,5 +49,6 @@ available a colab notebook to directly try them out.
 | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
 | [stable_diffusion](./api/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
 | [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation | 
+| [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation | 
 
 **Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers. 

docs/source/installation.mdx

@@ -12,9 +12,12 @@ specific language governing permissions and limitations under the License.
 
 # Installation
 
-Install Diffusers for with PyTorch. Support for other libraries will come in the future
+Install 🤗 Diffusers for whichever deep learning library you’re working with.
 
-🤗 Diffusers is tested on Python 3.7+, and PyTorch 1.7.0+.
+🤗 Diffusers is tested on Python 3.7+, PyTorch 1.7.0+ and flax. Follow the installation instructions below for the deep learning library you are using:
+
+- [PyTorch](https://pytorch.org/get-started/locally/) installation instructions.
+- [Flax](https://flax.readthedocs.io/en/latest/) installation instructions.
 
 ## Install with pip
 
@@ -36,12 +39,30 @@ source .env/bin/activate
 
 Now you're ready to install 🤗 Diffusers with the following command:
 
+**For PyTorch**
+
 ```bash
-pip install diffusers
+pip install diffusers["torch"]
+```
+
+**For Flax**
+
+```bash
+pip install diffusers["flax"]
 ```
 
 ## Install from source
 
+Before intsalling `diffusers` from source, make sure you have `torch` and `accelerate` installed.
+
+For `torch` installation refer to the `torch` [docs](https://pytorch.org/get-started/locally/#start-locally).
+
+To install `accelerate`
+
+```bash
+pip install accelerate
+```
+
 Install 🤗 Diffusers from source with the following command:
 
 ```bash
@@ -53,7 +74,7 @@ The `main` version is useful for staying up-to-date with the latest developments
 For instance, if a bug has been fixed since the last official release but a new release hasn't been rolled out yet.
 However, this means the `main` version may not always be stable.
 We strive to keep the `main` version operational, and most issues are usually resolved within a few hours or a day.
-If you run into a problem, please open an [Issue](https://github.com/huggingface/transformers/issues) so we can fix it even sooner!
+If you run into a problem, please open an [Issue](https://github.com/huggingface/transformers/issues), so we can fix it even sooner!
 
 ## Editable install
 
@@ -67,7 +88,18 @@ Clone the repository and install 🤗 Diffusers with the following commands:
 ```bash
 git clone https://github.com/huggingface/diffusers.git
 cd diffusers
-pip install -e .
+```
+
+**For PyTorch**
+
+```
+pip install -e ".[torch]"
+```
+
+**For Flax**
+
+```
+pip install -e ".[flax]"
 ```
 
 These commands will link the folder you cloned the repository to and your Python library paths.

docs/source/optimization/fp16.mdx

@@ -14,17 +14,21 @@ specific language governing permissions and limitations under the License.
 
 We present some techniques and ideas to optimize 🤗 Diffusers _inference_ for memory or speed.
 
-
 |                  | Latency | Speedup |
-|------------------|---------|---------|
+| ---------------- | ------- | ------- |
 | original         | 9.50s   | x1      |
 | cuDNN auto-tuner | 9.37s   | x1.01   |
-| autocast (fp16)  | 5.47s   | x1.91   |
-| fp16             | 3.61s   | x2.91   |
-| channels last    | 3.30s   | x2.87   |
+| autocast (fp16)  | 5.47s   | x1.74   |
+| fp16             | 3.61s   | x2.63   |
+| channels last    | 3.30s   | x2.88   |
 | traced UNet      | 3.21s   | x2.96   |
+| memory efficient attention  | 2.63s  | x3.61   |
 
-<em>obtained on NVIDIA TITAN RTX by generating a single image of size 512x512 from the prompt "a photo of an astronaut riding a horse on mars" with 50 DDIM steps.</em>
+<em>
+  obtained on NVIDIA TITAN RTX by generating a single image of size 512x512 from
+  the prompt "a photo of an astronaut riding a horse on mars" with 50 DDIM
+  steps.
+</em>
 
 ## Enable cuDNN auto-tuner
 
@@ -56,12 +60,12 @@ If you use a CUDA GPU, you can take advantage of `torch.autocast` to perform inf
 from torch import autocast
 from diffusers import StableDiffusionPipeline
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
 with autocast("cuda"):
-    image = pipe(prompt).images[0]  
+    image = pipe(prompt).images[0]
 ```
 
 Despite the precision loss, in our experience the final image results look the same as the `float32` versions. Feel free to experiment and report back!
@@ -72,14 +76,14 @@ To save more GPU memory and get even more speed, you can load and run the model
 
 ```Python
 pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     revision="fp16",
     torch_dtype=torch.float16,
 )
 pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
-image = pipe(prompt).images[0]  
+image = pipe(prompt).images[0]
 ```
 
 ## Sliced attention for additional memory savings
@@ -87,7 +91,10 @@ image = pipe(prompt).images[0]
 For even additional memory savings, you can use a sliced version of attention that performs the computation in steps instead of all at once.
 
 <Tip>
-Attention slicing is useful even if a batch size of just 1 is used - as long as the model uses more than one attention head. If there is more than one attention head the *QK^T* attention matrix can be computed sequentially for each head which can save a significant amount of memory.
+  Attention slicing is useful even if a batch size of just 1 is used - as long
+  as the model uses more than one attention head. If there is more than one
+  attention head the *QK^T* attention matrix can be computed sequentially for
+  each head which can save a significant amount of memory.
 </Tip>
 
 To perform the attention computation sequentially over each head, you only need to invoke [`~StableDiffusionPipeline.enable_attention_slicing`] in your pipeline before inference, like here:
@@ -97,7 +104,7 @@ import torch
 from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     revision="fp16",
     torch_dtype=torch.float16,
 )
@@ -105,11 +112,55 @@ pipe = pipe.to("cuda")
 
 prompt = "a photo of an astronaut riding a horse on mars"
 pipe.enable_attention_slicing()
-image = pipe(prompt).images[0]  
+image = pipe(prompt).images[0]
 ```
 
 There's a small performance penalty of about 10% slower inference times, but this method allows you to use Stable Diffusion in as little as 3.2 GB of VRAM!
 
+## Offloading to CPU with accelerate for memory savings
+
+For additional memory savings, you can offload the weights to CPU and load them to GPU when performing the forward pass.
+
+To perform CPU offloading, all you have to do is invoke [`~StableDiffusionPipeline.enable_sequential_cpu_offload`]:
+
+```Python
+import torch
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+pipe.enable_sequential_cpu_offload()
+image = pipe(prompt).images[0]
+```
+
+And you can get the memory consumption to < 2GB.
+
+If is also possible to chain it with attention slicing for minimal memory consumption, running it in as little as < 800mb of GPU vRAM:
+
+```Python
+import torch
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+pipe.enable_sequential_cpu_offload()
+pipe.enable_attention_slicing(1)
+
+image = pipe(prompt).images[0]
+```
+
 ## Using Channels Last memory format
 
 Channels last memory format is an alternative way of ordering NCHW tensors in memory preserving dimensions ordering. Channels last tensors ordered in such a way that channels become the densest dimension (aka storing images pixel-per-pixel). Since not all operators currently support channels last format it may result in a worst performance, so it's better to try it and see if it works for your model.
@@ -152,7 +203,7 @@ def generate_inputs():
 
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     revision="fp16",
     torch_dtype=torch.float16,
 ).to("cuda")
@@ -216,7 +267,7 @@ class UNet2DConditionOutput:
 
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     revision="fp16",
     torch_dtype=torch.float16,
 ).to("cuda")
@@ -240,3 +291,41 @@ pipe.unet = TracedUNet()
 with torch.inference_mode():
     image = pipe([prompt] * 1, num_inference_steps=50).images[0]
 ```
+
+
+## Memory Efficient Attention
+Recent work on optimizing the bandwitdh in the attention block have generated huge speed ups and gains in GPU memory usage. The most recent being Flash Attention (from @tridao, [code](https://github.com/HazyResearch/flash-attention), [paper](https://arxiv.org/pdf/2205.14135.pdf)) .
+Here are the speedups we obtain on a few Nvidia GPUs when running the inference at 512x512 with a batch size of 1 (one prompt):
+
+| GPU              	| Base Attention FP16 	| Memory Efficient Attention FP16 	|
+|------------------	|---------------------	|---------------------------------	|
+| NVIDIA Tesla T4  	| 3.5it/s             	| 5.5it/s                         	|
+| NVIDIA 3060 RTX  	| 4.6it/s             	| 7.8it/s                         	|
+| NVIDIA A10G      	| 8.88it/s            	| 15.6it/s                        	|
+| NVIDIA RTX A6000 	| 11.7it/s            	| 21.09it/s                       	|
+| NVIDIA TITAN RTX  | 12.51it/s         	| 18.22it/s                       	|
+| A100-SXM4-40GB    	| 18.6it/s            	| 29.it/s                        	|
+| A100-SXM-80GB    	| 18.7it/s            	| 29.5it/s                        	|
+
+To leverage it just make sure you have: 
+ - PyTorch > 1.12
+ - Cuda available
+ - Installed the [xformers](https://github.com/facebookresearch/xformers) library
+```python
+from diffusers import StableDiffusionPipeline
+import torch
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    revision="fp16",
+    torch_dtype=torch.float16,
+).to("cuda")
+
+pipe.enable_xformers_memory_efficient_attention()
+
+with torch.inference_mode():
+    sample = pipe("a small cat")
+
+# optional: You can disable it via
+# pipe.disable_xformers_memory_efficient_attention()
+```

docs/source/optimization/mps.mdx

@@ -17,9 +17,10 @@ specific language governing permissions and limitations under the License.
 ## Requirements
 
 - Mac computer with Apple silicon (M1/M2) hardware.
-- macOS 12.3 or later.
+- macOS 12.6 or later (13.0 or later recommended).
 - arm64 version of Python.
-- PyTorch [Preview (Nightly)](https://pytorch.org/get-started/locally/), version `1.13.0.dev20220830` or later.
+- PyTorch 1.13. You can install it with `pip` or `conda` using the instructions in https://pytorch.org/get-started/locally/.
+
 
 ## Inference Pipeline
 
@@ -31,9 +32,12 @@ We recommend to "prime" the pipeline using an additional one-time pass through i
 # make sure you're logged in with `huggingface-cli login`
 from diffusers import StableDiffusionPipeline
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
 pipe = pipe.to("mps")
 
+# Recommended if your computer has < 64 GB of RAM
+pipe.enable_attention_slicing()
+
 prompt = "a photo of an astronaut riding a horse on mars"
 
 # First-time "warmup" pass (see explanation above)
@@ -43,16 +47,17 @@ _ = pipe(prompt, num_inference_steps=1)
 image = pipe(prompt).images[0]
 ```
 
+## Performance Recommendations
+
+M1/M2 performance is very sensitive to memory pressure. The system will automatically swap if it needs to, but performance will degrade significantly when it does.
+
+We recommend you use _attention slicing_ to reduce memory pressure during inference and prevent swapping, particularly if your computer has lass than 64 GB of system RAM, or if you generate images at non-standard resolutions larger than 512 × 512 pixels. Attention slicing performs the costly attention operation in multiple steps instead of all at once. It usually has a performance impact of ~20% in computers without universal memory, but we have observed _better performance_ in most Apple Silicon computers, unless you have 64 GB or more.
+
+```python
+pipeline.enable_attention_slicing()
+```
+
 ## Known Issues
 
 - As mentioned above, we are investigating a strange [first-time inference issue](https://github.com/huggingface/diffusers/issues/372).
-- Generating multiple prompts in a batch [crashes or doesn't work reliably](https://github.com/huggingface/diffusers/issues/363). We believe this might be related to the [`mps` backend in PyTorch](https://github.com/pytorch/pytorch/issues/84039#issuecomment-1237735249), but we need to investigate in more depth. For now, we recommend to iterate instead of batching.
-
-## Performance
-
-These are the results we got on a M1 Max MacBook Pro with 64 GB of RAM, running macOS Ventura Version 13.0 Beta (22A5331f). We performed Stable Diffusion text-to-image generation of the same prompt for 50 inference steps, using a guidance scale of 7.5.
-
-| Device | Steps | Time    |
-|--------|-------|---------|
-| CPU    | 50    | 213.46s |
-| MPS    | 50    | 30.81s  |
+- Generating multiple prompts in a batch [crashes or doesn't work reliably](https://github.com/huggingface/diffusers/issues/363). We believe this is related to the [`mps` backend in PyTorch](https://github.com/pytorch/pytorch/issues/84039). This is being resolved, but for now we recommend to iterate instead of batching.

docs/source/optimization/onnx.mdx

@@ -28,7 +28,7 @@ The snippet below demonstrates how to use the ONNX runtime. You need to use `Sta
 from diffusers import StableDiffusionOnnxPipeline
 
 pipe = StableDiffusionOnnxPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     revision="onnx",
     provider="CUDAExecutionProvider",
 )

docs/source/quicktour.mdx

@@ -41,7 +41,7 @@ In this guide though, you'll use [`DiffusionPipeline`] for text-to-image generat
 ```python
 >>> from diffusers import DiffusionPipeline
 
->>> generator = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
+>>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
 ```
 
 The [`DiffusionPipeline`] downloads and caches all modeling, tokenization, and scheduling components. 
@@ -49,13 +49,13 @@ Because the model consists of roughly 1.4 billion parameters, we strongly recomm
 You can move the generator object to GPU, just like you would in PyTorch.
 
 ```python
->>> generator.to("cuda")
+>>> pipeline.to("cuda")
 ```
 
-Now you can use the `generator` on your text prompt:
+Now you can use the `pipeline` on your text prompt:
 
 ```python
->>> image = generator("An image of a squirrel in Picasso style").images[0]
+>>> image = pipeline("An image of a squirrel in Picasso style").images[0]
 ```
 
 The output is by default wrapped into a [PIL Image object](https://pillow.readthedocs.io/en/stable/reference/Image.html?highlight=image#the-image-class).
@@ -68,8 +68,7 @@ You can save the image by simply calling:
 
 More advanced models, like [Stable Diffusion](https://huggingface.co/CompVis/stable-diffusion) require you to accept a [license](https://huggingface.co/spaces/CompVis/stable-diffusion-license) before running the model.
 This is due to the improved image generation capabilities of the model and the potentially harmful content that could be produced with it.
-Long story short: Head over to your stable diffusion model of choice, *e.g.* [`CompVis/stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4), read through the license and click-accept to get 
-access to the model. 
+Please, head over to your stable diffusion model of choice, *e.g.*  [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license carefully and tick the checkbox if you agree.
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 Having "click-accepted" the license, you can save your token:
 
@@ -77,13 +76,13 @@ Having "click-accepted" the license, you can save your token:
 AUTH_TOKEN = "<please-fill-with-your-token>"
 ```
 
-You can then load [`CompVis/stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4) 
+You can then load [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5)
 just like we did before only that now you need to pass your `AUTH_TOKEN`:
 
 ```python
 >>> from diffusers import DiffusionPipeline
 
->>> generator = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", use_auth_token=AUTH_TOKEN)
+>>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", use_auth_token=AUTH_TOKEN)
 ```
 
 If you do not pass your authentication token you will see that the diffusion system will not be correctly 
@@ -95,15 +94,15 @@ the weights locally via:
 
 ```
 git lfs install
-git clone https://huggingface.co/CompVis/stable-diffusion-v1-4
+git clone https://huggingface.co/runwayml/stable-diffusion-v1-5
 ```
 
 and then load locally saved weights into the pipeline. This way, you do not need to pass an authentication
-token. Assuming that `"./stable-diffusion-v1-4"` is the local path to the cloned stable-diffusion-v1-4 repo,
+token. Assuming that `"./stable-diffusion-v1-5"` is the local path to the cloned stable-diffusion-v1-5 repo,
 you can also load the pipeline as follows:
 
 ```python
->>> generator = DiffusionPipeline.from_pretrained("./stable-diffusion-v1-4")
+>>> pipeline = DiffusionPipeline.from_pretrained("./stable-diffusion-v1-5")
 ```
 
 Running the pipeline is then identical to the code above as it's the same model architecture.
@@ -116,19 +115,20 @@ Running the pipeline is then identical to the code above as it's the same model
 
 Diffusion systems can be used with multiple different [schedulers](./api/schedulers) each with their
 pros and cons. By default, Stable Diffusion runs with [`PNDMScheduler`], but it's very simple to 
-use a different scheduler. *E.g.* if you would instead like to use the [`LMSDiscreteScheduler`] scheduler,
+use a different scheduler. *E.g.* if you would instead like to use the [`EulerDiscreteScheduler`] scheduler,
 you could use it as follows:
 
 ```python
->>> from diffusers import LMSDiscreteScheduler
+>>> from diffusers import EulerDiscreteScheduler
 
->>> scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
+>>> pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", use_auth_token=AUTH_TOKEN)
 
->>> generator = StableDiffusionPipeline.from_pretrained(
-...     "CompVis/stable-diffusion-v1-4", scheduler=scheduler, use_auth_token=AUTH_TOKEN
-... )
+>>> # change scheduler to Euler
+>>> pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
 ```
 
+For more in-detail information on how to change between schedulers, please refer to the [Using Schedulers](./using-diffusers/schedulers) guide.
+
 [Stability AI's](https://stability.ai/) Stable Diffusion model is an impressive image generation model
 and can do much more than just generating images from text. We have dedicated a whole documentation page,
 just for Stable Diffusion [here](./conceptual/stable_diffusion).

docs/source/training/dreambooth.mdx

@@ -0,0 +1,240 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# DreamBooth fine-tuning example
+
+[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text-to-image models like stable diffusion given just a few (3~5) images of a subject.
+
+![Dreambooth examples from the project's blog](https://dreambooth.github.io/DreamBooth_files/teaser_static.jpg)
+_Dreambooth examples from the [project's blog](https://dreambooth.github.io)._
+
+The [Dreambooth training script](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth) shows how to implement this training procedure on a pre-trained Stable Diffusion model.
+
+<Tip warning={true}>
+
+<!-- TODO: replace with our blog when it's done -->
+
+Dreambooth fine-tuning is very sensitive to hyperparameters and easy to overfit. We recommend you take a look at our [in-depth analysis](https://huggingface.co/blog/dreambooth) with recommended settings for different subjects, and go from there.
+
+</Tip>
+
+## Training locally 
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies. We also recommend to install `diffusers` from the `main` github branch.
+
+```bash
+pip install git+https://github.com/huggingface/diffusers
+pip install -U -r diffusers/examples/dreambooth/requirements.txt
+```
+
+Then initialize and configure a [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+
+You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
+
+Run the following command to authenticate your token
+
+```bash
+huggingface-cli login
+```
+
+If you have already cloned the repo, then you won't need to go through these steps. Instead, you can pass the path to your local checkout to the training script and it will be loaded from there.
+
+### Dog toy example
+
+In this example we'll use [these images](https://drive.google.com/drive/folders/1BO_dyz-p65qhBRRMRA4TbZ8qW4rB99JZ) to add a new concept to Stable Diffusion using the Dreambooth process. They will be our training data. Please, download them and place them somewhere in your system.
+
+Then you can launch the training script using:
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path_to_training_images"
+export OUTPUT_DIR="path_to_saved_model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=1 \
+  --learning_rate=5e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=400
+```
+
+### Training with a prior-preserving loss
+
+Prior preservation is used to avoid overfitting and language-drift. Please, refer to the paper to learn more about it if you are interested. For prior preservation, we use other images of the same class as part of the training process. The nice thing is that we can generate those images using the Stable Diffusion model itself! The training script will save the generated images to a local path we specify.
+
+According to the paper, it's recommended to generate `num_epochs * num_samples` images for prior preservation. 200-300 works well for most cases.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path_to_training_images"
+export CLASS_DIR="path_to_class_images"
+export OUTPUT_DIR="path_to_saved_model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=1 \
+  --learning_rate=5e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
+### Training on a 16GB GPU
+
+With the help of gradient checkpointing and the 8-bit optimizer from [bitsandbytes](https://github.com/TimDettmers/bitsandbytes), it's possible to train dreambooth on a 16GB GPU.
+
+```bash
+pip install bitsandbytes
+```
+
+Then pass the `--use_8bit_adam` option to the training script.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path_to_training_images"
+export CLASS_DIR="path_to_class_images"
+export OUTPUT_DIR="path_to_saved_model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=2 --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=5e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
+### Fine-tune the text encoder in addition to the UNet
+
+The script also allows to fine-tune the `text_encoder` along with the `unet`. It has been observed experimentally that this gives much better results, especially on faces. Please, refer to [our blog](https://huggingface.co/blog/dreambooth) for more details.
+
+To enable this option, pass the `--train_text_encoder` argument to the training script.
+
+<Tip>
+Training the text encoder requires additional memory, so training won't fit on a 16GB GPU. You'll need at least 24GB VRAM to use this option.
+</Tip>
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path_to_training_images"
+export CLASS_DIR="path_to_class_images"
+export OUTPUT_DIR="path_to_saved_model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --train_text_encoder \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --use_8bit_adam
+  --gradient_checkpointing \
+  --learning_rate=2e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
+### Training on a 8 GB GPU:
+
+Using [DeepSpeed](https://www.deepspeed.ai/) it's even possible to offload some
+tensors from VRAM to either CPU or NVME, allowing training to proceed with less GPU memory.
+
+DeepSpeed needs to be enabled with `accelerate config`. During configuration,
+answer yes to "Do you want to use DeepSpeed?". Combining DeepSpeed stage 2, fp16
+mixed precision, and offloading both the model parameters and the optimizer state to CPU, it's
+possible to train on under 8 GB VRAM. The drawback is that this requires more system RAM (about 25 GB). See [the DeepSpeed documentation](https://huggingface.co/docs/accelerate/usage_guides/deepspeed) for more configuration options.
+
+Changing the default Adam optimizer to DeepSpeed's special version of Adam
+`deepspeed.ops.adam.DeepSpeedCPUAdam` gives a substantial speedup, but enabling
+it requires the system's CUDA toolchain version to be the same as the one installed with PyTorch. 8-bit optimizers don't seem to be compatible with DeepSpeed at the moment.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path_to_training_images"
+export CLASS_DIR="path_to_class_images"
+export OUTPUT_DIR="path_to_saved_model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --sample_batch_size=1 \
+  --gradient_accumulation_steps=1 --gradient_checkpointing \
+  --learning_rate=5e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800 \
+  --mixed_precision=fp16
+```
+
+## Inference
+
+Once you have trained a model, inference can be done using the `StableDiffusionPipeline`, by simply indicating the path where the model was saved. Make sure that your prompts include the special `identifier` used during training (`sks` in the previous examples).
+
+```python
+from diffusers import StableDiffusionPipeline
+import torch
+
+model_id = "path_to_saved_model"
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+prompt = "A photo of sks dog in a bucket"
+image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
+
+image.save("dog-bucket.png")
+```

docs/source/training/overview.mdx

@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # 🧨 Diffusers Training Examples
 
-Diffusers examples are a collection of scripts to demonstrate how to effectively use the `diffusers` library
+Diffusers training examples are a collection of scripts to demonstrate how to effectively use the `diffusers` library
 for a variety of use cases.
 
 **Note**: If you are looking for **official** examples on how to use `diffusers` for inference, 
@@ -36,13 +36,15 @@ Training examples show how to pretrain or fine-tune diffusion models for a varie
 - [Unconditional Training](./unconditional_training)
 - [Text-to-Image Training](./text2image)
 - [Text Inversion](./text_inversion)
+- [Dreambooth](./dreambooth)
 
 
 | Task | 🤗 Accelerate | 🤗 Datasets | Colab
 |---|---|:---:|:---:|
 | [**Unconditional Image Generation**](./unconditional_training) | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
-| [**Text-to-Image**](./text2image) | - | - | 
-| [**Text-Inversion**](./text_inversion) | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
+| [**Text-to-Image fine-tuning**](./text2image) | ✅ | ✅ | 
+| [**Textual Inversion**](./text_inversion) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
+| [**Dreambooth**](./dreambooth) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_dreambooth_training.ipynb)
 
 ## Community
 

docs/source/training/text2image.mdx

@@ -11,6 +11,128 @@ specific language governing permissions and limitations under the License.
 -->
 
 
-# Text-to-Image Training
+# Stable Diffusion text-to-image fine-tuning
 
-Under construction 🚧
+The [`train_text_to_image.py`](https://github.com/huggingface/diffusers/tree/main/examples/text_to_image) script shows how to fine-tune the stable diffusion model on your own dataset.
+
+<Tip warning={true}>
+
+The text-to-image fine-tuning script is experimental. It's easy to overfit and run into issues like catastrophic forgetting. We recommend to explore different hyperparameters to get the best results on your dataset.
+
+</Tip>
+
+
+## Running locally 
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+```bash
+pip install git+https://github.com/huggingface/diffusers.git
+pip install -U -r requirements.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+
+You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
+
+Run the following command to authenticate your token
+
+```bash
+huggingface-cli login
+```
+
+If you have already cloned the repo, then you won't need to go through these steps. Instead, you can pass the path to your local checkout to the training script and it will be loaded from there.
+
+### Hardware Requirements for Fine-tuning
+
+Using `gradient_checkpointing` and `mixed_precision` it should be possible to fine tune the model on a single 24GB GPU. For higher `batch_size` and faster training it's better to use GPUs with more than 30GB of GPU memory. You can also use JAX / Flax for fine-tuning on TPUs or GPUs, see [below](#flax-jax-finetuning) for details.
+
+### Fine-tuning Example
+
+The following script will launch a fine-tuning run using [Justin Pinkneys' captioned Pokemon dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions), available in Hugging Face Hub.
+
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+accelerate launch train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir="sd-pokemon-model" 
+```
+
+To run on your own training files you need to prepare the dataset according to the format required by `datasets`. You can upload your dataset to the Hub, or you can prepare a local folder with your files. [This documentation](https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder-with-metadata) explains how to do it.
+
+You should modify the script if you wish to use custom loading logic. We have left pointers in the code in the appropriate places :)
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export TRAIN_DIR="path_to_your_dataset"
+export OUTPUT_DIR="path_to_save_model"
+
+accelerate launch train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --train_data_dir=$TRAIN_DIR \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir=${OUTPUT_DIR}
+```
+
+Once training is finished the model will be saved to the `OUTPUT_DIR` specified in the command. To load the fine-tuned model for inference, just pass that path to `StableDiffusionPipeline`:
+
+```python
+from diffusers import StableDiffusionPipeline
+
+model_path = "path_to_saved_model"
+pipe = StableDiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+pipe.to("cuda")
+
+image = pipe(prompt="yoda").images[0]
+image.save("yoda-pokemon.png")
+```
+
+### Flax / JAX fine-tuning
+
+Thanks to [@duongna211](https://github.com/duongna21) it's possible to fine-tune Stable Diffusion using Flax! This is very efficient on TPU hardware but works great on GPUs too. You can use the [Flax training script](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_flax.py) like this:
+
+```Python
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+python train_text_to_image_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --output_dir="sd-pokemon-model" 
+```

docs/source/training/text_inversion.mdx

@@ -49,7 +49,7 @@ The `textual_inversion.py` script [here](https://github.com/huggingface/diffuser
 
 ### Installing the dependencies
 
-Before running the scripts, make sure to install the library's training dependencies:
+Before running the scripts, make sure to install the library's training dependencies.
 
 ```bash
 pip install diffusers[training] accelerate transformers
@@ -64,7 +64,7 @@ accelerate config
 
 ### Cat toy example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree.
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -83,7 +83,7 @@ Now let's get our dataset.Download 3-4 images from [here](https://drive.google.c
 And launch the training using
 
 ```bash
-export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
 export DATA_DIR="path-to-dir-containing-images"
 
 accelerate launch textual_inversion.py \

docs/source/using-diffusers/audio.mdx

@@ -0,0 +1,16 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Using Diffusers for audio
+
+The [`DanceDiffusionPipeline`] can be used to generate audio rapidly!
+More coming soon!

docs/source/using-diffusers/conditional_image_generation.mdx

@@ -44,5 +44,3 @@ You can save the image by simply calling:
 ```python
 >>> image.save("image_of_squirrel_painting.png")
 ```
-
-

docs/source/using-diffusers/configuration.mdx

@@ -12,21 +12,10 @@ specific language governing permissions and limitations under the License.
 
 
 
-# Quicktour
+# Configuration
 
-Start using Diffusers🧨 quickly!
-To start, use the [`DiffusionPipeline`] for quick inference and sample generations!
-
-```
-pip install diffusers
-```
-
-## Main classes
-
-### Models
-
-### Schedulers
-
-### Pipelines
+The handling of configurations in Diffusers is with the `ConfigMixin` class.
 
+[[autodoc]] ConfigMixin
 
+Under further construction 🚧, open a [PR](https://github.com/huggingface/diffusers/compare) if you want to contribute!

docs/source/using-diffusers/contribute_pipeline.mdx

@@ -0,0 +1,169 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# How to build a community pipeline
+
+*Note*: this page was built from the GitHub Issue on Community Pipelines [#841](https://github.com/huggingface/diffusers/issues/841).
+
+Let's make an example!
+Say you want to define a pipeline that just does a single forward pass to a U-Net and then calls a scheduler only once (Note, this doesn't make any sense from a scientific point of view, but only represents an example of how things work under the hood).
+
+Cool! So you open your favorite IDE and start creating your pipeline 💻.
+First, what model weights and configurations do we need?
+We have a U-Net and a scheduler, so our pipeline should take a U-Net and a scheduler as an argument.
+Also, as stated above, you'd like to be able to load weights and the scheduler config for Hub and share your code with others, so we'll inherit from `DiffusionPipeline`:
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+
+class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+```
+
+Now, we must save the `unet` and `scheduler` in a config file so that you can save your pipeline with `save_pretrained`.
+Therefore, make sure you add every component that is save-able to the `register_modules` function:
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+
+class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+
+        self.register_modules(unet=unet, scheduler=scheduler)
+```
+
+Cool, the init is done! 🔥 Now, let's go into the forward pass, which we recommend defining as `__call__` . Here you're given all the creative freedom there is. For our amazing "one-step" pipeline, we simply create a random image and call the unet once and the scheduler once:
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+
+class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+
+        self.register_modules(unet=unet, scheduler=scheduler)
+
+    def __call__(self):
+        image = torch.randn(
+            (1, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size),
+        )
+        timestep = 1
+
+        model_output = self.unet(image, timestep).sample
+        scheduler_output = self.scheduler.step(model_output, timestep, image).prev_sample
+
+        return scheduler_output
+```
+
+Cool, that's it! 🚀 You can now run this pipeline by passing a `unet` and a `scheduler` to the init:
+
+```python
+from diffusers import DDPMScheduler, Unet2DModel
+
+scheduler = DDPMScheduler()
+unet = UNet2DModel()
+
+pipeline = UnetSchedulerOneForwardPipeline(unet=unet, scheduler=scheduler)
+
+output = pipeline()
+```
+
+But what's even better is that you can load pre-existing weights into the pipeline if they match exactly your pipeline structure. This is e.g. the case for [https://huggingface.co/google/ddpm-cifar10-32](https://huggingface.co/google/ddpm-cifar10-32) so that we can do the following:
+
+```python
+pipeline = UnetSchedulerOneForwardPipeline.from_pretrained("google/ddpm-cifar10-32")
+
+output = pipeline()
+```
+
+We want to share this amazing pipeline with the community, so we would open a PR request to add the following code under `one_step_unet.py` to [https://github.com/huggingface/diffusers/tree/main/examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) .
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+
+class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+
+        self.register_modules(unet=unet, scheduler=scheduler)
+
+    def __call__(self):
+        image = torch.randn(
+            (1, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size),
+        )
+        timestep = 1
+
+        model_output = self.unet(image, timestep).sample
+        scheduler_output = self.scheduler.step(model_output, timestep, image).prev_sample
+
+        return scheduler_output
+```
+
+Our amazing pipeline got merged here: [#840](https://github.com/huggingface/diffusers/pull/840).
+Now everybody that has `diffusers >= 0.4.0` installed can use our pipeline magically 🪄 as follows:
+
+```python
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
+pipe()
+```
+
+Another way to upload your custom_pipeline, besides sending a PR, is uploading the code that contains it to the Hugging Face Hub, [as exemplified here](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview#loading-custom-pipelines-from-the-hub).
+
+**Try it out now - it works!**
+
+In general, you will want to create much more sophisticated pipelines, so we recommend looking at existing pipelines here: [https://github.com/huggingface/diffusers/tree/main/examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community).
+
+IMPORTANT:
+You can use whatever package you want in your community pipeline file - as long as the user has it installed, everything will work fine. Make sure you have one and only one pipeline class that inherits from `DiffusionPipeline` as this will be automatically detected.
+
+## How do community pipelines work?
+A community pipeline is a class that has to inherit from ['DiffusionPipeline']:
+and that has been added to `examples/community` [files](https://github.com/huggingface/diffusers/tree/main/examples/community).
+The community can load the pipeline code via the custom_pipeline argument from DiffusionPipeline. See docs [here](https://huggingface.co/docs/diffusers/api/diffusion_pipeline#diffusers.DiffusionPipeline.from_pretrained.custom_pipeline):
+
+This means:
+The model weights and configs of the pipeline should be loaded from the `pretrained_model_name_or_path` [argument](https://huggingface.co/docs/diffusers/api/diffusion_pipeline#diffusers.DiffusionPipeline.from_pretrained.pretrained_model_name_or_path):
+whereas the code that powers the community pipeline is defined in a file added in [`examples/community`](https://github.com/huggingface/diffusers/tree/main/examples/community).
+
+Now, it might very well be that only some of your pipeline components weights can be downloaded from an official repo.
+The other components should then be passed directly to init as is the case for the ClIP guidance notebook [here](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb#scrollTo=z9Kglma6hjki).
+
+The magic behind all of this is that we load the code directly from GitHub. You can check it out in more detail if you follow the functionality defined here:
+
+```python
+# 2. Load the pipeline class, if using custom module then load it from the hub
+# if we load from explicit class, let's use it
+if custom_pipeline is not None:
+    pipeline_class = get_class_from_dynamic_module(
+        custom_pipeline, module_file=CUSTOM_PIPELINE_FILE_NAME, cache_dir=custom_pipeline
+    )
+elif cls != DiffusionPipeline:
+    pipeline_class = cls
+else:
+    diffusers_module = importlib.import_module(cls.__module__.split(".")[0])
+    pipeline_class = getattr(diffusers_module, config_dict["_class_name"])
+```
+
+This is why a community pipeline merged to GitHub will be directly available to all `diffusers` packages.
+

docs/source/using-diffusers/custom_pipeline_examples.mdx

@@ -0,0 +1,283 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Custom Pipelines
+
+> **For more information about community pipelines, please have a look at [this issue](https://github.com/huggingface/diffusers/issues/841).**
+
+**Community** examples consist of both inference and training examples that have been added by the community.
+Please have a look at the following table to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste ready code example that you can try out.
+If a community doesn't work as expected, please open an issue and ping the author on it.
+
+| Example                                | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                      | Colab                                                                                                                                                                                                              |                                                     Author |
+|:---------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------:|
+| CLIP Guided Stable Diffusion           | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)     | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |             [Suraj Patil](https://github.com/patil-suraj/) |
+| One Step U-Net (Dummy)                 | Example showcasing of how to use Community Pipelines (see https://github.com/huggingface/diffusers/issues/841)                                                                                                                                                                                                                                                                                                                                                                                           | [One Step U-Net](#one-step-unet)                                  | -                                                                                                                                                                                                                  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
+| Stable Diffusion Interpolation         | Interpolate the latent space of Stable Diffusion between different prompts/seeds                                                                                                                                                                                                                                                                                                                                                                                                                         | [Stable Diffusion Interpolation](#stable-diffusion-interpolation) | -                                                                                                                                                                                                                  |                    [Nate Raw](https://github.com/nateraw/) |
+| Stable Diffusion Mega                  | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega)                   | -                                                                                                                                                                                                                  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
+| Long Prompt Weighting Stable Diffusion | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [SkyTNT](https://github.com/SkyTNT) |
+| Speech to Image                        | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                               | -                                                                                                                                                                                                                  | [Mikail Duzenli](https://github.com/MikailINTech)
+
+To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
+```py
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4", custom_pipeline="filename_in_the_community_folder"
+)
+```
+
+## Example usages
+
+### CLIP Guided Stable Diffusion
+
+CLIP guided stable diffusion can help to generate more realistic images
+by guiding stable diffusion at every denoising step with an additional CLIP model.
+
+The following code requires roughly 12GB of GPU RAM.
+
+```python
+from diffusers import DiffusionPipeline
+from transformers import CLIPFeatureExtractor, CLIPModel
+import torch
+
+
+feature_extractor = CLIPFeatureExtractor.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K")
+clip_model = CLIPModel.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K", torch_dtype=torch.float16)
+
+
+guided_pipeline = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="clip_guided_stable_diffusion",
+    clip_model=clip_model,
+    feature_extractor=feature_extractor,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+guided_pipeline.enable_attention_slicing()
+guided_pipeline = guided_pipeline.to("cuda")
+
+prompt = "fantasy book cover, full moon, fantasy forest landscape, golden vector elements, fantasy magic, dark light night, intricate, elegant, sharp focus, illustration, highly detailed, digital painting, concept art, matte, art by WLOP and Artgerm and Albert Bierstadt, masterpiece"
+
+generator = torch.Generator(device="cuda").manual_seed(0)
+images = []
+for i in range(4):
+    image = guided_pipeline(
+        prompt,
+        num_inference_steps=50,
+        guidance_scale=7.5,
+        clip_guidance_scale=100,
+        num_cutouts=4,
+        use_cutouts=False,
+        generator=generator,
+    ).images[0]
+    images.append(image)
+
+# save images locally
+for i, img in enumerate(images):
+    img.save(f"./clip_guided_sd/image_{i}.png")
+```
+
+The `images` list contains a list of PIL images that can be saved locally or displayed directly in a google colab.
+Generated images tend to be of higher qualtiy than natively using stable diffusion. E.g. the above script generates the following images:
+
+![clip_guidance](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/clip_guidance/merged_clip_guidance.jpg).
+
+### One Step Unet
+
+The dummy "one-step-unet" can be run as follows:
+
+```python
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
+pipe()
+```
+
+**Note**: This community pipeline is not useful as a feature, but rather just serves as an example of how community pipelines can be added (see https://github.com/huggingface/diffusers/issues/841).
+
+### Stable Diffusion Interpolation
+
+The following code can be run on a GPU of at least 8GB VRAM and should take approximately 5 minutes.
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    revision="fp16",
+    torch_dtype=torch.float16,
+    safety_checker=None,  # Very important for videos...lots of false positives while interpolating
+    custom_pipeline="interpolate_stable_diffusion",
+).to("cuda")
+pipe.enable_attention_slicing()
+
+frame_filepaths = pipe.walk(
+    prompts=["a dog", "a cat", "a horse"],
+    seeds=[42, 1337, 1234],
+    num_interpolation_steps=16,
+    output_dir="./dreams",
+    batch_size=4,
+    height=512,
+    width=512,
+    guidance_scale=8.5,
+    num_inference_steps=50,
+)
+```
+
+The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion.
+
+> **Please have a look at https://github.com/nateraw/stable-diffusion-videos for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
+
+### Stable Diffusion Mega
+
+The Stable Diffusion Mega Pipeline lets you use the main use cases of the stable diffusion pipeline in a single class.
+
+```python
+#!/usr/bin/env python3
+from diffusers import DiffusionPipeline
+import PIL
+import requests
+from io import BytesIO
+import torch
+
+
+def download_image(url):
+    response = requests.get(url)
+    return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="stable_diffusion_mega",
+    torch_dtype=torch.float16,
+    revision="fp16",
+)
+pipe.to("cuda")
+pipe.enable_attention_slicing()
+
+
+### Text-to-Image
+
+images = pipe.text2img("An astronaut riding a horse").images
+
+### Image-to-Image
+
+init_image = download_image(
+    "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+)
+
+prompt = "A fantasy landscape, trending on artstation"
+
+images = pipe.img2img(prompt=prompt, init_image=init_image, strength=0.75, guidance_scale=7.5).images
+
+### Inpainting
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+init_image = download_image(img_url).resize((512, 512))
+mask_image = download_image(mask_url).resize((512, 512))
+
+prompt = "a cat sitting on a bench"
+images = pipe.inpaint(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images
+```
+
+As shown above this one pipeline can run all both "text-to-image", "image-to-image", and "inpainting" in one pipeline.
+
+### Long Prompt Weighting Stable Diffusion
+
+The Pipeline lets you input prompt without 77 token length limit. And you can increase words weighting by using "()" or decrease words weighting by using "[]"
+The Pipeline also lets you use the main use cases of the stable diffusion pipeline in a single class.
+
+#### pytorch
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    "hakurei/waifu-diffusion", custom_pipeline="lpw_stable_diffusion", revision="fp16", torch_dtype=torch.float16
+)
+pipe = pipe.to("cuda")
+
+prompt = "best_quality (1girl:1.3) bow bride brown_hair closed_mouth frilled_bow frilled_hair_tubes frills (full_body:1.3) fox_ear hair_bow hair_tubes happy hood japanese_clothes kimono long_sleeves red_bow smile solo tabi uchikake white_kimono wide_sleeves cherry_blossoms"
+neg_prompt = "lowres, bad_anatomy, error_body, error_hair, error_arm, error_hands, bad_hands, error_fingers, bad_fingers, missing_fingers, error_legs, bad_legs, multiple_legs, missing_legs, error_lighting, error_shadow, error_reflection, text, error, extra_digit, fewer_digits, cropped, worst_quality, low_quality, normal_quality, jpeg_artifacts, signature, watermark, username, blurry"
+
+pipe.text2img(prompt, negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
+```
+
+#### onnxruntime
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="lpw_stable_diffusion_onnx",
+    revision="onnx",
+    provider="CUDAExecutionProvider",
+)
+
+prompt = "a photo of an astronaut riding a horse on mars, best quality"
+neg_prompt = "lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
+
+pipe.text2img(prompt, negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
+```
+
+if you see `Token indices sequence length is longer than the specified maximum sequence length for this model ( *** > 77 ) . Running this sequence through the model will result in indexing errors`. Do not worry, it is normal.
+
+### Speech to Image
+
+The following code can generate an image from an audio sample using pre-trained OpenAI whisper-small and Stable Diffusion.
+
+```Python
+import torch
+
+import matplotlib.pyplot as plt
+from datasets import load_dataset
+from diffusers import DiffusionPipeline
+from transformers import (
+    WhisperForConditionalGeneration,
+    WhisperProcessor,
+)
+
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+audio_sample = ds[3]
+
+text = audio_sample["text"].lower()
+speech_data = audio_sample["audio"]["array"]
+
+model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small").to(device)
+processor = WhisperProcessor.from_pretrained("openai/whisper-small")
+
+diffuser_pipeline = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="speech_to_image_diffusion",
+    speech_model=model,
+    speech_processor=processor,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+
+diffuser_pipeline.enable_attention_slicing()
+diffuser_pipeline = diffuser_pipeline.to(device)
+
+output = diffuser_pipeline(speech_data)
+plt.imshow(output.images[0])
+```
+This example produces the following image:
+
+![image](https://user-images.githubusercontent.com/45072645/196901736-77d9c6fc-63ee-4072-90b0-dc8b903d63e3.png)

docs/source/using-diffusers/custom_pipeline_overview.mdx

@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Custom Pipelines
+# Loading and Adding Custom Pipelines
 
 Diffusers allows you to conveniently load any custom pipeline from the Hugging Face Hub as well as any [official community pipeline](https://github.com/huggingface/diffusers/tree/main/examples/community) 
 via the [`DiffusionPipeline`] class.
@@ -58,7 +58,7 @@ feature_extractor = CLIPFeatureExtractor.from_pretrained(clip_model_id)
 clip_model = CLIPModel.from_pretrained(clip_model_id)
 
 pipeline = DiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
+    "runwayml/stable-diffusion-v1-5",
     custom_pipeline="clip_guided_stable_diffusion",
     clip_model=clip_model,
     feature_extractor=feature_extractor,
@@ -91,24 +91,24 @@ class MyPipeline(DiffusionPipeline):
         # Sample gaussian noise to begin loop
         image = torch.randn((batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size))
 
-    image = image.to(self.device)
+        image = image.to(self.device)
 
-    # set step values
-    self.scheduler.set_timesteps(num_inference_steps)
+        # set step values
+        self.scheduler.set_timesteps(num_inference_steps)
 
-    for t in self.progress_bar(self.scheduler.timesteps):
-        # 1. predict noise model_output
-        model_output = self.unet(image, t).sample
+        for t in self.progress_bar(self.scheduler.timesteps):
+            # 1. predict noise model_output
+            model_output = self.unet(image, t).sample
 
-        # 2. predict previous mean of image x_t-1 and add variance depending on eta
-        # eta corresponds to η in paper and should be between [0, 1]
-        # do x_t -> x_t-1
-        image = self.scheduler.step(model_output, t, image, eta).prev_sample
+            # 2. predict previous mean of image x_t-1 and add variance depending on eta
+            # eta corresponds to η in paper and should be between [0, 1]
+            # do x_t -> x_t-1
+            image = self.scheduler.step(model_output, t, image, eta).prev_sample
 
-    image = (image / 2 + 0.5).clamp(0, 1)
-    image = image.cpu().permute(0, 2, 3, 1).numpy()
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).numpy()
 
-    return image
+        return image
 ```
 
 Now you can upload this short file under the name `pipeline.py` in your preferred [model repository](https://huggingface.co/docs/hub/models-uploading). For Stable Diffusion pipelines, you may also [join the community organisation for shared pipelines](https://huggingface.co/organizations/sd-diffusers-pipelines-library/share/BUPyDUuHcciGTOKaExlqtfFcyCZsVFdrjr) to upload yours.

docs/source/using-diffusers/img2img.mdx

@@ -15,6 +15,7 @@ specific language governing permissions and limitations under the License.
 The [`StableDiffusionImg2ImgPipeline`] lets you pass a text prompt and an initial image to condition the generation of new images.
 
 ```python
+import torch
 import requests
 from PIL import Image
 from io import BytesIO
@@ -24,7 +25,7 @@ from diffusers import StableDiffusionImg2ImgPipeline
 # load the pipeline
 device = "cuda"
 pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16
+    "runwayml/stable-diffusion-v1-5", revision="fp16", torch_dtype=torch.float16
 ).to(device)
 
 # let's download an initial image
@@ -32,7 +33,7 @@ url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/st
 
 response = requests.get(url)
 init_image = Image.open(BytesIO(response.content)).convert("RGB")
-init_image = init_image.resize((768, 512))
+init_image.thumbnail((768, 768))
 
 prompt = "A fantasy landscape, trending on artstation"
 

docs/source/using-diffusers/inpaint.mdx

@@ -12,13 +12,19 @@ specific language governing permissions and limitations under the License.
 
 # Text-Guided Image-Inpainting
 
-The [`StableDiffusionInpaintPipeline`] lets you edit specific parts of an image by providing a mask and text prompt.
+The [`StableDiffusionInpaintPipeline`] lets you edit specific parts of an image by providing a mask and a text prompt. It uses a version of Stable Diffusion specifically trained for in-painting tasks.
+
+<Tip warning={true}>
+Note that this model is distributed separately from the regular Stable Diffusion model, so you have to accept its license even if you accepted the Stable Diffusion one in the past.
+
+Please, visit the [model card](https://huggingface.co/runwayml/stable-diffusion-inpainting), read the license carefully and tick the checkbox if you agree. You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section](https://huggingface.co/docs/hub/security-tokens) of the documentation.
+</Tip>
 
 ```python
-from io import BytesIO
-
-import requests
 import PIL
+import requests
+import torch
+from io import BytesIO
 
 from diffusers import StableDiffusionInpaintPipeline
 
@@ -34,15 +40,24 @@ mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data
 init_image = download_image(img_url).resize((512, 512))
 mask_image = download_image(mask_url).resize((512, 512))
 
-device = "cuda"
 pipe = StableDiffusionInpaintPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16
-).to(device)
+    "runwayml/stable-diffusion-inpainting",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
 
-prompt = "a cat sitting on a bench"
-images = pipe(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images
-
-images[0].save("cat_on_bench.png")
+prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
 ```
 
+`image`          | `mask_image` | `prompt` | **Output** |
+:-------------------------:|:-------------------------:|:-------------------------:|-------------------------:|
+<img src="https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" alt="drawing" width="250"/> | <img src="https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" alt="drawing" width="250"/> | ***Face of a yellow cat, high resolution, sitting on a park bench*** | <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/test.png" alt="drawing" width="250"/> |
+
+
 You can also run this example on colab [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
+
+<Tip warning={true}>
+A previous experimental implementation of in-painting used a different, lower-quality process. To ensure backwards compatibility, loading a pretrained pipeline that doesn't contain the new model will still apply the old in-painting method.
+</Tip>

docs/source/using-diffusers/loading.mdx

@@ -10,6 +10,389 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Loading 
+# Loading
 
-Under construction 🚧
+A core premise of the diffusers library is to make diffusion models **as accessible as possible**.
+Accessibility is therefore achieved by providing an API to load complete diffusion pipelines as well as individual components with a single line of code. 
+
+In the following we explain in-detail how to easily load:
+
+- *Complete Diffusion Pipelines* via the [`DiffusionPipeline.from_pretrained`]
+- *Diffusion Models* via [`ModelMixin.from_pretrained`]
+- *Schedulers* via [`SchedulerMixin.from_pretrained`]
+
+## Loading pipelines
+
+The [`DiffusionPipeline`] class is the easiest way to access any diffusion model that is [available on the Hub](https://huggingface.co/models?library=diffusers). Let's look at an example on how to download [CompVis' Latent Diffusion model](https://huggingface.co/CompVis/ldm-text2im-large-256).
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "CompVis/ldm-text2im-large-256"
+ldm = DiffusionPipeline.from_pretrained(repo_id)
+```
+
+Here [`DiffusionPipeline`] automatically detects the correct pipeline (*i.e.* [`LDMTextToImagePipeline`]), downloads and caches all required configuration and weight files (if not already done so), and finally returns a pipeline instance, called `ldm`.
+The pipeline instance can then be called using [`LDMTextToImagePipeline.__call__`] (i.e., `ldm("image of a astronaut riding a horse")`) for text-to-image generation.
+
+Instead of using the generic [`DiffusionPipeline`] class for loading, you can also load the appropriate pipeline class directly. The code snippet above yields the same instance as when doing:
+
+```python
+from diffusers import LDMTextToImagePipeline
+
+repo_id = "CompVis/ldm-text2im-large-256"
+ldm = LDMTextToImagePipeline.from_pretrained(repo_id)
+```
+
+Diffusion pipelines like `LDMTextToImagePipeline` often consist of multiple components. These components can be both parameterized models, such as `"unet"`, `"vqvae"` and "bert", tokenizers or schedulers. These components can interact in complex ways with each other when using the pipeline in inference, *e.g.* for [`LDMTextToImagePipeline`] or [`StableDiffusionPipeline`] the inference call is explained [here](https://huggingface.co/blog/stable_diffusion#how-does-stable-diffusion-work).
+The purpose of the [pipeline classes](./api/overview#diffusers-summary) is to wrap the complexity of these diffusion systems and give the user an easy-to-use API while staying flexible for customization, as will be shown later.
+
+### Loading pipelines that require access request
+
+Due to the capabilities of diffusion models to generate extremely realistic images, there is a certain danger that such models might be misused for unwanted applications, *e.g.* generating pornography or violent images.
+In order to minimize the possibility of such unsolicited use cases, some of the most powerful diffusion models require users to acknowledge a license before being able to use the model. If the user does not agree to the license, the pipeline cannot be downloaded.
+If you try to load [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5) the same way as done previously:
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "runwayml/stable-diffusion-v1-5"
+stable_diffusion = DiffusionPipeline.from_pretrained(repo_id)
+```
+
+it will only work if you have both *click-accepted* the license on [the model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) and are logged into the Hugging Face Hub. Otherwise you will get an error message
+such as the following:
+
+```
+OSError: runwayml/stable-diffusion-v1-5 is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'
+If this is a private repository, make sure to pass a token having permission to this repo with `use_auth_token` or log in with `huggingface-cli login`
+```
+
+Therefore, we need to make sure to *click-accept* the license. You can do this by simply visiting 
+the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) and clicking on "Agree and access repository":
+
+<p align="center">
+    <br>
+    <img src="https://raw.githubusercontent.com/huggingface/diffusers/main/docs/source/imgs/access_request.png" width="400"/>
+    <br>
+</p>
+
+Second, you need to login with your access token:
+
+```
+huggingface-cli login
+```
+
+before trying to load the model. Or alternatively, you can pass [your access token](https://huggingface.co/docs/hub/security-tokens#user-access-tokens) directly via the flag `use_auth_token`. In this case you do **not** need 
+to run `huggingface-cli login` before:
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "runwayml/stable-diffusion-v1-5"
+stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, use_auth_token="<your-access-token>")
+```
+
+The final option to use pipelines that require access without having to rely on the Hugging Face Hub is to load the pipeline locally as explained in the next section.
+
+### Loading pipelines locally
+
+If you prefer to have complete control over the pipeline and its corresponding files or, as said before, if you want to use pipelines that require an access request without having to be connected to the Hugging Face Hub, 
+we recommend loading pipelines locally.
+
+To load a diffusion pipeline locally, you first need to manually download the whole folder structure on your local disk and then pass a local path to the [`DiffusionPipeline.from_pretrained`]. Let's again look at an example for 
+[CompVis' Latent Diffusion model](https://huggingface.co/CompVis/ldm-text2im-large-256).
+
+First, you should make use of [`git-lfs`](https://git-lfs.github.com/) to download the whole folder structure that has been uploaded to the [model repository](https://huggingface.co/CompVis/ldm-text2im-large-256/tree/main):
+
+```
+git lfs install
+git clone https://huggingface.co/runwayml/stable-diffusion-v1-5
+```
+
+The command above will create a local folder called `./stable-diffusion-v1-5` on your disk.
+Now, all you have to do is to simply pass the local folder path to `from_pretrained`:
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "./stable-diffusion-v1-5"
+stable_diffusion = DiffusionPipeline.from_pretrained(repo_id)
+```
+
+If `repo_id` is a local path, as it is the case here, [`DiffusionPipeline.from_pretrained`] will automatically detect it and therefore not try to download any files from the Hub.
+While we usually recommend to load weights directly from the Hub to be certain to stay up to date with the newest changes, loading pipelines locally should be preferred if one 
+wants to stay anonymous, self-contained applications, etc...
+
+### Loading customized pipelines
+
+Advanced users that want to load customized versions of diffusion pipelines can do so by swapping any of the default components, *e.g.* the scheduler, with other scheduler classes.
+A classical use case of this functionality is to swap the scheduler. [Stable Diffusion v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5) uses the [`PNDMScheduler`] by default which is generally not the most performant scheduler. Since the release
+of stable diffusion, multiple improved schedulers have been published. To use those, the user has to manually load their preferred scheduler and pass it into [`DiffusionPipeline.from_pretrained`].
+
+*E.g.* to use [`EulerDiscreteScheduler`] or [`DPMSolverMultistepScheduler`] to have a better quality vs. generation speed trade-off for inference, one could load them as follows:
+
+```python
+from diffusers import DiffusionPipeline, EulerDiscreteScheduler, DPMSolverMultistepScheduler
+
+repo_id = "runwayml/stable-diffusion-v1-5"
+
+scheduler = EulerDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
+# or
+# scheduler = DPMSolverMultistepScheduler.from_pretrained(repo_id, subfolder="scheduler")
+
+stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, scheduler=scheduler)
+```
+
+Three things are worth paying attention to here.
+- First, the scheduler is loaded with [`SchedulerMixin.from_pretrained`]
+- Second, the scheduler is loaded with a function argument, called `subfolder="scheduler"` as the configuration of stable diffusion's scheduling is defined in a [subfolder of the official pipeline repository](https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main/scheduler)
+- Third, the scheduler instance can simply be passed with the `scheduler` keyword argument to [`DiffusionPipeline.from_pretrained`]. This works because the [`StableDiffusionPipeline`] defines its scheduler with the `scheduler` attribute. It's not possible to use a different name, such as `sampler=scheduler` since `sampler` is not a defined keyword for [`StableDiffusionPipeline.__init__`]
+
+Not only the scheduler components can be customized for diffusion pipelines; in theory, all components of a pipeline can be customized. In practice, however, it often only makes sense to switch out a component that has **compatible** alternatives to what the pipeline expects.
+Many scheduler classes are compatible with each other as can be seen [here](https://github.com/huggingface/diffusers/blob/0dd8c6b4dbab4069de9ed1cafb53cbd495873879/src/diffusers/schedulers/scheduling_ddim.py#L112). This is not always the case for other components, such as the `"unet"`.
+
+One special case that can also be customized is the `"safety_checker"` of stable diffusion. If you believe the safety checker doesn't serve you any good, you can simply disable it by passing `None`:
+
+```python
+from diffusers import DiffusionPipeline, EulerDiscreteScheduler, DPMSolverMultistepScheduler
+
+stable_diffusion = DiffusionPipeline.from_pretrained(repo_id, safety_checker=None)
+```
+
+Another common use case is to reuse the same components in multiple pipelines, *e.g.* the weights and configurations of [`"runwayml/stable-diffusion-v1-5"`](https://huggingface.co/runwayml/stable-diffusion-v1-5) can be used for both [`StableDiffusionPipeline`] and [`StableDiffusionImg2ImgPipeline`] and we might not want to 
+use the exact same weights into RAM twice. In this case, customizing all the input instances would help us 
+to only load the weights into RAM once:
+
+```python
+from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
+
+model_id = "runwayml/stable-diffusion-v1-5"
+stable_diffusion_txt2img = StableDiffusionPipeline.from_pretrained(model_id)
+
+components = stable_diffusion_txt2img.components
+
+# weights are not reloaded into RAM
+stable_diffusion_img2img = StableDiffusionImg2ImgPipeline(**components)
+```
+
+Note how the above code snippet makes use of [`DiffusionPipeline.components`].
+
+### How does loading work?
+
+As a class method, [`DiffusionPipeline.from_pretrained`] is responsible for two things:
+- Download the latest version of the folder structure required to run the `repo_id` with `diffusers` and cache them. If the latest folder structure is available in the local cache, [`DiffusionPipeline.from_pretrained`] will simply reuse the cache and **not** re-download the files.
+- Load the cached weights into the _correct_ pipeline class – one of the [officially supported pipeline classes](./api/overview#diffusers-summary) - and return an instance of the class. The _correct_ pipeline class is thereby retrieved from the `model_index.json` file.
+
+The underlying folder structure of diffusion pipelines correspond 1-to-1 to their corresponding class instances, *e.g.* [`LDMTextToImagePipeline`] for [`CompVis/ldm-text2im-large-256`](https://huggingface.co/CompVis/ldm-text2im-large-256)
+This can be understood better by looking at an example. Let's print out pipeline class instance `pipeline` we just defined:
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "CompVis/ldm-text2im-large-256"
+ldm = DiffusionPipeline.from_pretrained(repo_id)
+print(ldm)
+```
+
+*Output*:
+```
+LDMTextToImagePipeline {
+  "bert": [
+    "latent_diffusion",
+    "LDMBertModel"
+  ],
+  "scheduler": [
+    "diffusers",
+    "DDIMScheduler"
+  ],
+  "tokenizer": [
+    "transformers",
+    "BertTokenizer"
+  ],
+  "unet": [
+    "diffusers",
+    "UNet2DConditionModel"
+  ],
+  "vqvae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}
+```
+
+First, we see that the official pipeline is the [`LDMTextToImagePipeline`], and second we see that the `LDMTextToImagePipeline` consists of 5 components:
+- `"bert"` of class `LDMBertModel` as defined [in the pipeline](https://github.com/huggingface/diffusers/blob/cd502b25cf0debac6f98d27a6638ef95208d1ea2/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py#L664)
+- `"scheduler"` of class [`DDIMScheduler`]
+- `"tokenizer"` of class `BertTokenizer` as defined [in `transformers`](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer)
+- `"unet"` of class [`UNet2DConditionModel`]
+- `"vqvae"` of class [`AutoencoderKL`]
+
+Let's now compare the pipeline instance to the folder structure of the model repository `CompVis/ldm-text2im-large-256`. Looking at the folder structure of [`CompVis/ldm-text2im-large-256`](https://huggingface.co/CompVis/ldm-text2im-large-256/tree/main) on the Hub, we can see it matches 1-to-1 the printed out instance of `LDMTextToImagePipeline` above:
+
+```
+.
+├── bert
+│   ├── config.json
+│   └── pytorch_model.bin
+├── model_index.json
+├── scheduler
+│   └── scheduler_config.json
+├── tokenizer
+│   ├── special_tokens_map.json
+│   ├── tokenizer_config.json
+│   └── vocab.txt
+├── unet
+│   ├── config.json
+│   └── diffusion_pytorch_model.bin
+└── vqvae
+    ├── config.json
+    └── diffusion_pytorch_model.bin
+```
+
+As we can see each attribute of the instance of `LDMTextToImagePipeline` has its configuration and possibly weights defined in a subfolder that is called **exactly** like the class attribute (`"bert"`, `"scheduler"`, `"tokenizer"`, `"unet"`, `"vqvae"`). Importantly, every pipeline expects a `model_index.json` file that tells the `DiffusionPipeline` both:
+- which pipeline class should be loaded, and
+- what sub-classes from which library are stored in which subfolders
+
+In the case of `CompVis/ldm-text2im-large-256` the `model_index.json` is therefore defined as follows:
+
+```
+{
+  "_class_name": "LDMTextToImagePipeline",
+  "_diffusers_version": "0.0.4",
+  "bert": [
+    "latent_diffusion",
+    "LDMBertModel"
+  ],
+  "scheduler": [
+    "diffusers",
+    "DDIMScheduler"
+  ],
+  "tokenizer": [
+    "transformers",
+    "BertTokenizer"
+  ],
+  "unet": [
+    "diffusers",
+    "UNet2DConditionModel"
+  ],
+  "vqvae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}
+```
+
+- `_class_name` tells `DiffusionPipeline` which pipeline class should be loaded. 
+- `_diffusers_version` can be useful to know under which `diffusers` version this model was created.
+- Every component of the pipeline is then defined under the form:
+```
+"name" : [
+  "library",
+  "class"
+]
+```
+	- The `"name"` field corresponds both to the name of the subfolder in which the configuration and weights are stored as well as the attribute name of the pipeline class (as can be seen [here](https://huggingface.co/CompVis/ldm-text2im-large-256/tree/main/bert) and [here](https://github.com/huggingface/diffusers/blob/cd502b25cf0debac6f98d27a6638ef95208d1ea2/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py#L42)
+	- The `"library"` field corresponds to the name of the library, *e.g.* `diffusers` or `transformers` from which the `"class"` should be loaded
+	- The `"class"` field corresponds to the name of the class, *e.g.* [`BertTokenizer`](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer) or [`UNet2DConditionModel`]
+
+
+## Loading models
+
+Models as defined under [src/diffusers/models](https://github.com/huggingface/diffusers/tree/main/src/diffusers/models) can be loaded via the [`ModelMixin.from_pretrained`] function. The API is very similar the [`DiffusionPipeline.from_pretrained`] and works in the same way:
+- Download the latest version of the model weights and configuration with `diffusers` and cache them. If the latest files are available in the local cache, [`ModelMixin.from_pretrained`] will simply reuse the cache and **not** re-download the files.
+- Load the cached weights into the _defined_ model class - one of [the existing model classes](./api/models) - and return an instance of the class.
+
+In constrast to [`DiffusionPipeline.from_pretrained`], models rely on fewer files that usually don't require a folder structure, but just a `diffusion_pytorch_model.bin` and `config.json` file.
+
+Let's look at an example:
+
+```python
+from diffusers import UNet2DConditionModel
+
+repo_id = "CompVis/ldm-text2im-large-256"
+model = UNet2DConditionModel.from_pretrained(repo_id, subfolder="unet")
+```
+
+Note how we have to define the `subfolder="unet"` argument to tell [`ModelMixin.from_pretrained`] that the model weights are located in a [subfolder of the repository](https://huggingface.co/CompVis/ldm-text2im-large-256/tree/main/unet).
+
+As explained in [Loading customized pipelines]("./using-diffusers/loading#loading-customized-pipelines"), one can pass a loaded model to a diffusion pipeline, via [`DiffusionPipeline.from_pretrained`]:
+
+```python
+from diffusers import DiffusionPipeline
+
+repo_id = "CompVis/ldm-text2im-large-256"
+ldm = DiffusionPipeline.from_pretrained(repo_id, unet=model)
+```
+
+If the model files can be found directly at the root level, which is usually only the case for some very simple diffusion models, such as [`google/ddpm-cifar10-32`](https://huggingface.co/google/ddpm-cifar10-32), we don't 
+need to pass a `subfolder` argument:
+
+```python
+from diffusers import UNet2DModel
+
+repo_id = "google/ddpm-cifar10-32"
+model = UNet2DModel.from_pretrained(repo_id)
+```
+
+## Loading schedulers
+
+Schedulers rely on [`SchedulerMixin.from_pretrained`]. Schedulers are **not parameterized** or **trained**, but instead purely defined by a configuration file.
+For consistency, we use the same method name as we do for models or pipelines, but no weights are loaded in this case.
+
+In constrast to pipelines or models, loading schedulers does not consume any significant amount of memory and the same configuration file can often be used for a variety of different schedulers.
+For example, all of:
+
+- [`DDPMScheduler`]
+- [`DDIMScheduler`]
+- [`PNDMScheduler`]
+- [`LMSDiscreteScheduler`]
+- [`EulerDiscreteScheduler`]
+- [`EulerAncestralDiscreteScheduler`]
+- [`DPMSolverMultistepScheduler`]
+
+are compatible with [`StableDiffusionPipeline`] and therefore the same scheduler configuration file can be loaded in any of those classes:
+
+```python
+from diffusers import StableDiffusionPipeline
+from diffusers import (
+    DDPMScheduler,
+    DDIMScheduler,
+    PNDMScheduler,
+    LMSDiscreteScheduler,
+    EulerDiscreteScheduler,
+    EulerAncestralDiscreteScheduler,
+    DPMSolverMultistepScheduler,
+)
+
+repo_id = "runwayml/stable-diffusion-v1-5"
+
+ddpm = DDPMScheduler.from_pretrained(repo_id, subfolder="scheduler")
+ddim = DDIMScheduler.from_pretrained(repo_id, subfolder="scheduler")
+pndm = PNDMScheduler.from_pretrained(repo_id, subfolder="scheduler")
+lms = LMSDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
+euler_anc = EulerAncestralDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
+euler = EulerDiscreteScheduler.from_pretrained(repo_id, subfolder="scheduler")
+dpm = DPMSolverMultistepScheduler.from_pretrained(repo_id, subfolder="scheduler")
+
+# replace `dpm` with any of `ddpm`, `ddim`, `pndm`, `lms`, `euler`, `euler_anc`
+pipeline = StableDiffusionPipeline.from_pretrained(repo_id, scheduler=dpm)
+```
+
+## API
+
+[[autodoc]] modeling_utils.ModelMixin
+    - from_pretrained
+    - save_pretrained
+
+[[autodoc]] pipeline_utils.DiffusionPipeline
+    - from_pretrained
+    - save_pretrained
+
+[[autodoc]] modeling_flax_utils.FlaxModelMixin
+    - from_pretrained
+    - save_pretrained
+
+[[autodoc]] pipeline_flax_utils.FlaxDiffusionPipeline
+    - from_pretrained
+    - save_pretrained

docs/source/using-diffusers/other-modalities.mdx

@@ -0,0 +1,20 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Using Diffusers with other modalities
+
+Diffusers is in the process of expanding to modalities other than images.
+
+Currently, one example is for [molecule conformation](https://www.nature.com/subjects/molecular-conformation#:~:text=Definition,to%20changes%20in%20their%20environment.) generation.
+* Generate conformations in Colab [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/geodiff_molecule_conformation.ipynb)
+
+More coming soon!

docs/source/using-diffusers/rl.mdx

@@ -0,0 +1,18 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Using Diffusers for reinforcement learning
+
+Support for one RL model and related pipelines is included in the `experimental` source of diffusers.
+
+To try some of this in colab, please look at the following example:
+* Model-based reinforcement learning on Colab [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/reinforcement_learning_with_diffusers.ipynb) ![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)

docs/source/using-diffusers/schedulers.mdx

@@ -0,0 +1,262 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Schedulers
+
+Diffusion pipelines are inherently a collection of diffusion models and schedulers that are partly independent from each other. This means that one is able to switch out parts of the pipeline to better customize 
+a pipeline to one's use case. The best example of this are the [Schedulers](../api/schedulers.mdx).
+
+Whereas diffusion models usually simply define the forward pass from noise to a less noisy sample, 
+schedulers define the whole denoising process, *i.e.*:
+- How many denoising steps?
+- Stochastic or deterministic?
+- What algorithm to use to find the denoised sample
+
+They can be quite complex and often define a trade-off between **denoising speed** and **denoising quality**.
+It is extremely difficult to measure quantitatively which scheduler works best for a given diffusion pipeline, so it is often recommended to simply try out which works best.
+
+The following paragraphs shows how to do so with the 🧨 Diffusers library.
+
+## Load pipeline
+
+Let's start by loading the stable diffusion pipeline.
+Remember that you have to be a registered user on the 🤗 Hugging Face Hub, and have "click-accepted" the [license](https://huggingface.co/runwayml/stable-diffusion-v1-5) in order to use stable diffusion.
+
+```python
+from huggingface_hub import login
+from diffusers import DiffusionPipeline
+import torch
+
+# first we need to login with our access token
+login()
+
+# Now we can download the pipeline
+pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
+```
+
+Next, we move it to GPU:
+
+```python
+pipeline.to("cuda")
+```
+
+## Access the scheduler
+
+The scheduler is always one of the components of the pipeline and is usually called `"scheduler"`.
+So it can be accessed via the `"scheduler"` property.
+
+```python
+pipeline.scheduler
+```
+
+**Output**:
+```
+PNDMScheduler {
+  "_class_name": "PNDMScheduler",
+  "_diffusers_version": "0.8.0.dev0",
+  "beta_end": 0.012,
+  "beta_schedule": "scaled_linear",
+  "beta_start": 0.00085,
+  "clip_sample": false,
+  "num_train_timesteps": 1000,
+  "set_alpha_to_one": false,
+  "skip_prk_steps": true,
+  "steps_offset": 1,
+  "trained_betas": null
+}
+```
+
+We can see that the scheduler is of type [`PNDMScheduler`]. 
+Cool, now let's compare the scheduler in its performance to other schedulers.
+First we define a prompt on which we will test all the different schedulers:
+
+```python
+prompt = "A photograph of an astronaut riding a horse on Mars, high resolution, high definition."
+```
+
+Next, we create a generator from a random seed that will ensure that we can generate similar images as well as run the pipeline:
+
+```python
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_pndm.png" width="400"/>
+    <br>
+</p>
+
+
+## Changing the scheduler
+
+Now we show how easy it is to change the scheduler of a pipeline. Every scheduler has a property [`SchedulerMixin.compatibles`] 
+which defines all compatible schedulers. You can take a look at all available, compatible schedulers for the Stable Diffusion pipeline as follows.
+
+```python
+pipeline.scheduler.compatibles
+```
+
+**Output**:
+```
+[diffusers.schedulers.scheduling_lms_discrete.LMSDiscreteScheduler,
+ diffusers.schedulers.scheduling_ddim.DDIMScheduler,
+ diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler,
+ diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler,
+ diffusers.schedulers.scheduling_pndm.PNDMScheduler,
+ diffusers.schedulers.scheduling_ddpm.DDPMScheduler,
+ diffusers.schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteScheduler]
+```
+
+Cool, lots of schedulers to look at. Feel free to have a look at their respective class definitions: 
+
+- [`LMSDiscreteScheduler`], 
+- [`DDIMScheduler`], 
+- [`DPMSolverMultistepScheduler`], 
+- [`EulerDiscreteScheduler`], 
+- [`PNDMScheduler`], 
+- [`DDPMScheduler`], 
+- [`EulerAncestralDiscreteScheduler`].
+
+We will now compare the input prompt with all other schedulers. To change the scheduler of the pipeline you can make use of the 
+convenient [`ConfigMixin.config`] property in combination with the [`ConfigMixin.from_config`] function.
+
+```python
+pipeline.scheduler.config
+```
+
+returns a dictionary of the configuration of the scheduler:
+
+**Output**:
+```
+FrozenDict([('num_train_timesteps', 1000),
+            ('beta_start', 0.00085),
+            ('beta_end', 0.012),
+            ('beta_schedule', 'scaled_linear'),
+            ('trained_betas', None),
+            ('skip_prk_steps', True),
+            ('set_alpha_to_one', False),
+            ('steps_offset', 1),
+            ('_class_name', 'PNDMScheduler'),
+            ('_diffusers_version', '0.8.0.dev0'),
+            ('clip_sample', False)])
+```
+
+This configuration can then be used to instantiate a scheduler
+of a different class that is compatible with the pipeline. Here, 
+we change the scheduler to the [`DDIMScheduler`].
+
+```python
+from diffusers import DDIMScheduler
+
+pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+```
+
+Cool, now we can run the pipeline again to compare the generation quality.
+
+```python
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_ddim.png" width="400"/>
+    <br>
+</p>
+
+
+## Compare schedulers
+
+So far we have tried running the stable diffusion pipeline with two schedulers: [`PNDMScheduler`] and [`DDIMScheduler`]. 
+A number of better schedulers have been released that can be run with much fewer steps, let's compare them here:
+
+[`LMSDiscreteScheduler`] usually leads to better results:
+
+```python
+from diffusers import LMSDiscreteScheduler
+
+pipeline.scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_lms.png" width="400"/>
+    <br>
+</p>
+
+
+[`EulerDiscreteScheduler`] and [`EulerAncestralDiscreteScheduler`] can generate high quality results with as little as 30 steps.
+
+```python
+from diffusers import EulerDiscreteScheduler
+
+pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator, num_inference_steps=30).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_euler_discrete.png" width="400"/>
+    <br>
+</p>
+
+
+and:
+
+```python
+from diffusers import EulerAncestralDiscreteScheduler
+
+pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator, num_inference_steps=30).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_euler_ancestral.png" width="400"/>
+    <br>
+</p>
+
+
+At the time of writing this doc [`DPMSolverMultistepScheduler`] gives arguably the best speed/quality trade-off and can be run with as little 
+as 20 steps.
+
+```python
+from diffusers import DPMSolverMultistepScheduler
+
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
+
+generator = torch.Generator(device="cuda").manual_seed(8)
+image = pipeline(prompt, generator=generator, num_inference_steps=20).images[0]
+image
+```
+
+<p align="center">
+    <br>
+    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_dpm.png" width="400"/>
+    <br>
+</p>
+
+As you can see most images look very similar and are arguably of very similar quality. It often really depends on the specific use case which scheduler to choose. A good approach is always to run multiple different
+schedulers to compare results.

examples/README.md

@@ -16,7 +16,7 @@ limitations under the License.
 # 🧨 Diffusers Examples
 
 Diffusers examples are a collection of scripts to demonstrate how to effectively use the `diffusers` library
-for a variety of use cases.
+for a variety of use cases involving training or fine-tuning.
 
 **Note**: If you are looking for **official** examples on how to use `diffusers` for inference, 
 please have a look at [src/diffusers/pipelines](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines)
@@ -38,7 +38,11 @@ Training examples show how to pretrain or fine-tune diffusion models for a varie
 
 | Task | 🤗 Accelerate | 🤗 Datasets | Colab
 |---|---|:---:|:---:|
-| [**Unconditional Image Generation**](https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/train_unconditional.py) | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [**Unconditional Image Generation**](./unconditional_image_generation) | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [**Text-to-Image fine-tuning**](./text_to_image) | ✅ | ✅ | 
+| [**Textual Inversion**](./textual_inversion) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb)
+| [**Dreambooth**](./dreambooth) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_dreambooth_training.ipynb)
+| [**Reinforcement Learning for Control**](https://github.com/huggingface/diffusers/blob/main/examples/rl/run_diffusers_locomotion.py)                    | - | - | coming soon.
 
 ## Community
 

examples/community/README.md

@@ -1,7 +1,728 @@
 # Community Examples
 
-**Community** examples consist of both inference and training examples that have been added by the community.
+> **For more information about community pipelines, please have a look at [this issue](https://github.com/huggingface/diffusers/issues/841).**
+
+**Community** examples consist of both inference and training examples that have been added by the community.
+Please have a look at the following table to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste ready code example that you can try out.
+If a community doesn't work as expected, please open an issue and ping the author on it.
+
+| Example                                | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                      | Colab                                                                                                                                                                                                              |                                                     Author |
+|:---------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------:|
+| CLIP Guided Stable Diffusion           | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)     | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |             [Suraj Patil](https://github.com/patil-suraj/) | 
+| One Step U-Net (Dummy)                 | Example showcasing of how to use Community Pipelines (see https://github.com/huggingface/diffusers/issues/841)                                                                                                                                                                                                                                                                                                                                                                                           | [One Step U-Net](#one-step-unet)                                  | -                                                                                                                                                                                                                  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
+| Stable Diffusion Interpolation         | Interpolate the latent space of Stable Diffusion between different prompts/seeds                                                                                                                                                                                                                                                                                                                                                                                                                         | [Stable Diffusion Interpolation](#stable-diffusion-interpolation) | -                                                                                                                                                                                                                  |                    [Nate Raw](https://github.com/nateraw/) |
+| Stable Diffusion Mega                  | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega)                   | -                                                                                                                                                                                                                  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
+| Long Prompt Weighting Stable Diffusion | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [SkyTNT](https://github.com/SkyTNT) |
+| Speech to Image                        | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                               | -                                                                                                                                                                                                                  | [Mikail Duzenli](https://github.com/MikailINTech)
+| Wild Card Stable Diffusion | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                                                     | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Shyam Sudhakaran](https://github.com/shyamsn97) |
+| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                     | [Composable Stable Diffusion](#composable-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
+| Seed Resizing Stable Diffusion| Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                     | [Seed Resizing](#seed-resizing)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
+| Imagic Stable Diffusion | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image| [Imagic Stable Diffusion](#imagic-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Mark Rich](https://github.com/MarkRich) |
+| Multilingual Stable Diffusion| Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                     | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                                                                 | -                                                                                                                                                                                                                  |                        [Juan Carlos Piñeros](https://github.com/juancopi81) |
+| Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting| [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Alex McKinney](https://github.com/vvvm23) |
+| Text Based Inpainting Stable Diffusion | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting| [Text Based Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion)                                                                 | -                                                                                                                                                                                                                  |                        [Dhruv Karan](https://github.com/unography) |
+| Bit Diffusion | Diffusion on discrete data | [Bit Diffusion](#bit-diffusion) | -  |[Stuti R.](https://github.com/kingstut) |
+| K-Diffusion Stable Diffusion | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py) | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion) | -  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
+
+
+
+To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
+```py
+pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="filename_in_the_community_folder")
+```
+
+## Example usages
+
+### CLIP Guided Stable Diffusion
+
+CLIP guided stable diffusion can help to generate more realistic images 
+by guiding stable diffusion at every denoising step with an additional CLIP model.
+
+The following code requires roughly 12GB of GPU RAM.
+
+```python
+from diffusers import DiffusionPipeline
+from transformers import CLIPFeatureExtractor, CLIPModel
+import torch
+
+
+feature_extractor = CLIPFeatureExtractor.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K")
+clip_model = CLIPModel.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K", torch_dtype=torch.float16)
+
+
+guided_pipeline = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="clip_guided_stable_diffusion",
+    clip_model=clip_model,
+    feature_extractor=feature_extractor,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+guided_pipeline.enable_attention_slicing()
+guided_pipeline = guided_pipeline.to("cuda")
+
+prompt = "fantasy book cover, full moon, fantasy forest landscape, golden vector elements, fantasy magic, dark light night, intricate, elegant, sharp focus, illustration, highly detailed, digital painting, concept art, matte, art by WLOP and Artgerm and Albert Bierstadt, masterpiece"
+
+generator = torch.Generator(device="cuda").manual_seed(0)
+images = []
+for i in range(4):
+    image = guided_pipeline(
+        prompt,
+        num_inference_steps=50,
+        guidance_scale=7.5,
+        clip_guidance_scale=100,
+        num_cutouts=4,
+        use_cutouts=False,
+        generator=generator,
+    ).images[0]
+    images.append(image)
+    
+# save images locally
+for i, img in enumerate(images):
+    img.save(f"./clip_guided_sd/image_{i}.png")
+```
+
+The `images` list contains a list of PIL images that can be saved locally or displayed directly in a google colab.
+Generated images tend to be of higher qualtiy than natively using stable diffusion. E.g. the above script generates the following images:
+
+![clip_guidance](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/clip_guidance/merged_clip_guidance.jpg).
+
+### One Step Unet
+
+The dummy "one-step-unet" can be run as follows:
+
+```python
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
+pipe()
+```
+
+**Note**: This community pipeline is not useful as a feature, but rather just serves as an example of how community pipelines can be added (see https://github.com/huggingface/diffusers/issues/841).
+
+### Stable Diffusion Interpolation
+
+The following code can be run on a GPU of at least 8GB VRAM and should take approximately 5 minutes.
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    revision='fp16',
+    torch_dtype=torch.float16,
+    safety_checker=None,  # Very important for videos...lots of false positives while interpolating
+    custom_pipeline="interpolate_stable_diffusion",
+).to('cuda')
+pipe.enable_attention_slicing()
+
+frame_filepaths = pipe.walk(
+    prompts=['a dog', 'a cat', 'a horse'],
+    seeds=[42, 1337, 1234],
+    num_interpolation_steps=16,
+    output_dir='./dreams',
+    batch_size=4,
+    height=512,
+    width=512,
+    guidance_scale=8.5,
+    num_inference_steps=50,
+)
+```
+
+The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion. 
+
+> **Please have a look at https://github.com/nateraw/stable-diffusion-videos for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
+
+### Stable Diffusion Mega
+
+The Stable Diffusion Mega Pipeline lets you use the main use cases of the stable diffusion pipeline in a single class.
+
+```python
+#!/usr/bin/env python3
+from diffusers import DiffusionPipeline
+import PIL
+import requests
+from io import BytesIO
+import torch
+
+
+def download_image(url):
+    response = requests.get(url)
+    return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_mega", torch_dtype=torch.float16, revision="fp16")
+pipe.to("cuda")
+pipe.enable_attention_slicing()
+
+
+### Text-to-Image
+
+images = pipe.text2img("An astronaut riding a horse").images
+
+### Image-to-Image
+
+init_image = download_image("https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg")
+
+prompt = "A fantasy landscape, trending on artstation"
+
+images = pipe.img2img(prompt=prompt, init_image=init_image, strength=0.75, guidance_scale=7.5).images
+
+### Inpainting
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+init_image = download_image(img_url).resize((512, 512))
+mask_image = download_image(mask_url).resize((512, 512))
+
+prompt = "a cat sitting on a bench"
+images = pipe.inpaint(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images
+```
+
+As shown above this one pipeline can run all both "text-to-image", "image-to-image", and "inpainting" in one pipeline.
+
+### Long Prompt Weighting Stable Diffusion
+Features of this custom pipeline:
+- Input a prompt without the 77 token length limit.
+- Includes tx2img, img2img. and inpainting pipelines.
+- Emphasize/weigh part of your prompt with parentheses as so: `a baby deer with (big eyes)`
+- De-emphasize part of your prompt as so: `a [baby] deer with big eyes`
+- Precisely weigh part of your prompt as so: `a baby deer with (big eyes:1.3)`
+
+Prompt weighting equivalents:
+- `a baby deer with` == `(a baby deer with:1.0)`
+- `(big eyes)` == `(big eyes:1.1)`
+- `((big eyes))` == `(big eyes:1.21)`
+- `[big eyes]` == `(big eyes:0.91)`
+
+You can run this custom pipeline as so:
+
+#### pytorch
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    'hakurei/waifu-diffusion',
+    custom_pipeline="lpw_stable_diffusion",
+    revision="fp16",
+    torch_dtype=torch.float16
+)
+pipe=pipe.to("cuda")
+
+prompt = "best_quality (1girl:1.3) bow bride brown_hair closed_mouth frilled_bow frilled_hair_tubes frills (full_body:1.3) fox_ear hair_bow hair_tubes happy hood japanese_clothes kimono long_sleeves red_bow smile solo tabi uchikake white_kimono wide_sleeves cherry_blossoms"
+neg_prompt = "lowres, bad_anatomy, error_body, error_hair, error_arm, error_hands, bad_hands, error_fingers, bad_fingers, missing_fingers, error_legs, bad_legs, multiple_legs, missing_legs, error_lighting, error_shadow, error_reflection, text, error, extra_digit, fewer_digits, cropped, worst_quality, low_quality, normal_quality, jpeg_artifacts, signature, watermark, username, blurry"
+
+pipe.text2img(prompt, negative_prompt=neg_prompt, width=512,height=512,max_embeddings_multiples=3).images[0]
+
+```
+
+#### onnxruntime
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    'CompVis/stable-diffusion-v1-4',
+    custom_pipeline="lpw_stable_diffusion_onnx",
+    revision="onnx",
+    provider="CUDAExecutionProvider"
+)
+
+prompt = "a photo of an astronaut riding a horse on mars, best quality"
+neg_prompt = "lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
+
+pipe.text2img(prompt,negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
+
+```
+
+if you see `Token indices sequence length is longer than the specified maximum sequence length for this model ( *** > 77 ) . Running this sequence through the model will result in indexing errors`. Do not worry, it is normal.
+
+### Speech to Image
+
+The following code can generate an image from an audio sample using pre-trained OpenAI whisper-small and Stable Diffusion.
+
+```Python
+import torch
+
+import matplotlib.pyplot as plt
+from datasets import load_dataset
+from diffusers import DiffusionPipeline
+from transformers import (
+    WhisperForConditionalGeneration,
+    WhisperProcessor,
+)
+
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+audio_sample = ds[3]
+
+text = audio_sample["text"].lower()
+speech_data = audio_sample["audio"]["array"]
+
+model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small").to(device)
+processor = WhisperProcessor.from_pretrained("openai/whisper-small")
+
+diffuser_pipeline = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="speech_to_image_diffusion",
+    speech_model=model,
+    speech_processor=processor,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+
+diffuser_pipeline.enable_attention_slicing()
+diffuser_pipeline = diffuser_pipeline.to(device)
+
+output = diffuser_pipeline(speech_data)
+plt.imshow(output.images[0])
+```
+This example produces the following image:
+
+![image](https://user-images.githubusercontent.com/45072645/196901736-77d9c6fc-63ee-4072-90b0-dc8b903d63e3.png)
+
+### Wildcard Stable Diffusion
+Following the great examples from https://github.com/jtkelm2/stable-diffusion-webui-1/blob/master/scripts/wildcards.py and https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#wildcards, here's a minimal implementation that allows for users to add "wildcards", denoted by `__wildcard__` to prompts that are used as placeholders for randomly sampled values given by either a dictionary or a `.txt` file. For example:
+
+Say we have a prompt:
+
+```
+prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
+```
+
+We can then define possible values to be sampled for `animal`, `object`, and `clothing`. These can either be from a `.txt` with the same name as the category.
+
+The possible values can also be defined / combined by using a dictionary like: `{"animal":["dog", "cat", mouse"]}`.
+
+The actual pipeline works just like `StableDiffusionPipeline`, except the `__call__` method takes in:
+
+`wildcard_files`: list of file paths for wild card replacement
+`wildcard_option_dict`: dict with key as `wildcard` and values as a list of possible replacements
+`num_prompt_samples`: number of prompts to sample, uniformly sampling wildcards
+
+A full example:
+
+create `animal.txt`, with contents like:
+
+```
+dog
+cat
+mouse
+```
+
+create `object.txt`, with contents like:
+
+```
+chair
+sofa
+bench
+```
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="wildcard_stable_diffusion",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
+out = pipe(
+    prompt,
+    wildcard_option_dict={
+        "clothing":["hat", "shirt", "scarf", "beret"]
+    },
+    wildcard_files=["object.txt", "animal.txt"],
+    num_prompt_samples=1
+)
+```
+
+### Composable Stable diffusion 
+
+[Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) proposes conjunction and negation (negative prompts) operators for compositional generation with conditional diffusion models.
+
+```python
+import torch as th
+import numpy as np
+import torchvision.utils as tvu
+from diffusers import DiffusionPipeline
+
+has_cuda = th.cuda.is_available()
+device = th.device('cpu' if not has_cuda else 'cuda')
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
+    custom_pipeline="composable_stable_diffusion",
+).to(device)
+
+
+def dummy(images, **kwargs):
+    return images, False
+
+pipe.safety_checker = dummy
+
+images = []
+generator = torch.Generator("cuda").manual_seed(0)
+
+seed = 0
+prompt = "a forest | a camel"
+weights = " 1 | 1"  # Equal weight to each prompt. Can be negative
+
+images = []
+for i in range(4):
+    res = pipe(
+        prompt,
+        guidance_scale=7.5,
+        num_inference_steps=50,
+        weights=weights,
+        generator=generator)
+    image = res.images[0]
+    images.append(image)
+
+for i, img in enumerate(images):
+    img.save(f"./composable_diffusion/image_{i}.png")
+```
+
+### Imagic Stable Diffusion
+Allows you to edit an image using stable diffusion. 
+
+```python
+import requests
+from PIL import Image
+from io import BytesIO
+import torch
+import os
+from diffusers import DiffusionPipeline, DDIMScheduler
+has_cuda = torch.cuda.is_available()
+device = torch.device('cpu' if not has_cuda else 'cuda')
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+        safety_checker=None,
+    use_auth_token=True,
+    custom_pipeline="imagic_stable_diffusion",
+    scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
+).to(device)
+generator = th.Generator("cuda").manual_seed(0)
+seed = 0
+prompt = "A photo of Barack Obama smiling with a big grin"
+url = 'https://www.dropbox.com/s/6tlwzr73jd1r9yk/obama.png?dl=1'
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((512, 512))
+res = pipe.train(
+    prompt,
+    init_image,
+    guidance_scale=7.5,
+    num_inference_steps=50,
+    generator=generator)
+res = pipe(alpha=1)
+os.makedirs("imagic", exist_ok=True)
+image = res.images[0]
+image.save('./imagic/imagic_image_alpha_1.png')
+res = pipe(alpha=1.5)
+image = res.images[0]
+image.save('./imagic/imagic_image_alpha_1_5.png')
+res = pipe(alpha=2)
+image = res.images[0]
+image.save('./imagic/imagic_image_alpha_2.png')
+```
+
+### Seed Resizing 
+Test seed resizing. Originally generate an image in 512 by 512, then generate image with same seed at 512 by 592 using seed resizing. Finally, generate 512 by 592 using original stable diffusion pipeline.
+
+```python
+import torch as th
+import numpy as np
+from diffusers import DiffusionPipeline
+
+has_cuda = th.cuda.is_available()
+device = th.device('cpu' if not has_cuda else 'cuda')
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
+    custom_pipeline="seed_resize_stable_diffusion"
+).to(device)
+
+def dummy(images, **kwargs):
+    return images, False
+
+pipe.safety_checker = dummy
+
+
+images = []
+th.manual_seed(0)
+generator = th.Generator("cuda").manual_seed(0)
+
+seed = 0
+prompt = "A painting of a futuristic cop"
+
+width = 512
+height = 512
+
+res = pipe(
+    prompt,
+    guidance_scale=7.5,
+    num_inference_steps=50,
+    height=height,
+    width=width,
+    generator=generator)
+image = res.images[0]
+image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=height))
+
+
+th.manual_seed(0)
+generator = th.Generator("cuda").manual_seed(0)
+
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
+    custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
+).to(device)
+
+width = 512
+height = 592
+
+res = pipe(
+    prompt,
+    guidance_scale=7.5,
+    num_inference_steps=50,
+    height=height,
+    width=width,
+    generator=generator)
+image = res.images[0]
+image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=height))
+
+pipe_compare = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    use_auth_token=True,
+    custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
+).to(device)
+
+res = pipe_compare(
+    prompt,
+    guidance_scale=7.5,
+    num_inference_steps=50,
+    height=height,
+    width=width,
+    generator=generator
+)
+
+image = res.images[0]
+image.save('./seed_resize/seed_resize_{w}_{h}_image_compare.png'.format(w=width, h=height))
+```
+
+### Multilingual Stable Diffusion Pipeline
+
+The following code can generate an images from texts in different languages using the pre-trained [mBART-50 many-to-one multilingual machine translation model](https://huggingface.co/facebook/mbart-large-50-many-to-one-mmt) and Stable Diffusion.
+
+```python
+from PIL import Image
+
+import torch
+
+from diffusers import DiffusionPipeline
+from transformers import (
+    pipeline,
+    MBart50TokenizerFast,
+    MBartForConditionalGeneration,
+)
+device = "cuda" if torch.cuda.is_available() else "cpu"
+device_dict = {"cuda": 0, "cpu": -1}
+
+# helper function taken from: https://huggingface.co/blog/stable_diffusion
+def image_grid(imgs, rows, cols):
+    assert len(imgs) == rows*cols
+
+    w, h = imgs[0].size
+    grid = Image.new('RGB', size=(cols*w, rows*h))
+    grid_w, grid_h = grid.size
+
+    for i, img in enumerate(imgs):
+        grid.paste(img, box=(i%cols*w, i//cols*h))
+    return grid
+
+# Add language detection pipeline
+language_detection_model_ckpt = "papluca/xlm-roberta-base-language-detection"
+language_detection_pipeline = pipeline("text-classification",
+                                       model=language_detection_model_ckpt,
+                                       device=device_dict[device])
+
+# Add model for language translation
+trans_tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-one-mmt")
+trans_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-one-mmt").to(device)
+
+diffuser_pipeline = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    custom_pipeline="multilingual_stable_diffusion",
+    detection_pipeline=language_detection_pipeline,
+    translation_model=trans_model,
+    translation_tokenizer=trans_tokenizer,
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+
+diffuser_pipeline.enable_attention_slicing()
+diffuser_pipeline = diffuser_pipeline.to(device)
+
+prompt = ["a photograph of an astronaut riding a horse", 
+          "Una casa en la playa",
+          "Ein Hund, der Orange isst",
+          "Un restaurant parisien"]
+
+output = diffuser_pipeline(prompt)
+
+images = output.images
+
+grid = image_grid(images, rows=2, cols=2)
+```
+
+This example produces the following images:
+![image](https://user-images.githubusercontent.com/4313860/198328706-295824a4-9856-4ce5-8e66-278ceb42fd29.png)
+
+### Image to Image Inpainting Stable Diffusion
+
+Similar to the standard stable diffusion inpainting example, except with the addition of an `inner_image` argument.
+
+`image`, `inner_image`, and `mask` should have the same dimensions. `inner_image` should have an alpha (transparency) channel.
+
+The aim is to overlay two images, then mask out the boundary between `image` and `inner_image` to allow stable diffusion to make the connection more seamless.
+For example, this could be used to place a logo on a shirt and make it blend seamlessly.
+
+```python
+import PIL
+import torch
+
+from diffusers import StableDiffusionInpaintPipeline
+
+image_path = "./path-to-image.png"
+inner_image_path = "./path-to-inner-image.png"
+mask_path = "./path-to-mask.png"
+
+init_image = PIL.Image.open(image_path).convert("RGB").resize((512, 512))
+inner_image = PIL.Image.open(inner_image_path).convert("RGBA").resize((512, 512))
+mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
+
+pipe = StableDiffusionInpaintPipeline.from_pretrained(
+    "runwayml/stable-diffusion-inpainting",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+prompt = "Your prompt here!"
+image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
+```
+
+### Text Based Inpainting Stable Diffusion
+
+Use a text prompt to generate the mask for the area to be inpainted.
+Currently uses the CLIPSeg model for mask generation, then calls the standard Stable Diffusion Inpainting pipeline to perform the inpainting.
+
+```python
+from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
+from diffusers import DiffusionPipeline
+
+from PIL import Image
+import requests
+from torch import autocast
+
+processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
+model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
+
+pipe = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-inpainting",
+    custom_pipeline="text_inpainting",
+    segmentation_model=model,
+    segmentation_processor=processor
+)
+pipe = pipe.to("cuda")
+
+
+url = "https://github.com/timojl/clipseg/blob/master/example_image.jpg?raw=true"
+image = Image.open(requests.get(url, stream=True).raw).resize((512, 512))
+text = "a glass"  # will mask out this text
+prompt = "a cup"  # the masked out region will be replaced with this
+
+with autocast("cuda"):
+    image = pipe(image=image, text=text, prompt=prompt).images[0]
+```
+
+### Bit Diffusion 
+Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this: 
+
+```python
+from diffusers import DiffusionPipeline
+pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="bit_diffusion")
+image = pipe().images[0]
+
+```
+
+### Stable Diffusion with K Diffusion
+
+Make sure you have @crowsonkb's https://github.com/crowsonkb/k-diffusion installed:
+
+```
+pip install k-diffusion
+```
+
+You can use the community pipeline as follows:
+
+```python
+from diffusers import DiffusionPipeline
+
+pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion")
+pipe = pipe.to("cuda")
+
+prompt = "an astronaut riding a horse on mars"
+pipe.set_sampler("sample_heun")
+generator = torch.Generator(device="cuda").manual_seed(seed)
+image = pipe(prompt, generator=generator, num_inference_steps=20).images[0]
+
+image.save("./astronaut_heun_k_diffusion.png")
+```
+
+To make sure that K Diffusion and `diffusers` yield the same results:
+
+**Diffusers**:
+```python
+from diffusers import DiffusionPipeline, EulerDiscreteScheduler
+
+seed = 33
+
+pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to("cuda")
+
+generator = torch.Generator(device="cuda").manual_seed(seed)
+image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
+```
+
+![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler.png)
+
+**K Diffusion**:
+```python
+from diffusers import DiffusionPipeline, EulerDiscreteScheduler
+
+seed = 33
+
+pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion")
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to("cuda")
+
+pipe.set_sampler("sample_euler")
+generator = torch.Generator(device="cuda").manual_seed(seed)
+image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
+```
+
+![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler_k_diffusion.png)
 
-| Example   |      Description      |      Author      |    Colab  |
-|:----------|:----------------------|:-----------------|----------:|
-| CLIP Guided Stable Diffusion | Doing CLIP guidance for text to image generation with Stable Diffusion| [Suraj Patil](https://github.com/patil-suraj/) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |

examples/community/bit_diffusion.py

@@ -0,0 +1,263 @@
+from typing import Optional, Tuple, Union
+
+import torch
+
+from diffusers import DDIMScheduler, DDPMScheduler, DiffusionPipeline, UNet2DConditionModel
+from diffusers.pipeline_utils import ImagePipelineOutput
+from diffusers.schedulers.scheduling_ddim import DDIMSchedulerOutput
+from diffusers.schedulers.scheduling_ddpm import DDPMSchedulerOutput
+from einops import rearrange, reduce
+
+
+BITS = 8
+
+
+# convert to bit representations and back taken from https://github.com/lucidrains/bit-diffusion/blob/main/bit_diffusion/bit_diffusion.py
+def decimal_to_bits(x, bits=BITS):
+    """expects image tensor ranging from 0 to 1, outputs bit tensor ranging from -1 to 1"""
+    device = x.device
+
+    x = (x * 255).int().clamp(0, 255)
+
+    mask = 2 ** torch.arange(bits - 1, -1, -1, device=device)
+    mask = rearrange(mask, "d -> d 1 1")
+    x = rearrange(x, "b c h w -> b c 1 h w")
+
+    bits = ((x & mask) != 0).float()
+    bits = rearrange(bits, "b c d h w -> b (c d) h w")
+    bits = bits * 2 - 1
+    return bits
+
+
+def bits_to_decimal(x, bits=BITS):
+    """expects bits from -1 to 1, outputs image tensor from 0 to 1"""
+    device = x.device
+
+    x = (x > 0).int()
+    mask = 2 ** torch.arange(bits - 1, -1, -1, device=device, dtype=torch.int32)
+
+    mask = rearrange(mask, "d -> d 1 1")
+    x = rearrange(x, "b (c d) h w -> b c d h w", d=8)
+    dec = reduce(x * mask, "b c d h w -> b c h w", "sum")
+    return (dec / 255).clamp(0.0, 1.0)
+
+
+# modified scheduler step functions for clamping the predicted x_0 between -bit_scale and +bit_scale
+def ddim_bit_scheduler_step(
+    self,
+    model_output: torch.FloatTensor,
+    timestep: int,
+    sample: torch.FloatTensor,
+    eta: float = 0.0,
+    use_clipped_model_output: bool = True,
+    generator=None,
+    return_dict: bool = True,
+) -> Union[DDIMSchedulerOutput, Tuple]:
+    """
+    Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+    process from the learned model outputs (most often the predicted noise).
+    Args:
+        model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+        timestep (`int`): current discrete timestep in the diffusion chain.
+        sample (`torch.FloatTensor`):
+            current instance of sample being created by diffusion process.
+        eta (`float`): weight of noise for added noise in diffusion step.
+        use_clipped_model_output (`bool`): TODO
+        generator: random number generator.
+        return_dict (`bool`): option for returning tuple rather than DDIMSchedulerOutput class
+    Returns:
+        [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] or `tuple`:
+        [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
+        returning a tuple, the first element is the sample tensor.
+    """
+    if self.num_inference_steps is None:
+        raise ValueError(
+            "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+        )
+
+    # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+    # Ideally, read DDIM paper in-detail understanding
+
+    # Notation (<variable name> -> <name in paper>
+    # - pred_noise_t -> e_theta(x_t, t)
+    # - pred_original_sample -> f_theta(x_t, t) or x_0
+    # - std_dev_t -> sigma_t
+    # - eta -> η
+    # - pred_sample_direction -> "direction pointing to x_t"
+    # - pred_prev_sample -> "x_t-1"
+
+    # 1. get previous step value (=t-1)
+    prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+    # 2. compute alphas, betas
+    alpha_prod_t = self.alphas_cumprod[timestep]
+    alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+
+    beta_prod_t = 1 - alpha_prod_t
+
+    # 3. compute predicted original sample from predicted noise also called
+    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+
+    # 4. Clip "predicted x_0"
+    scale = self.bit_scale
+    if self.config.clip_sample:
+        pred_original_sample = torch.clamp(pred_original_sample, -scale, scale)
+
+    # 5. compute variance: "sigma_t(η)" -> see formula (16)
+    # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+    variance = self._get_variance(timestep, prev_timestep)
+    std_dev_t = eta * variance ** (0.5)
+
+    if use_clipped_model_output:
+        # the model_output is always re-derived from the clipped x_0 in Glide
+        model_output = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+
+    # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output
+
+    # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+    if eta > 0:
+        # randn_like does not support generator https://github.com/pytorch/pytorch/issues/27072
+        device = model_output.device if torch.is_tensor(model_output) else "cpu"
+        noise = torch.randn(model_output.shape, dtype=model_output.dtype, generator=generator).to(device)
+        variance = self._get_variance(timestep, prev_timestep) ** (0.5) * eta * noise
+
+        prev_sample = prev_sample + variance
+
+    if not return_dict:
+        return (prev_sample,)
+
+    return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+
+def ddpm_bit_scheduler_step(
+    self,
+    model_output: torch.FloatTensor,
+    timestep: int,
+    sample: torch.FloatTensor,
+    predict_epsilon=True,
+    generator=None,
+    return_dict: bool = True,
+) -> Union[DDPMSchedulerOutput, Tuple]:
+    """
+    Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+    process from the learned model outputs (most often the predicted noise).
+    Args:
+        model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+        timestep (`int`): current discrete timestep in the diffusion chain.
+        sample (`torch.FloatTensor`):
+            current instance of sample being created by diffusion process.
+        predict_epsilon (`bool`):
+            optional flag to use when model predicts the samples directly instead of the noise, epsilon.
+        generator: random number generator.
+        return_dict (`bool`): option for returning tuple rather than DDPMSchedulerOutput class
+    Returns:
+        [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] or `tuple`:
+        [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
+        returning a tuple, the first element is the sample tensor.
+    """
+    t = timestep
+
+    if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
+        model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
+    else:
+        predicted_variance = None
+
+    # 1. compute alphas, betas
+    alpha_prod_t = self.alphas_cumprod[t]
+    alpha_prod_t_prev = self.alphas_cumprod[t - 1] if t > 0 else self.one
+    beta_prod_t = 1 - alpha_prod_t
+    beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+    # 2. compute predicted original sample from predicted noise also called
+    # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+    if predict_epsilon:
+        pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+    else:
+        pred_original_sample = model_output
+
+    # 3. Clip "predicted x_0"
+    scale = self.bit_scale
+    if self.config.clip_sample:
+        pred_original_sample = torch.clamp(pred_original_sample, -scale, scale)
+
+    # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+    # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+    pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * self.betas[t]) / beta_prod_t
+    current_sample_coeff = self.alphas[t] ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+    # 5. Compute predicted previous sample µ_t
+    # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+    pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+    # 6. Add noise
+    variance = 0
+    if t > 0:
+        noise = torch.randn(
+            model_output.size(), dtype=model_output.dtype, layout=model_output.layout, generator=generator
+        ).to(model_output.device)
+        variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * noise
+
+    pred_prev_sample = pred_prev_sample + variance
+
+    if not return_dict:
+        return (pred_prev_sample,)
+
+    return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
+
+
+class BitDiffusion(DiffusionPipeline):
+    def __init__(
+        self,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, DDPMScheduler],
+        bit_scale: Optional[float] = 1.0,
+    ):
+        super().__init__()
+        self.bit_scale = bit_scale
+        self.scheduler.step = (
+            ddim_bit_scheduler_step if isinstance(scheduler, DDIMScheduler) else ddpm_bit_scheduler_step
+        )
+
+        self.register_modules(unet=unet, scheduler=scheduler)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        height: Optional[int] = 256,
+        width: Optional[int] = 256,
+        num_inference_steps: Optional[int] = 50,
+        generator: Optional[torch.Generator] = None,
+        batch_size: Optional[int] = 1,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        **kwargs,
+    ) -> Union[Tuple, ImagePipelineOutput]:
+        latents = torch.randn(
+            (batch_size, self.unet.in_channels, height, width),
+            generator=generator,
+        )
+        latents = decimal_to_bits(latents) * self.bit_scale
+        latents = latents.to(self.device)
+
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        for t in self.progress_bar(self.scheduler.timesteps):
+            # predict the noise residual
+            noise_pred = self.unet(latents, t).sample
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+
+        image = bits_to_decimal(latents)
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

examples/community/clip_guided_stable_diffusion.py

@@ -5,7 +5,14 @@ import torch
 from torch import nn
 from torch.nn import functional as F
 
-from diffusers import AutoencoderKL, DiffusionPipeline, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DiffusionPipeline,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    UNet2DConditionModel,
+)
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from torchvision import transforms
 from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTokenizer
@@ -56,7 +63,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         clip_model: CLIPModel,
         tokenizer: CLIPTokenizer,
         unet: UNet2DConditionModel,
-        scheduler: Union[PNDMScheduler, LMSDiscreteScheduler],
+        scheduler: Union[PNDMScheduler, LMSDiscreteScheduler, DDIMScheduler],
         feature_extractor: CLIPFeatureExtractor,
     ):
         super().__init__()
@@ -123,7 +130,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         # predict the noise residual
         noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample
 
-        if isinstance(self.scheduler, PNDMScheduler):
+        if isinstance(self.scheduler, (PNDMScheduler, DDIMScheduler)):
             alpha_prod_t = self.scheduler.alphas_cumprod[timestep]
             beta_prod_t = 1 - alpha_prod_t
             # compute predicted original sample from predicted noise also called
@@ -176,6 +183,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         num_inference_steps: Optional[int] = 50,
         guidance_scale: Optional[float] = 7.5,
         num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
         clip_guidance_scale: Optional[float] = 100,
         clip_prompt: Optional[Union[str, List[str]]] = None,
         num_cutouts: Optional[int] = 4,
@@ -249,7 +257,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         latents_dtype = text_embeddings.dtype
         if latents is None:
             if self.device.type == "mps":
-                # randn does not exist on mps
+                # randn does not work reproducibly on mps
                 latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
                     self.device
                 )
@@ -275,6 +283,20 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         # scale the initial noise by the standard deviation required by the scheduler
         latents = latents * self.scheduler.init_noise_sigma
 
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+
         for i, t in enumerate(self.progress_bar(timesteps_tensor)):
             # expand the latents if we are doing classifier free guidance
             latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
@@ -306,7 +328,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
                 )
 
             # compute the previous noisy sample x_t -> x_t-1
-            latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
         # scale and decode the image latents with vae
         latents = 1 / 0.18215 * latents

examples/community/composable_stable_diffusion.py

@@ -0,0 +1,329 @@
+"""
+    modified based on diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+"""
+import inspect
+import warnings
+from typing import List, Optional, Union
+
+import torch
+
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+class ComposableStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offsensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        weights: Optional[str] = "",
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        if "torch_device" in kwargs:
+            device = kwargs.pop("torch_device")
+            warnings.warn(
+                "`torch_device` is deprecated as an input argument to `__call__` and will be removed in v0.3.0."
+                " Consider using `pipe.to(torch_device)` instead."
+            )
+
+            # Set device as before (to be removed in 0.3.0)
+            if device is None:
+                device = "cuda" if torch.cuda.is_available() else "cpu"
+            self.to(device)
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if "|" in prompt:
+            prompt = [x.strip() for x in prompt.split("|")]
+            print(f"composing {prompt}...")
+
+        # get prompt text embeddings
+        text_input = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
+
+        if not weights:
+            # specify weights for prompts (excluding the unconditional score)
+            print("using equal weights for all prompts...")
+            pos_weights = torch.tensor(
+                [1 / (text_embeddings.shape[0] - 1)] * (text_embeddings.shape[0] - 1), device=self.device
+            ).reshape(-1, 1, 1, 1)
+            neg_weights = torch.tensor([1.0], device=self.device).reshape(-1, 1, 1, 1)
+            mask = torch.tensor([False] + [True] * pos_weights.shape[0], dtype=torch.bool)
+        else:
+            # set prompt weight for each
+            num_prompts = len(prompt) if isinstance(prompt, list) else 1
+            weights = [float(w.strip()) for w in weights.split("|")]
+            if len(weights) < num_prompts:
+                weights.append(1.0)
+            weights = torch.tensor(weights, device=self.device)
+            assert len(weights) == text_embeddings.shape[0], "weights specified are not equal to the number of prompts"
+            pos_weights = []
+            neg_weights = []
+            mask = []  # first one is unconditional score
+            for w in weights:
+                if w > 0:
+                    pos_weights.append(w)
+                    mask.append(True)
+                else:
+                    neg_weights.append(abs(w))
+                    mask.append(False)
+            # normalize the weights
+            pos_weights = torch.tensor(pos_weights, device=self.device).reshape(-1, 1, 1, 1)
+            pos_weights = pos_weights / pos_weights.sum()
+            neg_weights = torch.tensor(neg_weights, device=self.device).reshape(-1, 1, 1, 1)
+            neg_weights = neg_weights / neg_weights.sum()
+            mask = torch.tensor(mask, device=self.device, dtype=torch.bool)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            max_length = text_input.input_ids.shape[-1]
+
+            if torch.all(mask):
+                # no negative prompts, so we use empty string as the negative prompt
+                uncond_input = self.tokenizer(
+                    [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+                )
+                uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+                # For classifier free guidance, we need to do two forward passes.
+                # Here we concatenate the unconditional and text embeddings into a single batch
+                # to avoid doing two forward passes
+                text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+                # update negative weights
+                neg_weights = torch.tensor([1.0], device=self.device)
+                mask = torch.tensor([False] + mask.detach().tolist(), device=self.device, dtype=torch.bool)
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_device = "cpu" if self.device.type == "mps" else self.device
+        latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
+        if latents is None:
+            latents = torch.randn(
+                latents_shape,
+                generator=generator,
+                device=latents_device,
+            )
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+        latents = latents.to(self.device)
+
+        # set timesteps
+        accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
+        extra_set_kwargs = {}
+        if accepts_offset:
+            extra_set_kwargs["offset"] = 1
+
+        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
+
+        # if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas
+        if isinstance(self.scheduler, LMSDiscreteScheduler):
+            latents = latents * self.scheduler.sigmas[0]
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = (
+                torch.cat([latents] * text_embeddings.shape[0]) if do_classifier_free_guidance else latents
+            )
+            if isinstance(self.scheduler, LMSDiscreteScheduler):
+                sigma = self.scheduler.sigmas[i]
+                # the model input needs to be scaled to match the continuous ODE formulation in K-LMS
+                latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
+
+            # reduce memory by predicting each score sequentially
+            noise_preds = []
+            # predict the noise residual
+            for latent_in, text_embedding_in in zip(
+                torch.chunk(latent_model_input, chunks=latent_model_input.shape[0], dim=0),
+                torch.chunk(text_embeddings, chunks=text_embeddings.shape[0], dim=0),
+            ):
+                noise_preds.append(self.unet(latent_in, t, encoder_hidden_states=text_embedding_in).sample)
+            noise_preds = torch.cat(noise_preds, dim=0)
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond = (noise_preds[~mask] * neg_weights).sum(dim=0, keepdims=True)
+                noise_pred_text = (noise_preds[mask] * pos_weights).sum(dim=0, keepdims=True)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            if isinstance(self.scheduler, LMSDiscreteScheduler):
+                latents = self.scheduler.step(noise_pred, i, latents, **extra_step_kwargs).prev_sample
+            else:
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+        # scale and decode the image latents with vae
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+        # run safety checker
+        safety_cheker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
+        image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_cheker_input.pixel_values)
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/imagic_stable_diffusion.py

@@ -0,0 +1,497 @@
+"""
+    modeled after the textual_inversion.py / train_dreambooth.py and the work
+    of justinpinkney here: https://github.com/justinpinkney/stable-diffusion/blob/main/notebooks/imagic.ipynb
+"""
+import inspect
+import warnings
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+import PIL
+from accelerate import Accelerator
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import logging
+
+# TODO: remove and import from diffusers.utils when the new version of diffusers is released
+from packaging import version
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.Resampling.BILINEAR,
+        "bilinear": PIL.Image.Resampling.BILINEAR,
+        "bicubic": PIL.Image.Resampling.BICUBIC,
+        "lanczos": PIL.Image.Resampling.LANCZOS,
+        "nearest": PIL.Image.Resampling.NEAREST,
+    }
+else:
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.LINEAR,
+        "bilinear": PIL.Image.BILINEAR,
+        "bicubic": PIL.Image.BICUBIC,
+        "lanczos": PIL.Image.LANCZOS,
+        "nearest": PIL.Image.NEAREST,
+    }
+# ------------------------------------------------------------------------------
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def preprocess(image):
+    w, h = image.size
+    w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
+    image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+    image = np.array(image).astype(np.float32) / 255.0
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image)
+    return 2.0 * image - 1.0
+
+
+class ImagicStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for imagic image editing.
+    See paper here: https://arxiv.org/pdf/2210.09276.pdf
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offsensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def train(
+        self,
+        prompt: Union[str, List[str]],
+        init_image: Union[torch.FloatTensor, PIL.Image.Image],
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        generator: Optional[torch.Generator] = None,
+        embedding_learning_rate: float = 0.001,
+        diffusion_model_learning_rate: float = 2e-6,
+        text_embedding_optimization_steps: int = 500,
+        model_fine_tuning_optimization_steps: int = 1000,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `nd.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+        accelerator = Accelerator(
+            gradient_accumulation_steps=1,
+            mixed_precision="fp16",
+        )
+
+        if "torch_device" in kwargs:
+            device = kwargs.pop("torch_device")
+            warnings.warn(
+                "`torch_device` is deprecated as an input argument to `__call__` and will be removed in v0.3.0."
+                " Consider using `pipe.to(torch_device)` instead."
+            )
+
+            if device is None:
+                device = "cuda" if torch.cuda.is_available() else "cpu"
+            self.to(device)
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        # Freeze vae and unet
+        self.vae.requires_grad_(False)
+        self.unet.requires_grad_(False)
+        self.text_encoder.requires_grad_(False)
+        self.unet.eval()
+        self.vae.eval()
+        self.text_encoder.eval()
+
+        if accelerator.is_main_process:
+            accelerator.init_trackers(
+                "imagic",
+                config={
+                    "embedding_learning_rate": embedding_learning_rate,
+                    "text_embedding_optimization_steps": text_embedding_optimization_steps,
+                },
+            )
+
+        # get text embeddings for prompt
+        text_input = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncaton=True,
+            return_tensors="pt",
+        )
+        text_embeddings = torch.nn.Parameter(
+            self.text_encoder(text_input.input_ids.to(self.device))[0], requires_grad=True
+        )
+        text_embeddings = text_embeddings.detach()
+        text_embeddings.requires_grad_()
+        text_embeddings_orig = text_embeddings.clone()
+
+        # Initialize the optimizer
+        optimizer = torch.optim.Adam(
+            [text_embeddings],  # only optimize the embeddings
+            lr=embedding_learning_rate,
+        )
+
+        if isinstance(init_image, PIL.Image.Image):
+            init_image = preprocess(init_image)
+
+        latents_dtype = text_embeddings.dtype
+        init_image = init_image.to(device=self.device, dtype=latents_dtype)
+        init_latent_image_dist = self.vae.encode(init_image).latent_dist
+        init_image_latents = init_latent_image_dist.sample(generator=generator)
+        init_image_latents = 0.18215 * init_image_latents
+
+        progress_bar = tqdm(range(text_embedding_optimization_steps), disable=not accelerator.is_local_main_process)
+        progress_bar.set_description("Steps")
+
+        global_step = 0
+
+        logger.info("First optimizing the text embedding to better reconstruct the init image")
+        for _ in range(text_embedding_optimization_steps):
+            with accelerator.accumulate(text_embeddings):
+                # Sample noise that we'll add to the latents
+                noise = torch.randn(init_image_latents.shape).to(init_image_latents.device)
+                timesteps = torch.randint(1000, (1,), device=init_image_latents.device)
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = self.scheduler.add_noise(init_image_latents, noise, timesteps)
+
+                # Predict the noise residual
+                noise_pred = self.unet(noisy_latents, timesteps, text_embeddings).sample
+
+                loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                accelerator.backward(loss)
+
+                optimizer.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+            logs = {"loss": loss.detach().item()}  # , "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+
+        accelerator.wait_for_everyone()
+
+        text_embeddings.requires_grad_(False)
+
+        # Now we fine tune the unet to better reconstruct the image
+        self.unet.requires_grad_(True)
+        self.unet.train()
+        optimizer = torch.optim.Adam(
+            self.unet.parameters(),  # only optimize unet
+            lr=diffusion_model_learning_rate,
+        )
+        progress_bar = tqdm(range(model_fine_tuning_optimization_steps), disable=not accelerator.is_local_main_process)
+
+        logger.info("Next fine tuning the entire model to better reconstruct the init image")
+        for _ in range(model_fine_tuning_optimization_steps):
+            with accelerator.accumulate(self.unet.parameters()):
+                # Sample noise that we'll add to the latents
+                noise = torch.randn(init_image_latents.shape).to(init_image_latents.device)
+                timesteps = torch.randint(1000, (1,), device=init_image_latents.device)
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = self.scheduler.add_noise(init_image_latents, noise, timesteps)
+
+                # Predict the noise residual
+                noise_pred = self.unet(noisy_latents, timesteps, text_embeddings).sample
+
+                loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                accelerator.backward(loss)
+
+                optimizer.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+            logs = {"loss": loss.detach().item()}  # , "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+
+        accelerator.wait_for_everyone()
+        self.text_embeddings_orig = text_embeddings_orig
+        self.text_embeddings = text_embeddings
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        alpha: float = 1.2,
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        num_inference_steps: Optional[int] = 50,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        guidance_scale: float = 7.5,
+        eta: float = 0.0,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `nd.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+        if self.text_embeddings is None:
+            raise ValueError("Please run the pipe.train() before trying to generate an image.")
+        if self.text_embeddings_orig is None:
+            raise ValueError("Please run the pipe.train() before trying to generate an image.")
+
+        text_embeddings = alpha * self.text_embeddings_orig + (1 - alpha) * self.text_embeddings
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens = [""]
+            max_length = self.tokenizer.model_max_length
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.view(1, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (1, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if self.device.type == "mps":
+            # randn does not exist on mps
+            latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                self.device
+            )
+        else:
+            latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/img2img_inpainting.py

@@ -0,0 +1,463 @@
+import inspect
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+import PIL
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, logging
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def prepare_mask_and_masked_image(image, mask):
+    image = np.array(image.convert("RGB"))
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+    mask = np.array(mask.convert("L"))
+    mask = mask.astype(np.float32) / 255.0
+    mask = mask[None, None]
+    mask[mask < 0.5] = 0
+    mask[mask >= 0.5] = 1
+    mask = torch.from_numpy(mask)
+
+    masked_image = image * (mask < 0.5)
+
+    return mask, masked_image
+
+
+def check_size(image, height, width):
+    if isinstance(image, PIL.Image.Image):
+        w, h = image.size
+    elif isinstance(image, torch.Tensor):
+        *_, h, w = image.shape
+
+    if h != height or w != width:
+        raise ValueError(f"Image size should be {height}x{width}, but got {h}x{w}")
+
+
+def overlay_inner_image(image, inner_image, paste_offset: Tuple[int] = (0, 0)):
+    inner_image = inner_image.convert("RGBA")
+    image = image.convert("RGB")
+
+    image.paste(inner_image, paste_offset, inner_image)
+    image = image.convert("RGB")
+
+    return image
+
+
+class ImageToImageInpaintingPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-guided image-to-image inpainting using Stable Diffusion. *This is an experimental feature*.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        image: Union[torch.FloatTensor, PIL.Image.Image],
+        inner_image: Union[torch.FloatTensor, PIL.Image.Image],
+        mask_image: Union[torch.FloatTensor, PIL.Image.Image],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            image (`torch.Tensor` or `PIL.Image.Image`):
+                `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+                be masked out with `mask_image` and repainted according to `prompt`.
+            inner_image (`torch.Tensor` or `PIL.Image.Image`):
+                `Image`, or tensor representing an image batch which will be overlayed onto `image`. Non-transparent
+                regions of `inner_image` must fit inside white pixels in `mask_image`. Expects four channels, with
+                the last channel representing the alpha channel, which will be used to blend `inner_image` with
+                `image`. If not provided, it will be forcibly cast to RGBA.
+            mask_image (`PIL.Image.Image`):
+                `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+                repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+                to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+                instead of 3, so the expected shape would be `(B, H, W, 1)`.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        # check if input sizes are correct
+        check_size(image, height, width)
+        check_size(inner_image, height, width)
+        check_size(mask_image, height, width)
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""]
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        num_channels_latents = self.vae.config.latent_channels
+        latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not exist on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # overlay the inner image
+        image = overlay_inner_image(image, inner_image)
+
+        # prepare mask and masked_image
+        mask, masked_image = prepare_mask_and_masked_image(image, mask_image)
+        mask = mask.to(device=self.device, dtype=text_embeddings.dtype)
+        masked_image = masked_image.to(device=self.device, dtype=text_embeddings.dtype)
+
+        # resize the mask to latents shape as we concatenate the mask to the latents
+        mask = torch.nn.functional.interpolate(mask, size=(height // 8, width // 8))
+
+        # encode the mask image into latents space so we can concatenate it to the latents
+        masked_image_latents = self.vae.encode(masked_image).latent_dist.sample(generator=generator)
+        masked_image_latents = 0.18215 * masked_image_latents
+
+        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+        mask = mask.repeat(batch_size * num_images_per_prompt, 1, 1, 1)
+        masked_image_latents = masked_image_latents.repeat(batch_size * num_images_per_prompt, 1, 1, 1)
+
+        mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+        masked_image_latents = (
+            torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+        )
+
+        num_channels_mask = mask.shape[1]
+        num_channels_masked_image = masked_image_latents.shape[1]
+
+        if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+            raise ValueError(
+                f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+                f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+                f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+                f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                " `pipeline.unet` or your `mask_image` or `image` input."
+            )
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+            # concat latents, mask, masked_image_latents in the channel dimension
+            latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/interpolate_stable_diffusion.py

@@ -0,0 +1,524 @@
+import inspect
+import time
+from pathlib import Path
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, logging
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
+    """helper function to spherically interpolate two arrays v1 v2"""
+
+    if not isinstance(v0, np.ndarray):
+        inputs_are_torch = True
+        input_device = v0.device
+        v0 = v0.cpu().numpy()
+        v1 = v1.cpu().numpy()
+
+    dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
+    if np.abs(dot) > DOT_THRESHOLD:
+        v2 = (1 - t) * v0 + t * v1
+    else:
+        theta_0 = np.arccos(dot)
+        sin_theta_0 = np.sin(theta_0)
+        theta_t = theta_0 * t
+        sin_theta_t = np.sin(theta_t)
+        s0 = np.sin(theta_0 - theta_t) / sin_theta_0
+        s1 = sin_theta_t / sin_theta_0
+        v2 = s0 * v0 + s1 * v1
+
+    if inputs_are_torch:
+        v2 = torch.from_numpy(v2).to(input_device)
+
+    return v2
+
+
+class StableDiffusionWalkPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        text_embeddings: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*, defaults to `None`):
+                The prompt or prompts to guide the image generation. If not provided, `text_embeddings` is required.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            text_embeddings (`torch.FloatTensor`, *optional*, defaults to `None`):
+                Pre-generated text embeddings to be used as inputs for image generation. Can be used in place of
+                `prompt` to avoid re-computing the embeddings. If not provided, the embeddings will be generated from
+                the supplied `prompt`.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if text_embeddings is None:
+            if isinstance(prompt, str):
+                batch_size = 1
+            elif isinstance(prompt, list):
+                batch_size = len(prompt)
+            else:
+                raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+            # get prompt text embeddings
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+
+            if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+                removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+                print(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+                text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+            text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+        else:
+            batch_size = text_embeddings.shape[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = self.tokenizer.model_max_length
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+    def embed_text(self, text):
+        """takes in text and turns it into text embeddings"""
+        text_input = self.tokenizer(
+            text,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        with torch.no_grad():
+            embed = self.text_encoder(text_input.input_ids.to(self.device))[0]
+        return embed
+
+    def get_noise(self, seed, dtype=torch.float32, height=512, width=512):
+        """Takes in random seed and returns corresponding noise vector"""
+        return torch.randn(
+            (1, self.unet.in_channels, height // 8, width // 8),
+            generator=torch.Generator(device=self.device).manual_seed(seed),
+            device=self.device,
+            dtype=dtype,
+        )
+
+    def walk(
+        self,
+        prompts: List[str],
+        seeds: List[int],
+        num_interpolation_steps: Optional[int] = 6,
+        output_dir: Optional[str] = "./dreams",
+        name: Optional[str] = None,
+        batch_size: Optional[int] = 1,
+        height: Optional[int] = 512,
+        width: Optional[int] = 512,
+        guidance_scale: Optional[float] = 7.5,
+        num_inference_steps: Optional[int] = 50,
+        eta: Optional[float] = 0.0,
+    ) -> List[str]:
+        """
+        Walks through a series of prompts and seeds, interpolating between them and saving the results to disk.
+
+        Args:
+            prompts (`List[str]`):
+                List of prompts to generate images for.
+            seeds (`List[int]`):
+                List of seeds corresponding to provided prompts. Must be the same length as prompts.
+            num_interpolation_steps (`int`, *optional*, defaults to 6):
+                Number of interpolation steps to take between prompts.
+            output_dir (`str`, *optional*, defaults to `./dreams`):
+                Directory to save the generated images to.
+            name (`str`, *optional*, defaults to `None`):
+                Subdirectory of `output_dir` to save the generated images to. If `None`, the name will
+                be the current time.
+            batch_size (`int`, *optional*, defaults to 1):
+                Number of images to generate at once.
+            height (`int`, *optional*, defaults to 512):
+                Height of the generated images.
+            width (`int`, *optional*, defaults to 512):
+                Width of the generated images.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+
+        Returns:
+            `List[str]`: List of paths to the generated images.
+        """
+        if not len(prompts) == len(seeds):
+            raise ValueError(
+                f"Number of prompts and seeds must be equalGot {len(prompts)} prompts and {len(seeds)} seeds"
+            )
+
+        name = name or time.strftime("%Y%m%d-%H%M%S")
+        save_path = Path(output_dir) / name
+        save_path.mkdir(exist_ok=True, parents=True)
+
+        frame_idx = 0
+        frame_filepaths = []
+        for prompt_a, prompt_b, seed_a, seed_b in zip(prompts, prompts[1:], seeds, seeds[1:]):
+            # Embed Text
+            embed_a = self.embed_text(prompt_a)
+            embed_b = self.embed_text(prompt_b)
+
+            # Get Noise
+            noise_dtype = embed_a.dtype
+            noise_a = self.get_noise(seed_a, noise_dtype, height, width)
+            noise_b = self.get_noise(seed_b, noise_dtype, height, width)
+
+            noise_batch, embeds_batch = None, None
+            T = np.linspace(0.0, 1.0, num_interpolation_steps)
+            for i, t in enumerate(T):
+                noise = slerp(float(t), noise_a, noise_b)
+                embed = torch.lerp(embed_a, embed_b, t)
+
+                noise_batch = noise if noise_batch is None else torch.cat([noise_batch, noise], dim=0)
+                embeds_batch = embed if embeds_batch is None else torch.cat([embeds_batch, embed], dim=0)
+
+                batch_is_ready = embeds_batch.shape[0] == batch_size or i + 1 == T.shape[0]
+                if batch_is_ready:
+                    outputs = self(
+                        latents=noise_batch,
+                        text_embeddings=embeds_batch,
+                        height=height,
+                        width=width,
+                        guidance_scale=guidance_scale,
+                        eta=eta,
+                        num_inference_steps=num_inference_steps,
+                    )
+                    noise_batch, embeds_batch = None, None
+
+                    for image in outputs["images"]:
+                        frame_filepath = str(save_path / f"frame_{frame_idx:06d}.png")
+                        image.save(frame_filepath)
+                        frame_filepaths.append(frame_filepath)
+                        frame_idx += 1
+        return frame_filepaths

examples/community/multilingual_stable_diffusion.py

@@ -0,0 +1,436 @@
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, logging
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    MBart50TokenizerFast,
+    MBartForConditionalGeneration,
+    pipeline,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def detect_language(pipe, prompt, batch_size):
+    """helper function to detect language(s) of prompt"""
+
+    if batch_size == 1:
+        preds = pipe(prompt, top_k=1, truncation=True, max_length=128)
+        return preds[0]["label"]
+    else:
+        detected_languages = []
+        for p in prompt:
+            preds = pipe(p, top_k=1, truncation=True, max_length=128)
+            detected_languages.append(preds[0]["label"])
+
+        return detected_languages
+
+
+def translate_prompt(prompt, translation_tokenizer, translation_model, device):
+    """helper function to translate prompt to English"""
+
+    encoded_prompt = translation_tokenizer(prompt, return_tensors="pt").to(device)
+    generated_tokens = translation_model.generate(**encoded_prompt, max_new_tokens=1000)
+    en_trans = translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+
+    return en_trans[0]
+
+
+class MultilingualStableDiffusion(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion in different languages.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        detection_pipeline ([`pipeline`]):
+            Transformers pipeline to detect prompt's language.
+        translation_model ([`MBartForConditionalGeneration`]):
+            Model to translate prompt to English, if necessary. Please refer to the
+            [model card](https://huggingface.co/docs/transformers/model_doc/mbart) for details.
+        translation_tokenizer ([`MBart50TokenizerFast`]):
+            Tokenizer of the translation model.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        detection_pipeline: pipeline,
+        translation_model: MBartForConditionalGeneration,
+        translation_tokenizer: MBart50TokenizerFast,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            detection_pipeline=detection_pipeline,
+            translation_model=translation_model,
+            translation_tokenizer=translation_tokenizer,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation. Can be in different languages.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        # detect language and translate if necessary
+        prompt_language = detect_language(self.detection_pipeline, prompt, batch_size)
+        if batch_size == 1 and prompt_language != "en":
+            prompt = translate_prompt(prompt, self.translation_tokenizer, self.translation_model, self.device)
+
+        if isinstance(prompt, list):
+            for index in range(batch_size):
+                if prompt_language[index] != "en":
+                    p = translate_prompt(
+                        prompt[index], self.translation_tokenizer, self.translation_model, self.device
+                    )
+                    prompt[index] = p
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                # detect language and translate it if necessary
+                negative_prompt_language = detect_language(self.detection_pipeline, negative_prompt, batch_size)
+                if negative_prompt_language != "en":
+                    negative_prompt = translate_prompt(
+                        negative_prompt, self.translation_tokenizer, self.translation_model, self.device
+                    )
+                if isinstance(negative_prompt, str):
+                    uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                # detect language and translate it if necessary
+                if isinstance(negative_prompt, list):
+                    negative_prompt_languages = detect_language(self.detection_pipeline, negative_prompt, batch_size)
+                    for index in range(batch_size):
+                        if negative_prompt_languages[index] != "en":
+                            p = translate_prompt(
+                                negative_prompt[index], self.translation_tokenizer, self.translation_model, self.device
+                            )
+                            negative_prompt[index] = p
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/one_step_unet.py

@@ -0,0 +1,22 @@
+#!/usr/bin/env python3
+import torch
+
+from diffusers import DiffusionPipeline
+
+
+class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+
+        self.register_modules(unet=unet, scheduler=scheduler)
+
+    def __call__(self):
+        image = torch.randn(
+            (1, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size),
+        )
+        timestep = 1
+
+        model_output = self.unet(image, timestep).sample
+        scheduler_output = self.scheduler.step(model_output, timestep, image).prev_sample
+
+        return scheduler_output

examples/community/sd_text2img_k_diffusion.py

@@ -0,0 +1,479 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from diffusers import LMSDiscreteScheduler
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.utils import is_accelerate_available, logging
+from k_diffusion.external import CompVisDenoiser
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class ModelWrapper:
+    def __init__(self, model, alphas_cumprod):
+        self.model = model
+        self.alphas_cumprod = alphas_cumprod
+
+    def apply_model(self, *args, **kwargs):
+        return self.model(*args, **kwargs).sample
+
+
+class StableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae,
+        text_encoder,
+        tokenizer,
+        unet,
+        scheduler,
+        safety_checker,
+        feature_extractor,
+    ):
+        super().__init__()
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        # get correct sigmas from LMS
+        scheduler = LMSDiscreteScheduler.from_config(scheduler.config)
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+        model = ModelWrapper(unet, scheduler.alphas_cumprod)
+        self.k_diffusion_model = CompVisDenoiser(model)
+
+    def set_sampler(self, scheduler_type: str):
+        library = importlib.import_module("k_diffusion")
+        sampling = getattr(library, "sampling")
+        self.sampler = getattr(sampling, scheduler_type)
+
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(True)
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(False)
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `list(int)`):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+        """
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+        if not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+
+        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        else:
+            has_nsfw_concept = None
+        return image, has_nsfw_concept
+
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    def check_inputs(self, prompt, height, width, callback_steps):
+        if not isinstance(prompt, str) and not isinstance(prompt, list):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // 8, width // 8)
+        if latents is None:
+            if device.type == "mps":
+                # randn does not work reproducibly on mps
+                latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
+            else:
+                latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, callback_steps)
+
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(prompt, str) else len(prompt)
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = True
+        if guidance_scale <= 1.0:
+            raise ValueError("has to use guidance_scale")
+
+        # 3. Encode input prompt
+        text_embeddings = self._encode_prompt(
+            prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=text_embeddings.device)
+        sigmas = self.scheduler.sigmas
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            text_embeddings.dtype,
+            device,
+            generator,
+            latents,
+        )
+        latents = latents * sigmas[0]
+        self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device)
+        self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device)
+
+        def model_fn(x, t):
+            latent_model_input = torch.cat([x] * 2)
+
+            noise_pred = self.k_diffusion_model(latent_model_input, t, encoder_hidden_states=text_embeddings)
+
+            noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+            return noise_pred
+
+        latents = self.sampler(model_fn, latents, sigmas)
+
+        # 8. Post-processing
+        image = self.decode_latents(latents)
+
+        # 9. Run safety checker
+        image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype)
+
+        # 10. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/seed_resize_stable_diffusion.py

@@ -0,0 +1,366 @@
+"""
+    modified based on diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+"""
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import logging
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class SeedResizeStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        text_embeddings: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+
+        if text_embeddings is None:
+            text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""]
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_shape_reference = (batch_size * num_images_per_prompt, self.unet.in_channels, 64, 64)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not exist on mps
+                latents_reference = torch.randn(
+                    latents_shape_reference, generator=generator, device="cpu", dtype=latents_dtype
+                ).to(self.device)
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents_reference = torch.randn(
+                    latents_shape_reference, generator=generator, device=self.device, dtype=latents_dtype
+                )
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents_reference.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents_reference = latents_reference.to(self.device)
+            latents = latents.to(self.device)
+
+        # This is the key part of the pipeline where we
+        # try to ensure that the generated images w/ the same seed
+        # but different sizes actually result in similar images
+        dx = (latents_shape[3] - latents_shape_reference[3]) // 2
+        dy = (latents_shape[2] - latents_shape_reference[2]) // 2
+        w = latents_shape_reference[3] if dx >= 0 else latents_shape_reference[3] + 2 * dx
+        h = latents_shape_reference[2] if dy >= 0 else latents_shape_reference[2] + 2 * dy
+        tx = 0 if dx < 0 else dx
+        ty = 0 if dy < 0 else dy
+        dx = max(-dx, 0)
+        dy = max(-dy, 0)
+        # import pdb
+        # pdb.set_trace()
+        latents[:, :, ty : ty + h, tx : tx + w] = latents_reference[:, :, dy : dy + h, dx : dx + w]
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/community/speech_to_image_diffusion.py

@@ -0,0 +1,261 @@
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DiffusionPipeline,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    UNet2DConditionModel,
+)
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.utils import logging
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    WhisperForConditionalGeneration,
+    WhisperProcessor,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class SpeechToImagePipeline(DiffusionPipeline):
+    def __init__(
+        self,
+        speech_model: WhisperForConditionalGeneration,
+        speech_processor: WhisperProcessor,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            speech_model=speech_model,
+            speech_processor=speech_processor,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        if slice_size == "auto":
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        audio,
+        sampling_rate=16_000,
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        inputs = self.speech_processor.feature_extractor(
+            audio, return_tensors="pt", sampling_rate=sampling_rate
+        ).input_features.to(self.device)
+        predicted_ids = self.speech_model.generate(inputs, max_length=480_000)
+
+        prompt = self.speech_processor.tokenizer.batch_decode(predicted_ids, skip_special_tokens=True, normalize=True)[
+            0
+        ]
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not exist on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return image
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)

examples/community/stable_diffusion_mega.py

@@ -0,0 +1,224 @@
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+
+import PIL.Image
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DiffusionPipeline,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    StableDiffusionImg2ImgPipeline,
+    StableDiffusionInpaintPipelineLegacy,
+    StableDiffusionPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.configuration_utils import FrozenDict
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.utils import deprecate, logging
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class StableDiffusionMegaPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionMegaSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    @property
+    def components(self) -> Dict[str, Any]:
+        return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    @torch.no_grad()
+    def inpaint(
+        self,
+        prompt: Union[str, List[str]],
+        init_image: Union[torch.FloatTensor, PIL.Image.Image],
+        mask_image: Union[torch.FloatTensor, PIL.Image.Image],
+        strength: float = 0.8,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        # For more information on how this function works, please see: https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion#diffusers.StableDiffusionImg2ImgPipeline
+        return StableDiffusionInpaintPipelineLegacy(**self.components)(
+            prompt=prompt,
+            init_image=init_image,
+            mask_image=mask_image,
+            strength=strength,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+        )
+
+    @torch.no_grad()
+    def img2img(
+        self,
+        prompt: Union[str, List[str]],
+        init_image: Union[torch.FloatTensor, PIL.Image.Image],
+        strength: float = 0.8,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        # For more information on how this function works, please see: https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion#diffusers.StableDiffusionImg2ImgPipeline
+        return StableDiffusionImg2ImgPipeline(**self.components)(
+            prompt=prompt,
+            init_image=init_image,
+            strength=strength,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+        )
+
+    @torch.no_grad()
+    def text2img(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        # For more information on how this function https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion#diffusers.StableDiffusionPipeline
+        return StableDiffusionPipeline(**self.components)(
+            prompt=prompt,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+        )

examples/community/text_inpainting.py

@@ -0,0 +1,320 @@
+from typing import Callable, List, Optional, Union
+
+import torch
+
+import PIL
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, is_accelerate_available, logging
+from transformers import (
+    CLIPFeatureExtractor,
+    CLIPSegForImageSegmentation,
+    CLIPSegProcessor,
+    CLIPTextModel,
+    CLIPTokenizer,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class TextInpainting(DiffusionPipeline):
+    r"""
+    Pipeline for text based inpainting using Stable Diffusion.
+    Uses CLIPSeg to get a mask from the given text, then calls the Inpainting pipeline with the generated mask
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        segmentation_model ([`CLIPSegForImageSegmentation`]):
+            CLIPSeg Model to generate mask from the given text. Please refer to the [model card]() for details.
+        segmentation_processor ([`CLIPSegProcessor`]):
+            CLIPSeg processor to get image, text features to translate prompt to English, if necessary. Please refer to the
+            [model card](https://huggingface.co/docs/transformers/model_doc/clipseg) for details.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        segmentation_model: CLIPSegForImageSegmentation,
+        segmentation_processor: CLIPSegProcessor,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if hasattr(scheduler.config, "skip_prk_steps") and scheduler.config.skip_prk_steps is False:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration"
+                " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make"
+                " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to"
+                " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face"
+                " Hub, it would be very nice if you could open a Pull request for the"
+                " `scheduler/scheduler_config.json` file"
+            )
+            deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["skip_prk_steps"] = True
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            segmentation_model=segmentation_model,
+            segmentation_processor=segmentation_processor,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device("cuda")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(True)
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(False)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        image: Union[torch.FloatTensor, PIL.Image.Image],
+        text: str,
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            image (`PIL.Image.Image`):
+                `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+                be masked out with `mask_image` and repainted according to `prompt`.
+            text (`str``):
+                The text to use to generate the mask.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # We use the input text to generate the mask
+        inputs = self.segmentation_processor(
+            text=[text], images=[image], padding="max_length", return_tensors="pt"
+        ).to(self.device)
+        outputs = self.segmentation_model(**inputs)
+        mask = torch.sigmoid(outputs.logits).cpu().detach().unsqueeze(-1).numpy()
+        mask_pil = self.numpy_to_pil(mask)[0].resize(image.size)
+
+        # Run inpainting pipeline with the generated mask
+        inpainting_pipeline = StableDiffusionInpaintPipeline(
+            vae=self.vae,
+            text_encoder=self.text_encoder,
+            tokenizer=self.tokenizer,
+            unet=self.unet,
+            scheduler=self.scheduler,
+            safety_checker=self.safety_checker,
+            feature_extractor=self.feature_extractor,
+        )
+        return inpainting_pipeline(
+            prompt=prompt,
+            image=image,
+            mask_image=mask_pil,
+            height=height,
+            width=width,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            eta=eta,
+            generator=generator,
+            latents=latents,
+            output_type=output_type,
+            return_dict=return_dict,
+            callback=callback,
+            callback_steps=callback_steps,
+        )

examples/community/wildcard_stable_diffusion.py

@@ -0,0 +1,418 @@
+import inspect
+import os
+import random
+import re
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from diffusers.utils import deprecate, logging
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+global_re_wildcard = re.compile(r"__([^_]*)__")
+
+
+def get_filename(path: str):
+    # this doesn't work on Windows
+    return os.path.basename(path).split(".txt")[0]
+
+
+def read_wildcard_values(path: str):
+    with open(path, encoding="utf8") as f:
+        return f.read().splitlines()
+
+
+def grab_wildcard_values(wildcard_option_dict: Dict[str, List[str]] = {}, wildcard_files: List[str] = []):
+    for wildcard_file in wildcard_files:
+        filename = get_filename(wildcard_file)
+        read_values = read_wildcard_values(wildcard_file)
+        if filename not in wildcard_option_dict:
+            wildcard_option_dict[filename] = []
+        wildcard_option_dict[filename].extend(read_values)
+    return wildcard_option_dict
+
+
+def replace_prompt_with_wildcards(
+    prompt: str, wildcard_option_dict: Dict[str, List[str]] = {}, wildcard_files: List[str] = []
+):
+    new_prompt = prompt
+
+    # get wildcard options
+    wildcard_option_dict = grab_wildcard_values(wildcard_option_dict, wildcard_files)
+
+    for m in global_re_wildcard.finditer(new_prompt):
+        wildcard_value = m.group()
+        replace_value = random.choice(wildcard_option_dict[wildcard_value.strip("__")])
+        new_prompt = new_prompt.replace(wildcard_value, replace_value, 1)
+
+    return new_prompt
+
+
+@dataclass
+class WildcardStableDiffusionOutput(StableDiffusionPipelineOutput):
+    prompts: List[str]
+
+
+class WildcardStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Example Usage:
+        pipe = WildcardStableDiffusionPipeline.from_pretrained(
+            "CompVis/stable-diffusion-v1-4",
+            revision="fp16",
+            torch_dtype=torch.float16,
+        )
+        prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
+        out = pipe(
+            prompt,
+            wildcard_option_dict={
+                "clothing":["hat", "shirt", "scarf", "beret"]
+            },
+            wildcard_files=["object.txt", "animal.txt"],
+            num_prompt_samples=1
+        )
+
+
+    Pipeline for text-to-image generation with wild cards using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: int = 512,
+        width: int = 512,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        wildcard_option_dict: Dict[str, List[str]] = {},
+        wildcard_files: List[str] = [],
+        num_prompt_samples: Optional[int] = 1,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to 512):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 512):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            wildcard_option_dict (Dict[str, List[str]]):
+                dict with key as `wildcard` and values as a list of possible replacements. For example if a prompt, "A __animal__ sitting on a chair". A wildcard_option_dict can provide possible values for "animal" like this: {"animal":["dog", "cat", "fox"]}
+            wildcard_files: (List[str])
+               List of filenames of txt files for wildcard replacements. For example if a prompt, "A __animal__ sitting on a chair". A file can be provided ["animal.txt"]
+            num_prompt_samples: int
+                Number of times to sample wildcards for each prompt provided
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        if isinstance(prompt, str):
+            prompt = [
+                replace_prompt_with_wildcards(prompt, wildcard_option_dict, wildcard_files)
+                for i in range(num_prompt_samples)
+            ]
+            batch_size = len(prompt)
+        elif isinstance(prompt, list):
+            prompt_list = []
+            for p in prompt:
+                for i in range(num_prompt_samples):
+                    prompt_list.append(replace_prompt_with_wildcards(p, wildcard_option_dict, wildcard_files))
+            prompt = prompt_list
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # get the initial random noise unless the user supplied it
+
+        # Unlike in other pipelines, latents need to be generated in the target device
+        # for 1-to-1 results reproducibility with the CompVis implementation.
+        # However this currently doesn't work in `mps`.
+        latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
+        latents_dtype = text_embeddings.dtype
+        if latents is None:
+            if self.device.type == "mps":
+                # randn does not exist on mps
+                latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
+                    self.device
+                )
+            else:
+                latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+        else:
+            if latents.shape != latents_shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+            latents = latents.to(self.device)
+
+        # set timesteps
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        # Some schedulers like PNDM have timesteps as arrays
+        # It's more optimized to move all timesteps to correct device beforehand
+        timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
+                self.device
+            )
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
+            )
+        else:
+            has_nsfw_concept = None
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return WildcardStableDiffusionOutput(images=image, nsfw_content_detected=has_nsfw_concept, prompts=prompt)

examples/dreambooth/README.md

@@ -4,13 +4,12 @@
 The `train_dreambooth.py` script shows how to implement the training procedure and adapt it for stable diffusion.
 
 
-## Running locally 
+## Running locally with PyTorch
 ### Installing the dependencies
 
 Before running the scripts, make sure to install the library's training dependencies:
 
 ```bash
-pip install git+https://github.com/huggingface/diffusers.git
 pip install -U -r requirements.txt
 ```
 
@@ -88,11 +87,12 @@ accelerate launch train_dreambooth.py \
   --max_train_steps=800
 ```
 
+
 ### Training on a 16GB GPU:
 
 With the help of gradient checkpointing and the 8-bit optimizer from bitsandbytes it's possible to run train dreambooth on a 16GB GPU.
 
-Install `bitsandbytes` with `pip install bitsandbytes`
+To install `bitandbytes` please refer to this [readme](https://github.com/TimDettmers/bitsandbytes#requirements--installation).
 
 ```bash
 export MODEL_NAME="CompVis/stable-diffusion-v1-4"
@@ -119,8 +119,82 @@ accelerate launch train_dreambooth.py \
   --max_train_steps=800
 ```
 
+### Training on a 8 GB GPU:
 
-## Inference
+By using [DeepSpeed](https://www.deepspeed.ai/) it's possible to offload some
+tensors from VRAM to either CPU or NVME allowing to train with less VRAM.
+
+DeepSpeed needs to be enabled with `accelerate config`. During configuration
+answer yes to "Do you want to use DeepSpeed?". With DeepSpeed stage 2, fp16
+mixed precision and offloading both parameters and optimizer state to cpu it's
+possible to train on under 8 GB VRAM with a drawback of requiring significantly
+more RAM (about 25 GB). See [documentation](https://huggingface.co/docs/accelerate/usage_guides/deepspeed) for more DeepSpeed configuration options.
+
+Changing the default Adam optimizer to DeepSpeed's special version of Adam
+`deepspeed.ops.adam.DeepSpeedCPUAdam` gives a substantial speedup but enabling
+it requires CUDA toolchain with the same version as pytorch. 8-bit optimizer
+does not seem to be compatible with DeepSpeed at the moment.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --sample_batch_size=1 \
+  --gradient_accumulation_steps=1 --gradient_checkpointing \
+  --learning_rate=5e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800 \
+  --mixed_precision=fp16
+```
+
+### Fine-tune text encoder with the UNet.
+
+The script also allows to fine-tune the `text_encoder` along with the `unet`. It's been observed experimentally that fine-tuning `text_encoder` gives much better results especially on faces. 
+Pass the `--train_text_encoder` argument to the script to enable training `text_encoder`.
+
+___Note: Training text encoder requires more memory, with this option the training won't fit on 16GB GPU. It needs at least 24GB VRAM.___
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --train_text_encoder \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --use_8bit_adam \
+  --gradient_checkpointing \
+  --learning_rate=2e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
+### Inference
 
 Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline`. Make sure to include the `identifier`(e.g. sks in above example) in your prompt.
 
@@ -136,3 +210,85 @@ image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
 
 image.save("dog-bucket.png")
 ```
+
+
+## Running with Flax/JAX
+
+For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.
+
+____Note: The flax example don't yet support features like gradient checkpoint, gradient accumulation etc, so to use flax for faster training we will need >30GB cards.___
+
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+```bash
+pip install -U -r requirements_flax.txt
+```
+
+
+### Training without prior preservation loss
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export INSTANCE_DIR="path-to-instance-images"
+export OUTPUT_DIR="path-to-save-model"
+
+python train_dreambooth_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --learning_rate=5e-6 \
+  --max_train_steps=400
+```
+
+
+### Training with prior preservation loss
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+
+python train_dreambooth_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --learning_rate=5e-6 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
+
+### Fine-tune text encoder with the UNet.
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+
+python train_dreambooth_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --train_text_encoder \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --learning_rate=2e-6 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```

examples/dreambooth/requirements.txt

@@ -1,3 +1,4 @@
+diffusers>==0.5.0
 accelerate
 torchvision
 transformers>=4.21.0

examples/dreambooth/requirements_flax.txt

@@ -0,0 +1,9 @@
+diffusers>==0.5.1
+transformers>=4.21.0
+flax
+optax
+torch
+torchvision
+ftfy
+tensorboard
+modelcards

examples/dreambooth/train_dreambooth.py

@@ -1,4 +1,6 @@
 import argparse
+import hashlib
+import itertools
 import math
 import os
 from pathlib import Path
@@ -24,7 +26,7 @@ from transformers import CLIPTextModel, CLIPTokenizer
 logger = get_logger(__name__)
 
 
-def parse_args():
+def parse_args(input_args=None):
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
     parser.add_argument(
         "--pretrained_model_name_or_path",
@@ -33,6 +35,13 @@ def parse_args():
         required=True,
         help="Path to pretrained model or model identifier from huggingface.co/models.",
     )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help="Revision of pretrained model identifier from huggingface.co/models.",
+    )
     parser.add_argument(
         "--tokenizer_name",
         type=str,
@@ -57,6 +66,7 @@ def parse_args():
         "--instance_prompt",
         type=str,
         default=None,
+        required=True,
         help="The prompt with identifier specifying the instance",
     )
     parser.add_argument(
@@ -77,7 +87,7 @@ def parse_args():
         type=int,
         default=100,
         help=(
-            "Minimal class images for prior perversation loss. If not have enough images, additional images will be"
+            "Minimal class images for prior preservation loss. If not have enough images, additional images will be"
             " sampled with class_prompt."
         ),
     )
@@ -100,6 +110,7 @@ def parse_args():
     parser.add_argument(
         "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution"
     )
+    parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder")
     parser.add_argument(
         "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
     )
@@ -186,19 +197,25 @@ def parse_args():
     )
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
 
-    args = parser.parse_args()
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
     if env_local_rank != -1 and env_local_rank != args.local_rank:
         args.local_rank = env_local_rank
 
-    if args.instance_data_dir is None:
-        raise ValueError("You must specify a train data directory.")
-
     if args.with_prior_preservation:
         if args.class_data_dir is None:
             raise ValueError("You must specify a data directory for class images.")
         if args.class_prompt is None:
             raise ValueError("You must specify prompt for class images.")
+    else:
+        if args.class_data_dir is not None:
+            logger.warning("You need not use --class_data_dir without --with_prior_preservation.")
+        if args.class_prompt is not None:
+            logger.warning("You need not use --class_prompt without --with_prior_preservation.")
 
     return args
 
@@ -309,8 +326,7 @@ def get_full_repo_name(model_id: str, organization: Optional[str] = None, token:
         return f"{organization}/{model_id}"
 
 
-def main():
-    args = parse_args()
+def main(args):
     logging_dir = Path(args.output_dir, args.logging_dir)
 
     accelerator = Accelerator(
@@ -320,6 +336,15 @@ def main():
         logging_dir=logging_dir,
     )
 
+    # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
+    # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
+    # TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate.
+    if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
+        raise ValueError(
+            "Gradient accumulation is not supported when training the text encoder in distributed training. "
+            "Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
+        )
+
     if args.seed is not None:
         set_seed(args.seed)
 
@@ -332,7 +357,10 @@ def main():
         if cur_class_images < args.num_class_images:
             torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
             pipeline = StableDiffusionPipeline.from_pretrained(
-                args.pretrained_model_name_or_path, torch_dtype=torch_dtype
+                args.pretrained_model_name_or_path,
+                torch_dtype=torch_dtype,
+                safety_checker=None,
+                revision=args.revision,
             )
             pipeline.set_progress_bar_config(disable=True)
 
@@ -351,7 +379,9 @@ def main():
                 images = pipeline(example["prompt"]).images
 
                 for i, image in enumerate(images):
-                    image.save(class_images_dir / f"{example['index'][i] + cur_class_images}.jpg")
+                    hash_image = hashlib.sha1(image.tobytes()).hexdigest()
+                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
+                    image.save(image_filename)
 
             del pipeline
             if torch.cuda.is_available():
@@ -376,17 +406,42 @@ def main():
 
     # Load the tokenizer
     if args.tokenizer_name:
-        tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name)
+        tokenizer = CLIPTokenizer.from_pretrained(
+            args.tokenizer_name,
+            revision=args.revision,
+        )
     elif args.pretrained_model_name_or_path:
-        tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+        tokenizer = CLIPTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+        )
 
     # Load models and create wrapper for stable diffusion
-    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
-    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
-    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    text_encoder = CLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="text_encoder",
+        revision=args.revision,
+    )
+    vae = AutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="vae",
+        revision=args.revision,
+    )
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="unet",
+        revision=args.revision,
+    )
+
+    vae.requires_grad_(False)
+    if not args.train_text_encoder:
+        text_encoder.requires_grad_(False)
 
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
+        if args.train_text_encoder:
+            text_encoder.gradient_checkpointing_enable()
 
     if args.scale_lr:
         args.learning_rate = (
@@ -406,17 +461,18 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
+    params_to_optimize = (
+        itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters()
+    )
     optimizer = optimizer_class(
-        unet.parameters(),  # only optimize unet
+        params_to_optimize,
         lr=args.learning_rate,
         betas=(args.adam_beta1, args.adam_beta2),
         weight_decay=args.adam_weight_decay,
         eps=args.adam_epsilon,
     )
 
-    noise_scheduler = DDPMScheduler(
-        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000
-    )
+    noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler")
 
     train_dataset = DreamBoothDataset(
         instance_data_root=args.instance_data_dir,
@@ -441,7 +497,12 @@ def main():
         pixel_values = torch.stack(pixel_values)
         pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
 
-        input_ids = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt").input_ids
+        input_ids = tokenizer.pad(
+            {"input_ids": input_ids},
+            padding="max_length",
+            max_length=tokenizer.model_max_length,
+            return_tensors="pt",
+        ).input_ids
 
         batch = {
             "input_ids": input_ids,
@@ -450,7 +511,7 @@ def main():
         return batch
 
     train_dataloader = torch.utils.data.DataLoader(
-        train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=collate_fn
+        train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=collate_fn, num_workers=1
     )
 
     # Scheduler and math around the number of training steps.
@@ -467,13 +528,27 @@ def main():
         num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
     )
 
-    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, optimizer, train_dataloader, lr_scheduler
-    )
+    if args.train_text_encoder:
+        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+        )
+    else:
+        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, optimizer, train_dataloader, lr_scheduler
+        )
 
-    # Move text_encode and vae to gpu
-    text_encoder.to(accelerator.device)
-    vae.to(accelerator.device)
+    weight_dtype = torch.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move text_encode and vae to gpu.
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models are only used for inference, keeping weights in full precision is not required.
+    vae.to(accelerator.device, dtype=weight_dtype)
+    if not args.train_text_encoder:
+        text_encoder.to(accelerator.device, dtype=weight_dtype)
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -505,15 +580,16 @@ def main():
 
     for epoch in range(args.num_train_epochs):
         unet.train()
+        if args.train_text_encoder:
+            text_encoder.train()
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
                 # Convert images to latent space
-                with torch.no_grad():
-                    latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
-                    latents = latents * 0.18215
+                latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
+                latents = latents * 0.18215
 
                 # Sample noise that we'll add to the latents
-                noise = torch.randn(latents.shape).to(latents.device)
+                noise = torch.randn_like(latents)
                 bsz = latents.shape[0]
                 # Sample a random timestep for each image
                 timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
@@ -524,8 +600,7 @@ def main():
                 noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
 
                 # Get the text embedding for conditioning
-                with torch.no_grad():
-                    encoder_hidden_states = text_encoder(batch["input_ids"])[0]
+                encoder_hidden_states = text_encoder(batch["input_ids"])[0]
 
                 # Predict the noise residual
                 noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
@@ -536,19 +611,24 @@ def main():
                     noise, noise_prior = torch.chunk(noise, 2, dim=0)
 
                     # Compute instance loss
-                    loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                    loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="none").mean([1, 2, 3]).mean()
 
                     # Compute prior loss
-                    prior_loss = F.mse_loss(noise_pred_prior, noise_prior, reduction="none").mean([1, 2, 3]).mean()
+                    prior_loss = F.mse_loss(noise_pred_prior.float(), noise_prior.float(), reduction="mean")
 
                     # Add the prior loss to the instance loss.
                     loss = loss + args.prior_loss_weight * prior_loss
                 else:
-                    loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                    loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+                    params_to_clip = (
+                        itertools.chain(unet.parameters(), text_encoder.parameters())
+                        if args.train_text_encoder
+                        else unet.parameters()
+                    )
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -570,17 +650,19 @@ def main():
     # Create the pipeline using using the trained modules and save it.
     if accelerator.is_main_process:
         pipeline = StableDiffusionPipeline.from_pretrained(
-            args.pretrained_model_name_or_path, unet=accelerator.unwrap_model(unet)
+            args.pretrained_model_name_or_path,
+            unet=accelerator.unwrap_model(unet),
+            text_encoder=accelerator.unwrap_model(text_encoder),
+            revision=args.revision,
         )
         pipeline.save_pretrained(args.output_dir)
 
         if args.push_to_hub:
-            repo.push_to_hub(
-                args, pipeline, repo, commit_message="End of training", blocking=False, auto_lfs_prune=True
-            )
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
 
     accelerator.end_training()
 
 
 if __name__ == "__main__":
-    main()
+    args = parse_args()
+    main(args)

examples/dreambooth/train_dreambooth_flax.py

@@ -0,0 +1,652 @@
+import argparse
+import hashlib
+import logging
+import math
+import os
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch.utils.data import Dataset
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from diffusers import (
+    FlaxAutoencoderKL,
+    FlaxDDPMScheduler,
+    FlaxPNDMScheduler,
+    FlaxStableDiffusionPipeline,
+    FlaxUNet2DConditionModel,
+)
+from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionSafetyChecker
+from flax import jax_utils
+from flax.training import train_state
+from flax.training.common_utils import shard
+from huggingface_hub import HfFolder, Repository, whoami
+from PIL import Image
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel, set_seed
+
+
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--instance_data_dir",
+        type=str,
+        default=None,
+        required=True,
+        help="A folder containing the training data of instance images.",
+    )
+    parser.add_argument(
+        "--class_data_dir",
+        type=str,
+        default=None,
+        required=False,
+        help="A folder containing the training data of class images.",
+    )
+    parser.add_argument(
+        "--instance_prompt",
+        type=str,
+        default=None,
+        help="The prompt with identifier specifying the instance",
+    )
+    parser.add_argument(
+        "--class_prompt",
+        type=str,
+        default=None,
+        help="The prompt to specify images in the same class as provided instance images.",
+    )
+    parser.add_argument(
+        "--with_prior_preservation",
+        default=False,
+        action="store_true",
+        help="Flag to add prior preservation loss.",
+    )
+    parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
+    parser.add_argument(
+        "--num_class_images",
+        type=int,
+        default=100,
+        help=(
+            "Minimal class images for prior preservation loss. If not have enough images, additional images will be"
+            " sampled with class_prompt."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="text-inversion-model",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--seed", type=int, default=0, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution"
+    )
+    parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder")
+    parser.add_argument(
+        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=1)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=5e-6,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.instance_data_dir is None:
+        raise ValueError("You must specify a train data directory.")
+
+    if args.with_prior_preservation:
+        if args.class_data_dir is None:
+            raise ValueError("You must specify a data directory for class images.")
+        if args.class_prompt is None:
+            raise ValueError("You must specify prompt for class images.")
+
+    return args
+
+
+class DreamBoothDataset(Dataset):
+    """
+    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
+    It pre-processes the images and the tokenizes prompts.
+    """
+
+    def __init__(
+        self,
+        instance_data_root,
+        instance_prompt,
+        tokenizer,
+        class_data_root=None,
+        class_prompt=None,
+        size=512,
+        center_crop=False,
+    ):
+        self.size = size
+        self.center_crop = center_crop
+        self.tokenizer = tokenizer
+
+        self.instance_data_root = Path(instance_data_root)
+        if not self.instance_data_root.exists():
+            raise ValueError("Instance images root doesn't exists.")
+
+        self.instance_images_path = list(Path(instance_data_root).iterdir())
+        self.num_instance_images = len(self.instance_images_path)
+        self.instance_prompt = instance_prompt
+        self._length = self.num_instance_images
+
+        if class_data_root is not None:
+            self.class_data_root = Path(class_data_root)
+            self.class_data_root.mkdir(parents=True, exist_ok=True)
+            self.class_images_path = list(self.class_data_root.iterdir())
+            self.num_class_images = len(self.class_images_path)
+            self._length = max(self.num_class_images, self.num_instance_images)
+            self.class_prompt = class_prompt
+        else:
+            self.class_data_root = None
+
+        self.image_transforms = transforms.Compose(
+            [
+                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, index):
+        example = {}
+        instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
+        if not instance_image.mode == "RGB":
+            instance_image = instance_image.convert("RGB")
+        example["instance_images"] = self.image_transforms(instance_image)
+        example["instance_prompt_ids"] = self.tokenizer(
+            self.instance_prompt,
+            padding="do_not_pad",
+            truncation=True,
+            max_length=self.tokenizer.model_max_length,
+        ).input_ids
+
+        if self.class_data_root:
+            class_image = Image.open(self.class_images_path[index % self.num_class_images])
+            if not class_image.mode == "RGB":
+                class_image = class_image.convert("RGB")
+            example["class_images"] = self.image_transforms(class_image)
+            example["class_prompt_ids"] = self.tokenizer(
+                self.class_prompt,
+                padding="do_not_pad",
+                truncation=True,
+                max_length=self.tokenizer.model_max_length,
+            ).input_ids
+
+        return example
+
+
+class PromptDataset(Dataset):
+    "A simple dataset to prepare the prompts to generate class images on multiple GPUs."
+
+    def __init__(self, prompt, num_samples):
+        self.prompt = prompt
+        self.num_samples = num_samples
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, index):
+        example = {}
+        example["prompt"] = self.prompt
+        example["index"] = index
+        return example
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+def get_params_to_save(params):
+    return jax.device_get(jax.tree_util.tree_map(lambda x: x[0], params))
+
+
+def main():
+    args = parse_args()
+
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    rng = jax.random.PRNGKey(args.seed)
+
+    if args.with_prior_preservation:
+        class_images_dir = Path(args.class_data_dir)
+        if not class_images_dir.exists():
+            class_images_dir.mkdir(parents=True)
+        cur_class_images = len(list(class_images_dir.iterdir()))
+
+        if cur_class_images < args.num_class_images:
+            pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
+                args.pretrained_model_name_or_path, safety_checker=None
+            )
+            pipeline.set_progress_bar_config(disable=True)
+
+            num_new_images = args.num_class_images - cur_class_images
+            logger.info(f"Number of class images to sample: {num_new_images}.")
+
+            sample_dataset = PromptDataset(args.class_prompt, num_new_images)
+            total_sample_batch_size = args.sample_batch_size * jax.local_device_count()
+            sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=total_sample_batch_size)
+
+            for example in tqdm(
+                sample_dataloader, desc="Generating class images", disable=not jax.process_index() == 0
+            ):
+                prompt_ids = pipeline.prepare_inputs(example["prompt"])
+                prompt_ids = shard(prompt_ids)
+                p_params = jax_utils.replicate(params)
+                rng = jax.random.split(rng)[0]
+                sample_rng = jax.random.split(rng, jax.device_count())
+                images = pipeline(prompt_ids, p_params, sample_rng, jit=True).images
+                images = images.reshape((images.shape[0] * images.shape[1],) + images.shape[-3:])
+                images = pipeline.numpy_to_pil(np.array(images))
+
+                for i, image in enumerate(images):
+                    hash_image = hashlib.sha1(image.tobytes()).hexdigest()
+                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
+                    image.save(image_filename)
+
+            del pipeline
+
+    # Handle the repository creation
+    if jax.process_index() == 0:
+        if args.push_to_hub:
+            if args.hub_model_id is None:
+                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
+            else:
+                repo_name = args.hub_model_id
+            repo = Repository(args.output_dir, clone_from=repo_name)
+
+            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "step_*" not in gitignore:
+                    gitignore.write("step_*\n")
+                if "epoch_*" not in gitignore:
+                    gitignore.write("epoch_*\n")
+        elif args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load the tokenizer and add the placeholder token as a additional special token
+    if args.tokenizer_name:
+        tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+
+    train_dataset = DreamBoothDataset(
+        instance_data_root=args.instance_data_dir,
+        instance_prompt=args.instance_prompt,
+        class_data_root=args.class_data_dir if args.with_prior_preservation else None,
+        class_prompt=args.class_prompt,
+        tokenizer=tokenizer,
+        size=args.resolution,
+        center_crop=args.center_crop,
+    )
+
+    def collate_fn(examples):
+        input_ids = [example["instance_prompt_ids"] for example in examples]
+        pixel_values = [example["instance_images"] for example in examples]
+
+        # Concat class and instance examples for prior preservation.
+        # We do this to avoid doing two forward passes.
+        if args.with_prior_preservation:
+            input_ids += [example["class_prompt_ids"] for example in examples]
+            pixel_values += [example["class_images"] for example in examples]
+
+        pixel_values = torch.stack(pixel_values)
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+        input_ids = tokenizer.pad(
+            {"input_ids": input_ids}, padding="max_length", max_length=tokenizer.model_max_length, return_tensors="pt"
+        ).input_ids
+
+        batch = {
+            "input_ids": input_ids,
+            "pixel_values": pixel_values,
+        }
+        batch = {k: v.numpy() for k, v in batch.items()}
+        return batch
+
+    total_train_batch_size = args.train_batch_size * jax.local_device_count()
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset, batch_size=total_train_batch_size, shuffle=True, collate_fn=collate_fn, drop_last=True
+    )
+
+    weight_dtype = jnp.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = jnp.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = jnp.bfloat16
+
+    # Load models and create wrapper for stable diffusion
+    text_encoder = FlaxCLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", dtype=weight_dtype
+    )
+    vae, vae_params = FlaxAutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="vae", dtype=weight_dtype
+    )
+    unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", dtype=weight_dtype
+    )
+
+    # Optimization
+    if args.scale_lr:
+        args.learning_rate = args.learning_rate * total_train_batch_size
+
+    constant_scheduler = optax.constant_schedule(args.learning_rate)
+
+    adamw = optax.adamw(
+        learning_rate=constant_scheduler,
+        b1=args.adam_beta1,
+        b2=args.adam_beta2,
+        eps=args.adam_epsilon,
+        weight_decay=args.adam_weight_decay,
+    )
+
+    optimizer = optax.chain(
+        optax.clip_by_global_norm(args.max_grad_norm),
+        adamw,
+    )
+
+    unet_state = train_state.TrainState.create(apply_fn=unet.__call__, params=unet_params, tx=optimizer)
+    text_encoder_state = train_state.TrainState.create(
+        apply_fn=text_encoder.__call__, params=text_encoder.params, tx=optimizer
+    )
+
+    noise_scheduler = FlaxDDPMScheduler(
+        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000
+    )
+
+    # Initialize our training
+    train_rngs = jax.random.split(rng, jax.local_device_count())
+
+    def train_step(unet_state, text_encoder_state, vae_params, batch, train_rng):
+        dropout_rng, sample_rng, new_train_rng = jax.random.split(train_rng, 3)
+
+        if args.train_text_encoder:
+            params = {"text_encoder": text_encoder_state.params, "unet": unet_state.params}
+        else:
+            params = {"unet": unet_state.params}
+
+        def compute_loss(params):
+            # Convert images to latent space
+            vae_outputs = vae.apply(
+                {"params": vae_params}, batch["pixel_values"], deterministic=True, method=vae.encode
+            )
+            latents = vae_outputs.latent_dist.sample(sample_rng)
+            # (NHWC) -> (NCHW)
+            latents = jnp.transpose(latents, (0, 3, 1, 2))
+            latents = latents * 0.18215
+
+            # Sample noise that we'll add to the latents
+            noise_rng, timestep_rng = jax.random.split(sample_rng)
+            noise = jax.random.normal(noise_rng, latents.shape)
+            # Sample a random timestep for each image
+            bsz = latents.shape[0]
+            timesteps = jax.random.randint(
+                timestep_rng,
+                (bsz,),
+                0,
+                noise_scheduler.config.num_train_timesteps,
+            )
+
+            # Add noise to the latents according to the noise magnitude at each timestep
+            # (this is the forward diffusion process)
+            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+            # Get the text embedding for conditioning
+            if args.train_text_encoder:
+                encoder_hidden_states = text_encoder_state.apply_fn(
+                    batch["input_ids"], params=params["text_encoder"], dropout_rng=dropout_rng, train=True
+                )[0]
+            else:
+                encoder_hidden_states = text_encoder(
+                    batch["input_ids"], params=text_encoder_state.params, train=False
+                )[0]
+
+            # Predict the noise residual
+            unet_outputs = unet.apply(
+                {"params": params["unet"]}, noisy_latents, timesteps, encoder_hidden_states, train=True
+            )
+            noise_pred = unet_outputs.sample
+
+            if args.with_prior_preservation:
+                # Chunk the noise and noise_pred into two parts and compute the loss on each part separately.
+                noise_pred, noise_pred_prior = jnp.split(noise_pred, 2, axis=0)
+                noise, noise_prior = jnp.split(noise, 2, axis=0)
+
+                # Compute instance loss
+                loss = (noise - noise_pred) ** 2
+                loss = loss.mean()
+
+                # Compute prior loss
+                prior_loss = (noise_prior - noise_pred_prior) ** 2
+                prior_loss = prior_loss.mean()
+
+                # Add the prior loss to the instance loss.
+                loss = loss + args.prior_loss_weight * prior_loss
+            else:
+                loss = (noise - noise_pred) ** 2
+                loss = loss.mean()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_unet_state = unet_state.apply_gradients(grads=grad["unet"])
+        if args.train_text_encoder:
+            new_text_encoder_state = text_encoder_state.apply_gradients(grads=grad["text_encoder"])
+        else:
+            new_text_encoder_state = text_encoder_state
+
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_unet_state, new_text_encoder_state, metrics, new_train_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0, 1))
+
+    # Replicate the train state on each device
+    unet_state = jax_utils.replicate(unet_state)
+    text_encoder_state = jax_utils.replicate(text_encoder_state)
+    vae_params = jax_utils.replicate(vae_params)
+
+    # Train!
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader))
+
+    # Scheduler and math around the number of training steps.
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {total_train_batch_size}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+
+    global_step = 0
+
+    epochs = tqdm(range(args.num_train_epochs), desc="Epoch ... ", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+
+        train_metrics = []
+
+        steps_per_epoch = len(train_dataset) // total_train_batch_size
+        train_step_progress_bar = tqdm(total=steps_per_epoch, desc="Training...", position=1, leave=False)
+        # train
+        for batch in train_dataloader:
+            batch = shard(batch)
+            unet_state, text_encoder_state, train_metric, train_rngs = p_train_step(
+                unet_state, text_encoder_state, vae_params, batch, train_rngs
+            )
+            train_metrics.append(train_metric)
+
+            train_step_progress_bar.update(1)
+
+            global_step += 1
+            if global_step >= args.max_train_steps:
+                break
+
+        train_metric = jax_utils.unreplicate(train_metric)
+
+        train_step_progress_bar.close()
+        epochs.write(f"Epoch... ({epoch + 1}/{args.num_train_epochs} | Loss: {train_metric['loss']})")
+
+    # Create the pipeline using using the trained modules and save it.
+    if jax.process_index() == 0:
+        scheduler = FlaxPNDMScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
+        )
+        safety_checker = FlaxStableDiffusionSafetyChecker.from_pretrained(
+            "CompVis/stable-diffusion-safety-checker", from_pt=True
+        )
+        pipeline = FlaxStableDiffusionPipeline(
+            text_encoder=text_encoder,
+            vae=vae,
+            unet=unet,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"),
+        )
+
+        pipeline.save_pretrained(
+            args.output_dir,
+            params={
+                "text_encoder": get_params_to_save(text_encoder_state.params),
+                "vae": get_params_to_save(vae_params),
+                "unet": get_params_to_save(unet_state.params),
+                "safety_checker": safety_checker.params,
+            },
+        )
+
+        if args.push_to_hub:
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
+
+
+if __name__ == "__main__":
+    main()

examples/rl/README.md

@@ -0,0 +1,19 @@
+# Overview
+
+These examples show how to run (Diffuser)[https://arxiv.org/abs/2205.09991] in Diffusers. 
+There are four scripts, 
+1. `run_diffuser_locomotion.py` to sample actions and run them in the environment,
+2. and `run_diffuser_gen_trajectories.py` to just sample actions from the pre-trained diffusion model.
+
+You will need some RL specific requirements to run the examples:
+
+```
+pip install -f https://download.pytorch.org/whl/torch_stable.html \
+                free-mujoco-py \
+                einops \
+                gym==0.24.1 \
+                protobuf==3.20.1 \
+                git+https://github.com/rail-berkeley/d4rl.git \
+                mediapy \
+                Pillow==9.0.0
+```

examples/rl/run_diffuser_gen_trajectories.py

@@ -0,0 +1,57 @@
+import d4rl  # noqa
+import gym
+import tqdm
+from diffusers.experimental import ValueGuidedRLPipeline
+
+
+config = dict(
+    n_samples=64,
+    horizon=32,
+    num_inference_steps=20,
+    n_guide_steps=0,
+    scale_grad_by_std=True,
+    scale=0.1,
+    eta=0.0,
+    t_grad_cutoff=2,
+    device="cpu",
+)
+
+
+if __name__ == "__main__":
+    env_name = "hopper-medium-v2"
+    env = gym.make(env_name)
+
+    pipeline = ValueGuidedRLPipeline.from_pretrained(
+        "bglick13/hopper-medium-v2-value-function-hor32",
+        env=env,
+    )
+
+    env.seed(0)
+    obs = env.reset()
+    total_reward = 0
+    total_score = 0
+    T = 1000
+    rollout = [obs.copy()]
+    try:
+        for t in tqdm.tqdm(range(T)):
+            # Call the policy
+            denorm_actions = pipeline(obs, planning_horizon=32)
+
+            # execute action in environment
+            next_observation, reward, terminal, _ = env.step(denorm_actions)
+            score = env.get_normalized_score(total_reward)
+            # update return
+            total_reward += reward
+            total_score += score
+            print(
+                f"Step: {t}, Reward: {reward}, Total Reward: {total_reward}, Score: {score}, Total Score:"
+                f" {total_score}"
+            )
+            # save observations for rendering
+            rollout.append(next_observation.copy())
+
+            obs = next_observation
+    except KeyboardInterrupt:
+        pass
+
+    print(f"Total reward: {total_reward}")

examples/rl/run_diffuser_locomotion.py

@@ -0,0 +1,57 @@
+import d4rl  # noqa
+import gym
+import tqdm
+from diffusers.experimental import ValueGuidedRLPipeline
+
+
+config = dict(
+    n_samples=64,
+    horizon=32,
+    num_inference_steps=20,
+    n_guide_steps=2,
+    scale_grad_by_std=True,
+    scale=0.1,
+    eta=0.0,
+    t_grad_cutoff=2,
+    device="cpu",
+)
+
+
+if __name__ == "__main__":
+    env_name = "hopper-medium-v2"
+    env = gym.make(env_name)
+
+    pipeline = ValueGuidedRLPipeline.from_pretrained(
+        "bglick13/hopper-medium-v2-value-function-hor32",
+        env=env,
+    )
+
+    env.seed(0)
+    obs = env.reset()
+    total_reward = 0
+    total_score = 0
+    T = 1000
+    rollout = [obs.copy()]
+    try:
+        for t in tqdm.tqdm(range(T)):
+            # call the policy
+            denorm_actions = pipeline(obs, planning_horizon=32)
+
+            # execute action in environment
+            next_observation, reward, terminal, _ = env.step(denorm_actions)
+            score = env.get_normalized_score(total_reward)
+            # update return
+            total_reward += reward
+            total_score += score
+            print(
+                f"Step: {t}, Reward: {reward}, Total Reward: {total_reward}, Score: {score}, Total Score:"
+                f" {total_score}"
+            )
+            # save observations for rendering
+            rollout.append(next_observation.copy())
+
+            obs = next_observation
+    except KeyboardInterrupt:
+        pass
+
+    print(f"Total reward: {total_reward}")

examples/test_examples.py

@@ -101,7 +101,7 @@ class ExamplesTestsAccelerate(unittest.TestCase):
         with tempfile.TemporaryDirectory() as tmpdir:
             test_args = f"""
                 examples/textual_inversion/textual_inversion.py
-                --pretrained_model_name_or_path CompVis/stable-diffusion-v1-4
+                --pretrained_model_name_or_path runwayml/stable-diffusion-v1-5
                 --train_data_dir docs/source/imgs
                 --learnable_property object
                 --placeholder_token <cat-toy>

examples/text_to_image/README.md

@@ -0,0 +1,153 @@
+# Stable Diffusion text-to-image fine-tuning
+
+The `train_text_to_image.py` script shows how to fine-tune stable diffusion model on your own dataset.
+
+___Note___:
+
+___This script is experimental. The script fine-tunes the whole model and often times the model overfits and runs into issues like catastrophic forgetting. It's recommended to try different hyperparamters to get the best result on your dataset.___
+
+
+## Running locally with PyTorch
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+```bash
+pip install git+https://github.com/huggingface/diffusers.git
+pip install -U -r requirements.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+### Pokemon example
+
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+
+You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
+
+Run the following command to authenticate your token
+
+```bash
+huggingface-cli login
+```
+
+If you have already cloned the repo, then you won't need to go through these steps.
+
+<br>
+
+#### Hardware
+With `gradient_checkpointing` and `mixed_precision` it should be possible to fine tune the model on a single 24GB GPU. For higher `batch_size` and faster training it's better to use GPUs with >30GB memory.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+accelerate launch train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir="sd-pokemon-model" 
+```
+
+
+To run on your own training files prepare the dataset according to the format required by `datasets`, you can find the instructions for how to do that in this [document](https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder-with-metadata).
+If you wish to use custom loading logic, you should modify the script, we have left pointers for that in the training script.
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export TRAIN_DIR="path_to_your_dataset"
+
+accelerate launch train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --train_data_dir=$TRAIN_DIR \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir="sd-pokemon-model"
+```
+
+
+Once the training is finished the model will be saved in the `output_dir` specified in the command. In this example it's `sd-pokemon-model`. To load the fine-tuned model for inference just pass that path to `StableDiffusionPipeline`
+
+
+```python
+from diffusers import StableDiffusionPipeline
+
+model_path = "path_to_saved_model"
+pipe = StableDiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
+pipe.to("cuda")
+
+image = pipe(prompt="yoda").images[0]
+image.save("yoda-pokemon.png")
+```
+
+
+
+## Training with Flax/JAX
+
+For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.
+
+____Note: The flax example don't yet support features like gradient checkpoint, gradient accumulation etc, so to use flax for faster training we will need >30GB cards.___
+
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+```bash
+pip install -U -r requirements_flax.txt
+```
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+python train_text_to_image_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --output_dir="sd-pokemon-model" 
+```
+
+
+To run on your own training files prepare the dataset according to the format required by `datasets`, you can find the instructions for how to do that in this [document](https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder-with-metadata).
+If you wish to use custom loading logic, you should modify the script, we have left pointers for that in the training script.
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export TRAIN_DIR="path_to_your_dataset"
+
+python train_text_to_image_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --train_data_dir=$TRAIN_DIR \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --mixed_precision="fp16" \
+  --max_train_steps=15000 \
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --output_dir="sd-pokemon-model"
+```

examples/text_to_image/requirements.txt

@@ -0,0 +1,7 @@
+diffusers==0.4.1
+accelerate
+torchvision
+transformers>=4.21.0
+ftfy
+tensorboard
+modelcards

examples/text_to_image/requirements_flax.txt

@@ -0,0 +1,9 @@
+diffusers>==0.5.1
+transformers>=4.21.0
+flax
+optax
+torch
+torchvision
+ftfy
+tensorboard
+modelcards

examples/text_to_image/train_text_to_image.py

@@ -0,0 +1,621 @@
+import argparse
+import logging
+import math
+import os
+import random
+from pathlib import Path
+from typing import Iterable, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import set_seed
+from datasets import load_dataset
+from diffusers import AutoencoderKL, DDPMScheduler, PNDMScheduler, StableDiffusionPipeline, UNet2DConditionModel
+from diffusers.optimization import get_scheduler
+from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
+from huggingface_hub import HfFolder, Repository, whoami
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+logger = get_logger(__name__)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default="text",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-model-finetuned",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        action="store_true",
+        help="Whether to center crop images before resizing to resolution (if not set, random crop will be used)",
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
+            ' `"wandb"` and `"comet_ml"`. Use `"all"` (default) to report to all integrations.'
+            "Only applicable when `--with_tracking` is passed."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Need either a dataset name or a training folder.")
+
+    return args
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+dataset_name_mapping = {
+    "lambdalabs/pokemon-blip-captions": ("image", "text"),
+}
+
+
+# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
+class EMAModel:
+    """
+    Exponential Moving Average of models weights
+    """
+
+    def __init__(self, parameters: Iterable[torch.nn.Parameter], decay=0.9999):
+        parameters = list(parameters)
+        self.shadow_params = [p.clone().detach() for p in parameters]
+
+        self.decay = decay
+        self.optimization_step = 0
+
+    def get_decay(self, optimization_step):
+        """
+        Compute the decay factor for the exponential moving average.
+        """
+        value = (1 + optimization_step) / (10 + optimization_step)
+        return 1 - min(self.decay, value)
+
+    @torch.no_grad()
+    def step(self, parameters):
+        parameters = list(parameters)
+
+        self.optimization_step += 1
+        self.decay = self.get_decay(self.optimization_step)
+
+        for s_param, param in zip(self.shadow_params, parameters):
+            if param.requires_grad:
+                tmp = self.decay * (s_param - param)
+                s_param.sub_(tmp)
+            else:
+                s_param.copy_(param)
+
+        torch.cuda.empty_cache()
+
+    def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+        """
+        Copy current averaged parameters into given collection of parameters.
+
+        Args:
+            parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+                updated with the stored moving averages. If `None`, the
+                parameters with which this `ExponentialMovingAverage` was
+                initialized will be used.
+        """
+        parameters = list(parameters)
+        for s_param, param in zip(self.shadow_params, parameters):
+            param.data.copy_(s_param.data)
+
+    def to(self, device=None, dtype=None) -> None:
+        r"""Move internal buffers of the ExponentialMovingAverage to `device`.
+
+        Args:
+            device: like `device` argument to `torch.Tensor.to`
+        """
+        # .to() on the tensors handles None correctly
+        self.shadow_params = [
+            p.to(device=device, dtype=dtype) if p.is_floating_point() else p.to(device=device)
+            for p in self.shadow_params
+        ]
+
+
+def main():
+    args = parse_args()
+    logging_dir = os.path.join(args.output_dir, args.logging_dir)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        logging_dir=logging_dir,
+    )
+
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.push_to_hub:
+            if args.hub_model_id is None:
+                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
+            else:
+                repo_name = args.hub_model_id
+            repo = Repository(args.output_dir, clone_from=repo_name)
+
+            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "step_*" not in gitignore:
+                    gitignore.write("step_*\n")
+                if "epoch_*" not in gitignore:
+                    gitignore.write("epoch_*\n")
+        elif args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load models and create wrapper for stable diffusion
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+
+    # Freeze vae and text_encoder
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+
+    if args.gradient_checkpointing:
+        unet.enable_gradient_checkpointing()
+
+    if args.scale_lr:
+        args.learning_rate = (
+            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Initialize the optimizer
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError(
+                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
+            )
+
+        optimizer_cls = bnb.optim.AdamW8bit
+    else:
+        optimizer_cls = torch.optim.AdamW
+
+    optimizer = optimizer_cls(
+        unet.parameters(),
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+    noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler")
+
+    # Get the datasets: you can either provide your own training and evaluation files (see below)
+    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
+
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            args.dataset_name,
+            args.dataset_config_name,
+            cache_dir=args.cache_dir,
+        )
+    else:
+        data_files = {}
+        if args.train_data_dir is not None:
+            data_files["train"] = os.path.join(args.train_data_dir, "**")
+        dataset = load_dataset(
+            "imagefolder",
+            data_files=data_files,
+            cache_dir=args.cache_dir,
+        )
+        # See more about loading custom images at
+        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    column_names = dataset["train"].column_names
+
+    # 6. Get the column names for input/target.
+    dataset_columns = dataset_name_mapping.get(args.dataset_name, None)
+    if args.image_column is None:
+        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
+    else:
+        image_column = args.image_column
+        if image_column not in column_names:
+            raise ValueError(
+                f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
+            )
+    if args.caption_column is None:
+        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
+    else:
+        caption_column = args.caption_column
+        if caption_column not in column_names:
+            raise ValueError(
+                f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
+            )
+
+    # Preprocessing the datasets.
+    # We need to tokenize input captions and transform the images.
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if isinstance(caption, str):
+                captions.append(caption)
+            elif isinstance(caption, (list, np.ndarray)):
+                # take a random caption if there are multiple
+                captions.append(random.choice(caption) if is_train else caption[0])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        inputs = tokenizer(captions, max_length=tokenizer.model_max_length, padding="do_not_pad", truncation=True)
+        input_ids = inputs.input_ids
+        return input_ids
+
+    train_transforms = transforms.Compose(
+        [
+            transforms.Resize((args.resolution, args.resolution), interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
+            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    def preprocess_train(examples):
+        images = [image.convert("RGB") for image in examples[image_column]]
+        examples["pixel_values"] = [train_transforms(image) for image in images]
+        examples["input_ids"] = tokenize_captions(examples)
+
+        return examples
+
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
+        # Set the training transforms
+        train_dataset = dataset["train"].with_transform(preprocess_train)
+
+    def collate_fn(examples):
+        pixel_values = torch.stack([example["pixel_values"] for example in examples])
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+        input_ids = [example["input_ids"] for example in examples]
+        padded_tokens = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt")
+        return {
+            "pixel_values": pixel_values,
+            "input_ids": padded_tokens.input_ids,
+            "attention_mask": padded_tokens.attention_mask,
+        }
+
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset, shuffle=True, collate_fn=collate_fn, batch_size=args.train_batch_size
+    )
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
+        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+    )
+
+    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        unet, optimizer, train_dataloader, lr_scheduler
+    )
+
+    weight_dtype = torch.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move text_encode and vae to gpu.
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models are only used for inference, keeping weights in full precision is not required.
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+
+    # Create EMA for the unet.
+    if args.use_ema:
+        ema_unet = EMAModel(unet.parameters())
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        accelerator.init_trackers("text2image-fine-tune", config=vars(args))
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+
+    # Only show the progress bar once on each machine.
+    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
+    progress_bar.set_description("Steps")
+    global_step = 0
+
+    for epoch in range(args.num_train_epochs):
+        unet.train()
+        train_loss = 0.0
+        for step, batch in enumerate(train_dataloader):
+            with accelerator.accumulate(unet):
+                # Convert images to latent space
+                latents = vae.encode(batch["pixel_values"].to(weight_dtype)).latent_dist.sample()
+                latents = latents * 0.18215
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                # Get the text embedding for conditioning
+                encoder_hidden_states = text_encoder(batch["input_ids"])[0]
+
+                # Predict the noise residual and compute loss
+                noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
+                loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
+                train_loss += avg_loss.item() / args.gradient_accumulation_steps
+
+                # Backpropagate
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                if args.use_ema:
+                    ema_unet.step(unet.parameters())
+                progress_bar.update(1)
+                global_step += 1
+                accelerator.log({"train_loss": train_loss}, step=global_step)
+                train_loss = 0.0
+
+            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+    # Create the pipeline using the trained modules and save it.
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        unet = accelerator.unwrap_model(unet)
+        if args.use_ema:
+            ema_unet.copy_to(unet.parameters())
+
+        pipeline = StableDiffusionPipeline(
+            text_encoder=text_encoder,
+            vae=vae,
+            unet=unet,
+            tokenizer=tokenizer,
+            scheduler=PNDMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler"),
+            safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker"),
+            feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"),
+        )
+        pipeline.save_pretrained(args.output_dir)
+
+        if args.push_to_hub:
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    main()

examples/text_to_image/train_text_to_image_flax.py

@@ -0,0 +1,560 @@
+import argparse
+import logging
+import math
+import os
+import random
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from datasets import load_dataset
+from diffusers import (
+    FlaxAutoencoderKL,
+    FlaxDDPMScheduler,
+    FlaxPNDMScheduler,
+    FlaxStableDiffusionPipeline,
+    FlaxUNet2DConditionModel,
+)
+from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionSafetyChecker
+from flax import jax_utils
+from flax.training import train_state
+from flax.training.common_utils import shard
+from huggingface_hub import HfFolder, Repository, whoami
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel, set_seed
+
+
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default="text",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-model-finetuned",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=0, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        action="store_true",
+        help="Whether to center crop images before resizing to resolution (if not set, random crop will be used)",
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
+            ' `"wandb"` and `"comet_ml"`. Use `"all"` (default) to report to all integrations.'
+            "Only applicable when `--with_tracking` is passed."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Need either a dataset name or a training folder.")
+
+    return args
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+dataset_name_mapping = {
+    "lambdalabs/pokemon-blip-captions": ("image", "text"),
+}
+
+
+def get_params_to_save(params):
+    return jax.device_get(jax.tree_util.tree_map(lambda x: x[0], params))
+
+
+def main():
+    args = parse_args()
+
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if jax.process_index() == 0:
+        if args.push_to_hub:
+            if args.hub_model_id is None:
+                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
+            else:
+                repo_name = args.hub_model_id
+            repo = Repository(args.output_dir, clone_from=repo_name)
+
+            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "step_*" not in gitignore:
+                    gitignore.write("step_*\n")
+                if "epoch_*" not in gitignore:
+                    gitignore.write("epoch_*\n")
+        elif args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Get the datasets: you can either provide your own training and evaluation files (see below)
+    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
+
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            args.dataset_name,
+            args.dataset_config_name,
+            cache_dir=args.cache_dir,
+        )
+    else:
+        data_files = {}
+        if args.train_data_dir is not None:
+            data_files["train"] = os.path.join(args.train_data_dir, "**")
+        dataset = load_dataset(
+            "imagefolder",
+            data_files=data_files,
+            cache_dir=args.cache_dir,
+        )
+        # See more about loading custom images at
+        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    column_names = dataset["train"].column_names
+
+    # 6. Get the column names for input/target.
+    dataset_columns = dataset_name_mapping.get(args.dataset_name, None)
+    if args.image_column is None:
+        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
+    else:
+        image_column = args.image_column
+        if image_column not in column_names:
+            raise ValueError(
+                f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
+            )
+    if args.caption_column is None:
+        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
+    else:
+        caption_column = args.caption_column
+        if caption_column not in column_names:
+            raise ValueError(
+                f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
+            )
+
+    # Preprocessing the datasets.
+    # We need to tokenize input captions and transform the images.
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if isinstance(caption, str):
+                captions.append(caption)
+            elif isinstance(caption, (list, np.ndarray)):
+                # take a random caption if there are multiple
+                captions.append(random.choice(caption) if is_train else caption[0])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        inputs = tokenizer(captions, max_length=tokenizer.model_max_length, padding="do_not_pad", truncation=True)
+        input_ids = inputs.input_ids
+        return input_ids
+
+    train_transforms = transforms.Compose(
+        [
+            transforms.Resize((args.resolution, args.resolution), interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
+            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    def preprocess_train(examples):
+        images = [image.convert("RGB") for image in examples[image_column]]
+        examples["pixel_values"] = [train_transforms(image) for image in images]
+        examples["input_ids"] = tokenize_captions(examples)
+
+        return examples
+
+    if jax.process_index() == 0:
+        if args.max_train_samples is not None:
+            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
+        # Set the training transforms
+        train_dataset = dataset["train"].with_transform(preprocess_train)
+
+    def collate_fn(examples):
+        pixel_values = torch.stack([example["pixel_values"] for example in examples])
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+        input_ids = [example["input_ids"] for example in examples]
+
+        padded_tokens = tokenizer.pad(
+            {"input_ids": input_ids}, padding="max_length", max_length=tokenizer.model_max_length, return_tensors="pt"
+        )
+        batch = {
+            "pixel_values": pixel_values,
+            "input_ids": padded_tokens.input_ids,
+        }
+        batch = {k: v.numpy() for k, v in batch.items()}
+
+        return batch
+
+    total_train_batch_size = args.train_batch_size * jax.local_device_count()
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset, shuffle=True, collate_fn=collate_fn, batch_size=total_train_batch_size, drop_last=True
+    )
+
+    weight_dtype = jnp.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = jnp.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = jnp.bfloat16
+
+    # Load models and create wrapper for stable diffusion
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = FlaxCLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", dtype=weight_dtype
+    )
+    vae, vae_params = FlaxAutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="vae", dtype=weight_dtype
+    )
+    unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", dtype=weight_dtype
+    )
+
+    # Optimization
+    if args.scale_lr:
+        args.learning_rate = args.learning_rate * total_train_batch_size
+
+    constant_scheduler = optax.constant_schedule(args.learning_rate)
+
+    adamw = optax.adamw(
+        learning_rate=constant_scheduler,
+        b1=args.adam_beta1,
+        b2=args.adam_beta2,
+        eps=args.adam_epsilon,
+        weight_decay=args.adam_weight_decay,
+    )
+
+    optimizer = optax.chain(
+        optax.clip_by_global_norm(args.max_grad_norm),
+        adamw,
+    )
+
+    state = train_state.TrainState.create(apply_fn=unet.__call__, params=unet_params, tx=optimizer)
+
+    noise_scheduler = FlaxDDPMScheduler(
+        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000
+    )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(args.seed)
+    train_rngs = jax.random.split(rng, jax.local_device_count())
+
+    def train_step(state, text_encoder_params, vae_params, batch, train_rng):
+        dropout_rng, sample_rng, new_train_rng = jax.random.split(train_rng, 3)
+
+        def compute_loss(params):
+            # Convert images to latent space
+            vae_outputs = vae.apply(
+                {"params": vae_params}, batch["pixel_values"], deterministic=True, method=vae.encode
+            )
+            latents = vae_outputs.latent_dist.sample(sample_rng)
+            # (NHWC) -> (NCHW)
+            latents = jnp.transpose(latents, (0, 3, 1, 2))
+            latents = latents * 0.18215
+
+            # Sample noise that we'll add to the latents
+            noise_rng, timestep_rng = jax.random.split(sample_rng)
+            noise = jax.random.normal(noise_rng, latents.shape)
+            # Sample a random timestep for each image
+            bsz = latents.shape[0]
+            timesteps = jax.random.randint(
+                timestep_rng,
+                (bsz,),
+                0,
+                noise_scheduler.config.num_train_timesteps,
+            )
+
+            # Add noise to the latents according to the noise magnitude at each timestep
+            # (this is the forward diffusion process)
+            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+            # Get the text embedding for conditioning
+            encoder_hidden_states = text_encoder(
+                batch["input_ids"],
+                params=text_encoder_params,
+                train=False,
+            )[0]
+
+            # Predict the noise residual and compute loss
+            unet_outputs = unet.apply({"params": params}, noisy_latents, timesteps, encoder_hidden_states, train=True)
+            noise_pred = unet_outputs.sample
+            loss = (noise - noise_pred) ** 2
+            loss = loss.mean()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics, new_train_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+    text_encoder_params = jax_utils.replicate(text_encoder.params)
+    vae_params = jax_utils.replicate(vae_params)
+
+    # Train!
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader))
+
+    # Scheduler and math around the number of training steps.
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {total_train_batch_size}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+
+    global_step = 0
+
+    epochs = tqdm(range(args.num_train_epochs), desc="Epoch ... ", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+
+        train_metrics = []
+
+        steps_per_epoch = len(train_dataset) // total_train_batch_size
+        train_step_progress_bar = tqdm(total=steps_per_epoch, desc="Training...", position=1, leave=False)
+        # train
+        for batch in train_dataloader:
+            batch = shard(batch)
+            state, train_metric, train_rngs = p_train_step(state, text_encoder_params, vae_params, batch, train_rngs)
+            train_metrics.append(train_metric)
+
+            train_step_progress_bar.update(1)
+
+            global_step += 1
+            if global_step >= args.max_train_steps:
+                break
+
+        train_metric = jax_utils.unreplicate(train_metric)
+
+        train_step_progress_bar.close()
+        epochs.write(f"Epoch... ({epoch + 1}/{args.num_train_epochs} | Loss: {train_metric['loss']})")
+
+    # Create the pipeline using using the trained modules and save it.
+    if jax.process_index() == 0:
+        scheduler = FlaxPNDMScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
+        )
+        safety_checker = FlaxStableDiffusionSafetyChecker.from_pretrained(
+            "CompVis/stable-diffusion-safety-checker", from_pt=True
+        )
+        pipeline = FlaxStableDiffusionPipeline(
+            text_encoder=text_encoder,
+            vae=vae,
+            unet=unet,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"),
+        )
+
+        pipeline.save_pretrained(
+            args.output_dir,
+            params={
+                "text_encoder": get_params_to_save(text_encoder_params),
+                "vae": get_params_to_save(vae_params),
+                "unet": get_params_to_save(state.params),
+                "safety_checker": safety_checker.params,
+            },
+        )
+
+        if args.push_to_hub:
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
+
+
+if __name__ == "__main__":
+    main()

examples/textual_inversion/README.md

@@ -11,7 +11,7 @@ Colab for training
 Colab for inference
 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_conceptualizer_inference.ipynb)
 
-## Running locally 
+## Running locally with PyTorch
 ### Installing the dependencies
 
 Before running the scripts, make sure to install the library's training dependencies:
@@ -29,7 +29,7 @@ accelerate config
 
 ### Cat toy example
 
-You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-4`, so you'll need to visit [its card](https://huggingface.co/CompVis/stable-diffusion-v1-4), read the license and tick the checkbox if you agree. 
+You need to accept the model license before downloading or using the weights. In this example we'll use model version `v1-5`, so you'll need to visit [its card](https://huggingface.co/runwayml/stable-diffusion-v1-5), read the license and tick the checkbox if you agree. 
 
 You have to be a registered user in 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to [this section of the documentation](https://huggingface.co/docs/hub/security-tokens).
 
@@ -48,7 +48,7 @@ Now let's get our dataset.Download 3-4 images from [here](https://drive.google.c
 And launch the training using
 
 ```bash
-export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
 export DATA_DIR="path-to-dir-containing-images"
 
 accelerate launch textual_inversion.py \
@@ -68,7 +68,6 @@ accelerate launch textual_inversion.py \
 
 A full training run takes ~1 hour on one V100 GPU.
 
-
 ### Inference
 
 Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline`. Make sure to include the `placeholder_token` in your prompt.
@@ -85,3 +84,31 @@ image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
 
 image.save("cat-backpack.png")
 ```
+
+
+## Training with Flax/JAX
+
+For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+```bash
+pip install -U -r requirements_flax.txt
+```
+
+```bash
+export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
+export DATA_DIR="path-to-dir-containing-images"
+
+python textual_inversion_flax.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --train_data_dir=$DATA_DIR \
+  --learnable_property="object" \
+  --placeholder_token="<cat-toy>" --initializer_token="toy" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --max_train_steps=3000 \
+  --learning_rate=5.0e-04 --scale_lr \
+  --output_dir="textual_inversion_cat"
+```
+It should be at least 70% faster than the PyTorch script with the same configuration.

examples/textual_inversion/requirements_flax.txt

@@ -0,0 +1,9 @@
+diffusers>==0.5.1
+transformers>=4.21.0
+flax
+optax
+torch
+torchvision
+ftfy
+tensorboard
+modelcards

examples/textual_inversion/textual_inversion.py

@@ -20,12 +20,34 @@ from diffusers import AutoencoderKL, DDPMScheduler, PNDMScheduler, StableDiffusi
 from diffusers.optimization import get_scheduler
 from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 from huggingface_hub import HfFolder, Repository, whoami
+
+# TODO: remove and import from diffusers.utils when the new version of diffusers is released
+from packaging import version
 from PIL import Image
 from torchvision import transforms
 from tqdm.auto import tqdm
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 
+if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.Resampling.BILINEAR,
+        "bilinear": PIL.Image.Resampling.BILINEAR,
+        "bicubic": PIL.Image.Resampling.BICUBIC,
+        "lanczos": PIL.Image.Resampling.LANCZOS,
+        "nearest": PIL.Image.Resampling.NEAREST,
+    }
+else:
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.LINEAR,
+        "bilinear": PIL.Image.BILINEAR,
+        "bicubic": PIL.Image.BICUBIC,
+        "lanczos": PIL.Image.LANCZOS,
+        "nearest": PIL.Image.NEAREST,
+    }
+# ------------------------------------------------------------------------------
+
+
 logger = get_logger(__name__)
 
 
@@ -260,10 +282,10 @@ class TextualInversionDataset(Dataset):
             self._length = self.num_images * repeats
 
         self.interpolation = {
-            "linear": PIL.Image.LINEAR,
-            "bilinear": PIL.Image.BILINEAR,
-            "bicubic": PIL.Image.BICUBIC,
-            "lanczos": PIL.Image.LANCZOS,
+            "linear": PIL_INTERPOLATION["linear"],
+            "bilinear": PIL_INTERPOLATION["bilinear"],
+            "bicubic": PIL_INTERPOLATION["bicubic"],
+            "lanczos": PIL_INTERPOLATION["lanczos"],
         }[interpolation]
 
         self.templates = imagenet_style_templates_small if learnable_property == "style" else imagenet_templates_small
@@ -419,13 +441,7 @@ def main():
         eps=args.adam_epsilon,
     )
 
-    # TODO (patil-suraj): load scheduler using args
-    noise_scheduler = DDPMScheduler(
-        beta_start=0.00085,
-        beta_end=0.012,
-        beta_schedule="scaled_linear",
-        num_train_timesteps=1000,
-    )
+    noise_scheduler = DDPMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler")
 
     train_dataset = TextualInversionDataset(
         data_root=args.train_data_dir,
@@ -558,9 +574,7 @@ def main():
             vae=vae,
             unet=unet,
             tokenizer=tokenizer,
-            scheduler=PNDMScheduler(
-                beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
-            ),
+            scheduler=PNDMScheduler.from_config(args.pretrained_model_name_or_path, subfolder="scheduler"),
             safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker"),
             feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"),
         )
@@ -569,9 +583,7 @@ def main():
         save_progress(text_encoder, placeholder_token_id, accelerator, args)
 
         if args.push_to_hub:
-            repo.push_to_hub(
-                args, pipeline, repo, commit_message="End of training", blocking=False, auto_lfs_prune=True
-            )
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
 
     accelerator.end_training()
 

examples/textual_inversion/textual_inversion_flax.py

@@ -0,0 +1,650 @@
+import argparse
+import logging
+import math
+import os
+import random
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch.utils.data import Dataset
+
+import jax
+import jax.numpy as jnp
+import optax
+import PIL
+import transformers
+from diffusers import (
+    FlaxAutoencoderKL,
+    FlaxDDPMScheduler,
+    FlaxPNDMScheduler,
+    FlaxStableDiffusionPipeline,
+    FlaxUNet2DConditionModel,
+)
+from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionSafetyChecker
+from flax import jax_utils
+from flax.training import train_state
+from flax.training.common_utils import shard
+from huggingface_hub import HfFolder, Repository, whoami
+
+# TODO: remove and import from diffusers.utils when the new version of diffusers is released
+from packaging import version
+from PIL import Image
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel, set_seed
+
+
+if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.Resampling.BILINEAR,
+        "bilinear": PIL.Image.Resampling.BILINEAR,
+        "bicubic": PIL.Image.Resampling.BICUBIC,
+        "lanczos": PIL.Image.Resampling.LANCZOS,
+        "nearest": PIL.Image.Resampling.NEAREST,
+    }
+else:
+    PIL_INTERPOLATION = {
+        "linear": PIL.Image.LINEAR,
+        "bilinear": PIL.Image.BILINEAR,
+        "bicubic": PIL.Image.BICUBIC,
+        "lanczos": PIL.Image.LANCZOS,
+        "nearest": PIL.Image.NEAREST,
+    }
+# ------------------------------------------------------------------------------
+
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--train_data_dir", type=str, default=None, required=True, help="A folder containing the training data."
+    )
+    parser.add_argument(
+        "--placeholder_token",
+        type=str,
+        default=None,
+        required=True,
+        help="A token to use as a placeholder for the concept.",
+    )
+    parser.add_argument(
+        "--initializer_token", type=str, default=None, required=True, help="A token to use as initializer word."
+    )
+    parser.add_argument("--learnable_property", type=str, default="object", help="Choose between 'object' and 'style'")
+    parser.add_argument("--repeats", type=int, default=100, help="How many times to repeat the training data.")
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="text-inversion-model",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution"
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=5000,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=True,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument(
+        "--use_auth_token",
+        action="store_true",
+        help=(
+            "Will use the token generated when running `huggingface-cli login` (necessary to use this script with"
+            " private models)."
+        ),
+    )
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.train_data_dir is None:
+        raise ValueError("You must specify a train data directory.")
+
+    return args
+
+
+imagenet_templates_small = [
+    "a photo of a {}",
+    "a rendering of a {}",
+    "a cropped photo of the {}",
+    "the photo of a {}",
+    "a photo of a clean {}",
+    "a photo of a dirty {}",
+    "a dark photo of the {}",
+    "a photo of my {}",
+    "a photo of the cool {}",
+    "a close-up photo of a {}",
+    "a bright photo of the {}",
+    "a cropped photo of a {}",
+    "a photo of the {}",
+    "a good photo of the {}",
+    "a photo of one {}",
+    "a close-up photo of the {}",
+    "a rendition of the {}",
+    "a photo of the clean {}",
+    "a rendition of a {}",
+    "a photo of a nice {}",
+    "a good photo of a {}",
+    "a photo of the nice {}",
+    "a photo of the small {}",
+    "a photo of the weird {}",
+    "a photo of the large {}",
+    "a photo of a cool {}",
+    "a photo of a small {}",
+]
+
+imagenet_style_templates_small = [
+    "a painting in the style of {}",
+    "a rendering in the style of {}",
+    "a cropped painting in the style of {}",
+    "the painting in the style of {}",
+    "a clean painting in the style of {}",
+    "a dirty painting in the style of {}",
+    "a dark painting in the style of {}",
+    "a picture in the style of {}",
+    "a cool painting in the style of {}",
+    "a close-up painting in the style of {}",
+    "a bright painting in the style of {}",
+    "a cropped painting in the style of {}",
+    "a good painting in the style of {}",
+    "a close-up painting in the style of {}",
+    "a rendition in the style of {}",
+    "a nice painting in the style of {}",
+    "a small painting in the style of {}",
+    "a weird painting in the style of {}",
+    "a large painting in the style of {}",
+]
+
+
+class TextualInversionDataset(Dataset):
+    def __init__(
+        self,
+        data_root,
+        tokenizer,
+        learnable_property="object",  # [object, style]
+        size=512,
+        repeats=100,
+        interpolation="bicubic",
+        flip_p=0.5,
+        set="train",
+        placeholder_token="*",
+        center_crop=False,
+    ):
+        self.data_root = data_root
+        self.tokenizer = tokenizer
+        self.learnable_property = learnable_property
+        self.size = size
+        self.placeholder_token = placeholder_token
+        self.center_crop = center_crop
+        self.flip_p = flip_p
+
+        self.image_paths = [os.path.join(self.data_root, file_path) for file_path in os.listdir(self.data_root)]
+
+        self.num_images = len(self.image_paths)
+        self._length = self.num_images
+
+        if set == "train":
+            self._length = self.num_images * repeats
+
+        self.interpolation = {
+            "linear": PIL_INTERPOLATION["linear"],
+            "bilinear": PIL_INTERPOLATION["bilinear"],
+            "bicubic": PIL_INTERPOLATION["bicubic"],
+            "lanczos": PIL_INTERPOLATION["lanczos"],
+        }[interpolation]
+
+        self.templates = imagenet_style_templates_small if learnable_property == "style" else imagenet_templates_small
+        self.flip_transform = transforms.RandomHorizontalFlip(p=self.flip_p)
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, i):
+        example = {}
+        image = Image.open(self.image_paths[i % self.num_images])
+
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        placeholder_string = self.placeholder_token
+        text = random.choice(self.templates).format(placeholder_string)
+
+        example["input_ids"] = self.tokenizer(
+            text,
+            padding="max_length",
+            truncation=True,
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        ).input_ids[0]
+
+        # default to score-sde preprocessing
+        img = np.array(image).astype(np.uint8)
+
+        if self.center_crop:
+            crop = min(img.shape[0], img.shape[1])
+            h, w, = (
+                img.shape[0],
+                img.shape[1],
+            )
+            img = img[(h - crop) // 2 : (h + crop) // 2, (w - crop) // 2 : (w + crop) // 2]
+
+        image = Image.fromarray(img)
+        image = image.resize((self.size, self.size), resample=self.interpolation)
+
+        image = self.flip_transform(image)
+        image = np.array(image).astype(np.uint8)
+        image = (image / 127.5 - 1.0).astype(np.float32)
+
+        example["pixel_values"] = torch.from_numpy(image).permute(2, 0, 1)
+        return example
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+def resize_token_embeddings(model, new_num_tokens, initializer_token_id, placeholder_token_id, rng):
+    if model.config.vocab_size == new_num_tokens or new_num_tokens is None:
+        return
+    model.config.vocab_size = new_num_tokens
+
+    params = model.params
+    old_embeddings = params["text_model"]["embeddings"]["token_embedding"]["embedding"]
+    old_num_tokens, emb_dim = old_embeddings.shape
+
+    initializer = jax.nn.initializers.normal()
+
+    new_embeddings = initializer(rng, (new_num_tokens, emb_dim))
+    new_embeddings = new_embeddings.at[:old_num_tokens].set(old_embeddings)
+    new_embeddings = new_embeddings.at[placeholder_token_id].set(new_embeddings[initializer_token_id])
+    params["text_model"]["embeddings"]["token_embedding"]["embedding"] = new_embeddings
+
+    model.params = params
+    return model
+
+
+def get_params_to_save(params):
+    return jax.device_get(jax.tree_util.tree_map(lambda x: x[0], params))
+
+
+def main():
+    args = parse_args()
+
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    if jax.process_index() == 0:
+        if args.push_to_hub:
+            if args.hub_model_id is None:
+                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
+            else:
+                repo_name = args.hub_model_id
+            repo = Repository(args.output_dir, clone_from=repo_name)
+
+            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "step_*" not in gitignore:
+                    gitignore.write("step_*\n")
+                if "epoch_*" not in gitignore:
+                    gitignore.write("epoch_*\n")
+        elif args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+
+    # Load the tokenizer and add the placeholder token as a additional special token
+    if args.tokenizer_name:
+        tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+
+    # Add the placeholder token in tokenizer
+    num_added_tokens = tokenizer.add_tokens(args.placeholder_token)
+    if num_added_tokens == 0:
+        raise ValueError(
+            f"The tokenizer already contains the token {args.placeholder_token}. Please pass a different"
+            " `placeholder_token` that is not already in the tokenizer."
+        )
+
+    # Convert the initializer_token, placeholder_token to ids
+    token_ids = tokenizer.encode(args.initializer_token, add_special_tokens=False)
+    # Check if initializer_token is a single token or a sequence of tokens
+    if len(token_ids) > 1:
+        raise ValueError("The initializer token must be a single token.")
+
+    initializer_token_id = token_ids[0]
+    placeholder_token_id = tokenizer.convert_tokens_to_ids(args.placeholder_token)
+
+    # Load models and create wrapper for stable diffusion
+    text_encoder = FlaxCLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    vae, vae_params = FlaxAutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+
+    # Create sampling rng
+    rng = jax.random.PRNGKey(args.seed)
+    rng, _ = jax.random.split(rng)
+    # Resize the token embeddings as we are adding new special tokens to the tokenizer
+    text_encoder = resize_token_embeddings(
+        text_encoder, len(tokenizer), initializer_token_id, placeholder_token_id, rng
+    )
+    original_token_embeds = text_encoder.params["text_model"]["embeddings"]["token_embedding"]["embedding"]
+
+    train_dataset = TextualInversionDataset(
+        data_root=args.train_data_dir,
+        tokenizer=tokenizer,
+        size=args.resolution,
+        placeholder_token=args.placeholder_token,
+        repeats=args.repeats,
+        learnable_property=args.learnable_property,
+        center_crop=args.center_crop,
+        set="train",
+    )
+
+    def collate_fn(examples):
+        pixel_values = torch.stack([example["pixel_values"] for example in examples])
+        input_ids = torch.stack([example["input_ids"] for example in examples])
+
+        batch = {"pixel_values": pixel_values, "input_ids": input_ids}
+        batch = {k: v.numpy() for k, v in batch.items()}
+
+        return batch
+
+    total_train_batch_size = args.train_batch_size * jax.local_device_count()
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset, batch_size=total_train_batch_size, shuffle=True, drop_last=True, collate_fn=collate_fn
+    )
+
+    # Optimization
+    if args.scale_lr:
+        args.learning_rate = args.learning_rate * total_train_batch_size
+
+    constant_scheduler = optax.constant_schedule(args.learning_rate)
+
+    optimizer = optax.adamw(
+        learning_rate=constant_scheduler,
+        b1=args.adam_beta1,
+        b2=args.adam_beta2,
+        eps=args.adam_epsilon,
+        weight_decay=args.adam_weight_decay,
+    )
+
+    def create_mask(params, label_fn):
+        def _map(params, mask, label_fn):
+            for k in params:
+                if label_fn(k):
+                    mask[k] = "token_embedding"
+                else:
+                    if isinstance(params[k], dict):
+                        mask[k] = {}
+                        _map(params[k], mask[k], label_fn)
+                    else:
+                        mask[k] = "zero"
+
+        mask = {}
+        _map(params, mask, label_fn)
+        return mask
+
+    def zero_grads():
+        # from https://github.com/deepmind/optax/issues/159#issuecomment-896459491
+        def init_fn(_):
+            return ()
+
+        def update_fn(updates, state, params=None):
+            return jax.tree_util.tree_map(jnp.zeros_like, updates), ()
+
+        return optax.GradientTransformation(init_fn, update_fn)
+
+    # Zero out gradients of layers other than the token embedding layer
+    tx = optax.multi_transform(
+        {"token_embedding": optimizer, "zero": zero_grads()},
+        create_mask(text_encoder.params, lambda s: s == "token_embedding"),
+    )
+
+    state = train_state.TrainState.create(apply_fn=text_encoder.__call__, params=text_encoder.params, tx=tx)
+
+    noise_scheduler = FlaxDDPMScheduler(
+        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000
+    )
+
+    # Initialize our training
+    train_rngs = jax.random.split(rng, jax.local_device_count())
+
+    # Define gradient train step fn
+    def train_step(state, vae_params, unet_params, batch, train_rng):
+        dropout_rng, sample_rng, new_train_rng = jax.random.split(train_rng, 3)
+
+        def compute_loss(params):
+            vae_outputs = vae.apply(
+                {"params": vae_params}, batch["pixel_values"], deterministic=True, method=vae.encode
+            )
+            latents = vae_outputs.latent_dist.sample(sample_rng)
+            # (NHWC) -> (NCHW)
+            latents = jnp.transpose(latents, (0, 3, 1, 2))
+            latents = latents * 0.18215
+
+            noise_rng, timestep_rng = jax.random.split(sample_rng)
+            noise = jax.random.normal(noise_rng, latents.shape)
+            bsz = latents.shape[0]
+            timesteps = jax.random.randint(
+                timestep_rng,
+                (bsz,),
+                0,
+                noise_scheduler.config.num_train_timesteps,
+            )
+            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+            encoder_hidden_states = state.apply_fn(
+                batch["input_ids"], params=params, dropout_rng=dropout_rng, train=True
+            )[0]
+            unet_outputs = unet.apply(
+                {"params": unet_params}, noisy_latents, timesteps, encoder_hidden_states, train=False
+            )
+            noise_pred = unet_outputs.sample
+            loss = (noise - noise_pred) ** 2
+            loss = loss.mean()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        # Keep the token embeddings fixed except the newly added embeddings for the concept,
+        # as we only want to optimize the concept embeddings
+        token_embeds = original_token_embeds.at[placeholder_token_id].set(
+            new_state.params["text_model"]["embeddings"]["token_embedding"]["embedding"][placeholder_token_id]
+        )
+        new_state.params["text_model"]["embeddings"]["token_embedding"]["embedding"] = token_embeds
+
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return new_state, metrics, new_train_rng
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+    vae_params = jax_utils.replicate(vae_params)
+    unet_params = jax_utils.replicate(unet_params)
+
+    # Train!
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader))
+
+    # Scheduler and math around the number of training steps.
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {total_train_batch_size}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+
+    global_step = 0
+
+    epochs = tqdm(range(args.num_train_epochs), desc=f"Epoch ... (1/{args.num_train_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+
+        train_metrics = []
+
+        steps_per_epoch = len(train_dataset) // total_train_batch_size
+        train_step_progress_bar = tqdm(total=steps_per_epoch, desc="Training...", position=1, leave=False)
+        # train
+        for batch in train_dataloader:
+            batch = shard(batch)
+            state, train_metric, train_rngs = p_train_step(state, vae_params, unet_params, batch, train_rngs)
+            train_metrics.append(train_metric)
+
+            train_step_progress_bar.update(1)
+            global_step += 1
+
+            if global_step >= args.max_train_steps:
+                break
+
+        train_metric = jax_utils.unreplicate(train_metric)
+
+        train_step_progress_bar.close()
+        epochs.write(f"Epoch... ({epoch + 1}/{args.num_train_epochs} | Loss: {train_metric['loss']})")
+
+    # Create the pipeline using using the trained modules and save it.
+    if jax.process_index() == 0:
+        scheduler = FlaxPNDMScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
+        )
+        safety_checker = FlaxStableDiffusionSafetyChecker.from_pretrained(
+            "CompVis/stable-diffusion-safety-checker", from_pt=True
+        )
+        pipeline = FlaxStableDiffusionPipeline(
+            text_encoder=text_encoder,
+            vae=vae,
+            unet=unet,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=CLIPFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32"),
+        )
+
+        pipeline.save_pretrained(
+            args.output_dir,
+            params={
+                "text_encoder": get_params_to_save(state.params),
+                "vae": get_params_to_save(vae_params),
+                "unet": get_params_to_save(unet_params),
+                "safety_checker": safety_checker.params,
+            },
+        )
+
+        # Also save the newly trained embeddings
+        learned_embeds = get_params_to_save(state.params)["text_model"]["embeddings"]["token_embedding"]["embedding"][
+            placeholder_token_id
+        ]
+        learned_embeds_dict = {args.placeholder_token: learned_embeds}
+        jnp.save(os.path.join(args.output_dir, "learned_embeds.npy"), learned_embeds_dict)
+
+        if args.push_to_hub:
+            repo.push_to_hub(commit_message="End of training", blocking=False, auto_lfs_prune=True)
+
+
+if __name__ == "__main__":
+    main()

examples/unconditional_image_generation/README.md

@@ -7,7 +7,7 @@ Creating a training image set is [described in a different document](https://hug
 Before running the scripts, make sure to install the library's training dependencies:
 
 ```bash
-pip install diffusers[training] accelerate datasets
+pip install diffusers[training] accelerate datasets tensorboard
 ```
 
 And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
@@ -127,3 +127,24 @@ dataset.push_to_hub("name_of_your_dataset", private=True)
 and that's it! You can now train your model by simply setting the `--dataset_name` argument to the name of your dataset on the hub.
 
 More on this can also be found in [this blog post](https://huggingface.co/blog/image-search-datasets).
+
+#### Use ONNXRuntime to accelerate training
+
+In order to leverage onnxruntime to accelerate training, please use train_unconditional_ort.py
+
+The command to train a DDPM UNet model on the Oxford Flowers dataset with onnxruntime:
+
+```bash
+accelerate launch train_unconditional_ort.py \
+  --dataset_name="huggan/flowers-102-categories" \
+  --resolution=64 \
+  --output_dir="ddpm-ema-flowers-64" \
+  --train_batch_size=16 \
+  --num_epochs=1 \
+  --gradient_accumulation_steps=1 \
+  --learning_rate=1e-4 \
+  --lr_warmup_steps=500 \
+  --mixed_precision=fp16
+  ```
+
+Please contact Prathik Rao (prathikr), Sunghoon Choi (hanbitmyths), Ashwini Khade (askhade), or Peng Wang (pengwa) on github with any questions.

examples/unconditional_image_generation/train_unconditional.py

@@ -1,6 +1,9 @@
 import argparse
+import inspect
 import math
 import os
+from pathlib import Path
+from typing import Optional
 
 import torch
 import torch.nn.functional as F
@@ -8,10 +11,12 @@ import torch.nn.functional as F
 from accelerate import Accelerator
 from accelerate.logging import get_logger
 from datasets import load_dataset
-from diffusers import DDPMPipeline, DDPMScheduler, UNet2DModel
-from diffusers.hub_utils import init_git_repo, push_to_hub
+from diffusers import DDPMPipeline, DDPMScheduler, UNet2DModel, __version__
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import EMAModel
+from diffusers.utils import deprecate
+from huggingface_hub import HfFolder, Repository, whoami
+from packaging import version
 from torchvision.transforms import (
     CenterCrop,
     Compose,
@@ -25,6 +30,198 @@ from tqdm.auto import tqdm
 
 
 logger = get_logger(__name__)
+diffusers_version = version.parse(version.parse(__version__).base_version)
+
+
+def _extract_into_tensor(arr, timesteps, broadcast_shape):
+    """
+    Extract values from a 1-D numpy array for a batch of indices.
+
+    :param arr: the 1-D numpy array.
+    :param timesteps: a tensor of indices into the array to extract.
+    :param broadcast_shape: a larger shape of K dimensions with the batch
+                            dimension equal to the length of timesteps.
+    :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims.
+    """
+    if not isinstance(arr, torch.Tensor):
+        arr = torch.from_numpy(arr)
+    res = arr[timesteps].float().to(timesteps.device)
+    while len(res.shape) < len(broadcast_shape):
+        res = res[..., None]
+    return res.expand(broadcast_shape)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that HF Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="ddpm-model-64",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--overwrite_output_dir", action="store_true")
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=64,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--eval_batch_size", type=int, default=16, help="The number of images to generate for evaluation."
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "The number of subprocesses to use for data loading. 0 means that the data will be loaded in the main"
+            " process."
+        ),
+    )
+    parser.add_argument("--num_epochs", type=int, default=100)
+    parser.add_argument("--save_images_epochs", type=int, default=10, help="How often to save images during training.")
+    parser.add_argument(
+        "--save_model_epochs", type=int, default=10, help="How often to save the model during training."
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="cosine",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.95, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument(
+        "--adam_weight_decay", type=float, default=1e-6, help="Weight decay magnitude for the Adam optimizer."
+    )
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer.")
+    parser.add_argument(
+        "--use_ema",
+        action="store_true",
+        default=True,
+        help="Whether to use Exponential Moving Average for the final model weights.",
+    )
+    parser.add_argument("--ema_inv_gamma", type=float, default=1.0, help="The inverse gamma value for the EMA decay.")
+    parser.add_argument("--ema_power", type=float, default=3 / 4, help="The power value for the EMA decay.")
+    parser.add_argument("--ema_max_decay", type=float, default=0.9999, help="The maximum decay magnitude for EMA.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--hub_private_repo", action="store_true", help="Whether or not to create a private repository."
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."
+        ),
+    )
+
+    parser.add_argument(
+        "--predict_epsilon",
+        action="store_true",
+        default=True,
+        help="Whether the model should predict the 'epsilon'/noise error or directly the reconstructed image 'x0'.",
+    )
+
+    parser.add_argument("--ddpm_num_steps", type=int, default=1000)
+    parser.add_argument("--ddpm_beta_schedule", type=str, default="linear")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("You must specify either a dataset name from the hub or a train data directory.")
+
+    return args
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
 
 
 def main(args):
@@ -59,7 +256,17 @@ def main(args):
             "UpBlock2D",
         ),
     )
-    noise_scheduler = DDPMScheduler(num_train_timesteps=1000)
+    accepts_predict_epsilon = "predict_epsilon" in set(inspect.signature(DDPMScheduler.__init__).parameters.keys())
+
+    if accepts_predict_epsilon:
+        noise_scheduler = DDPMScheduler(
+            num_train_timesteps=args.ddpm_num_steps,
+            beta_schedule=args.ddpm_beta_schedule,
+            predict_epsilon=args.predict_epsilon,
+        )
+    else:
+        noise_scheduler = DDPMScheduler(num_train_timesteps=args.ddpm_num_steps, beta_schedule=args.ddpm_beta_schedule)
+
     optimizer = torch.optim.AdamW(
         model.parameters(),
         lr=args.learning_rate,
@@ -92,8 +299,12 @@ def main(args):
         images = [augmentations(image.convert("RGB")) for image in examples["image"]]
         return {"input": images}
 
+    logger.info(f"Dataset size: {len(dataset)}")
+
     dataset.set_transform(transforms)
-    train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.train_batch_size, shuffle=True)
+    train_dataloader = torch.utils.data.DataLoader(
+        dataset, batch_size=args.train_batch_size, shuffle=True, num_workers=args.dataloader_num_workers
+    )
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
@@ -110,8 +321,22 @@ def main(args):
 
     ema_model = EMAModel(model, inv_gamma=args.ema_inv_gamma, power=args.ema_power, max_value=args.ema_max_decay)
 
-    if args.push_to_hub:
-        repo = init_git_repo(args, at_init=True)
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.push_to_hub:
+            if args.hub_model_id is None:
+                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
+            else:
+                repo_name = args.hub_model_id
+            repo = Repository(args.output_dir, clone_from=repo_name)
+
+            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
+                if "step_*" not in gitignore:
+                    gitignore.write("step_*\n")
+                if "epoch_*" not in gitignore:
+                    gitignore.write("epoch_*\n")
+        elif args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
 
     if accelerator.is_main_process:
         run = os.path.split(__file__)[-1].split(".")[0]
@@ -138,8 +363,20 @@ def main(args):
 
             with accelerator.accumulate(model):
                 # Predict the noise residual
-                noise_pred = model(noisy_images, timesteps).sample
-                loss = F.mse_loss(noise_pred, noise)
+                model_output = model(noisy_images, timesteps).sample
+
+                if args.predict_epsilon:
+                    loss = F.mse_loss(model_output, noise)  # this could have different weights!
+                else:
+                    alpha_t = _extract_into_tensor(
+                        noise_scheduler.alphas_cumprod, timesteps, (clean_images.shape[0], 1, 1, 1)
+                    )
+                    snr_weights = alpha_t / (1 - alpha_t)
+                    loss = snr_weights * F.mse_loss(
+                        model_output, clean_images, reduction="none"
+                    )  # use SNR weighting from distillation paper
+                    loss = loss.mean()
+
                 accelerator.backward(loss)
 
                 if accelerator.sync_gradients:
@@ -172,9 +409,17 @@ def main(args):
                     scheduler=noise_scheduler,
                 )
 
-                generator = torch.manual_seed(0)
+                deprecate("todo: remove this check", "0.10.0", "when the most used version is >= 0.8.0")
+                if diffusers_version < version.parse("0.8.0"):
+                    generator = torch.manual_seed(0)
+                else:
+                    generator = torch.Generator(device=pipeline.device).manual_seed(0)
                 # run pipeline in inference (sample random noise and denoise)
-                images = pipeline(generator=generator, batch_size=args.eval_batch_size, output_type="numpy").images
+                images = pipeline(
+                    generator=generator,
+                    batch_size=args.eval_batch_size,
+                    output_type="numpy",
+                ).images
 
                 # denormalize the images and save to tensorboard
                 images_processed = (images * 255).round().astype("uint8")
@@ -184,65 +429,14 @@ def main(args):
 
             if epoch % args.save_model_epochs == 0 or epoch == args.num_epochs - 1:
                 # save the model
+                pipeline.save_pretrained(args.output_dir)
                 if args.push_to_hub:
-                    push_to_hub(args, pipeline, repo, commit_message=f"Epoch {epoch}", blocking=False)
-                else:
-                    pipeline.save_pretrained(args.output_dir)
+                    repo.push_to_hub(commit_message=f"Epoch {epoch}", blocking=False)
         accelerator.wait_for_everyone()
 
     accelerator.end_training()
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Simple example of a training script.")
-    parser.add_argument("--local_rank", type=int, default=-1)
-    parser.add_argument("--dataset_name", type=str, default=None)
-    parser.add_argument("--dataset_config_name", type=str, default=None)
-    parser.add_argument("--train_data_dir", type=str, default=None, help="A folder containing the training data.")
-    parser.add_argument("--output_dir", type=str, default="ddpm-model-64")
-    parser.add_argument("--overwrite_output_dir", action="store_true")
-    parser.add_argument("--cache_dir", type=str, default=None)
-    parser.add_argument("--resolution", type=int, default=64)
-    parser.add_argument("--train_batch_size", type=int, default=16)
-    parser.add_argument("--eval_batch_size", type=int, default=16)
-    parser.add_argument("--num_epochs", type=int, default=100)
-    parser.add_argument("--save_images_epochs", type=int, default=10)
-    parser.add_argument("--save_model_epochs", type=int, default=10)
-    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
-    parser.add_argument("--learning_rate", type=float, default=1e-4)
-    parser.add_argument("--lr_scheduler", type=str, default="cosine")
-    parser.add_argument("--lr_warmup_steps", type=int, default=500)
-    parser.add_argument("--adam_beta1", type=float, default=0.95)
-    parser.add_argument("--adam_beta2", type=float, default=0.999)
-    parser.add_argument("--adam_weight_decay", type=float, default=1e-6)
-    parser.add_argument("--adam_epsilon", type=float, default=1e-08)
-    parser.add_argument("--use_ema", action="store_true", default=True)
-    parser.add_argument("--ema_inv_gamma", type=float, default=1.0)
-    parser.add_argument("--ema_power", type=float, default=3 / 4)
-    parser.add_argument("--ema_max_decay", type=float, default=0.9999)
-    parser.add_argument("--push_to_hub", action="store_true")
-    parser.add_argument("--hub_token", type=str, default=None)
-    parser.add_argument("--hub_model_id", type=str, default=None)
-    parser.add_argument("--hub_private_repo", action="store_true")
-    parser.add_argument("--logging_dir", type=str, default="logs")
-    parser.add_argument(
-        "--mixed_precision",
-        type=str,
-        default="no",
-        choices=["no", "fp16", "bf16"],
-        help=(
-            "Whether to use mixed precision. Choose"
-            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
-            "and an Nvidia Ampere GPU."
-        ),
-    )
-
-    args = parser.parse_args()
-    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
-    if env_local_rank != -1 and env_local_rank != args.local_rank:
-        args.local_rank = env_local_rank
-
-    if args.dataset_name is None and args.train_data_dir is None:
-        raise ValueError("You must specify either a dataset name from the hub or a train data directory.")
-
+    args = parse_args()
     main(args)

examples/unconditional_image_generation/train_unconditional_ort.py

@@ -0,0 +1,251 @@
+import argparse
+import math
+import os
+
+import torch
+import torch.nn.functional as F
+
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from datasets import load_dataset
+from diffusers import DDPMPipeline, DDPMScheduler, UNet2DModel
+from diffusers.hub_utils import init_git_repo, push_to_hub
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import EMAModel
+from onnxruntime.training.ortmodule import ORTModule
+from torchvision.transforms import (
+    CenterCrop,
+    Compose,
+    InterpolationMode,
+    Normalize,
+    RandomHorizontalFlip,
+    Resize,
+    ToTensor,
+)
+from tqdm.auto import tqdm
+
+
+logger = get_logger(__name__)
+
+
+def main(args):
+    logging_dir = os.path.join(args.output_dir, args.logging_dir)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with="tensorboard",
+        logging_dir=logging_dir,
+    )
+
+    model = UNet2DModel(
+        sample_size=args.resolution,
+        in_channels=3,
+        out_channels=3,
+        layers_per_block=2,
+        block_out_channels=(128, 128, 256, 256, 512, 512),
+        down_block_types=(
+            "DownBlock2D",
+            "DownBlock2D",
+            "DownBlock2D",
+            "DownBlock2D",
+            "AttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        up_block_types=(
+            "UpBlock2D",
+            "AttnUpBlock2D",
+            "UpBlock2D",
+            "UpBlock2D",
+            "UpBlock2D",
+            "UpBlock2D",
+        ),
+    )
+    model = ORTModule(model)
+    noise_scheduler = DDPMScheduler(num_train_timesteps=1000, tensor_format="pt")
+    optimizer = torch.optim.AdamW(
+        model.parameters(),
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    augmentations = Compose(
+        [
+            Resize(args.resolution, interpolation=InterpolationMode.BILINEAR),
+            CenterCrop(args.resolution),
+            RandomHorizontalFlip(),
+            ToTensor(),
+            Normalize([0.5], [0.5]),
+        ]
+    )
+
+    if args.dataset_name is not None:
+        dataset = load_dataset(
+            args.dataset_name,
+            args.dataset_config_name,
+            cache_dir=args.cache_dir,
+            use_auth_token=True if args.use_auth_token else None,
+            split="train",
+        )
+    else:
+        dataset = load_dataset("imagefolder", data_dir=args.train_data_dir, cache_dir=args.cache_dir, split="train")
+
+    def transforms(examples):
+        images = [augmentations(image.convert("RGB")) for image in examples["image"]]
+        return {"input": images}
+
+    dataset.set_transform(transforms)
+    train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.train_batch_size, shuffle=True)
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps,
+        num_training_steps=(len(train_dataloader) * args.num_epochs) // args.gradient_accumulation_steps,
+    )
+
+    model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        model, optimizer, train_dataloader, lr_scheduler
+    )
+
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+
+    ema_model = EMAModel(model, inv_gamma=args.ema_inv_gamma, power=args.ema_power, max_value=args.ema_max_decay)
+
+    if args.push_to_hub:
+        repo = init_git_repo(args, at_init=True)
+
+    if accelerator.is_main_process:
+        run = os.path.split(__file__)[-1].split(".")[0]
+        accelerator.init_trackers(run)
+
+    global_step = 0
+    for epoch in range(args.num_epochs):
+        model.train()
+        progress_bar = tqdm(total=num_update_steps_per_epoch, disable=not accelerator.is_local_main_process)
+        progress_bar.set_description(f"Epoch {epoch}")
+        for step, batch in enumerate(train_dataloader):
+            clean_images = batch["input"]
+            # Sample noise that we'll add to the images
+            noise = torch.randn(clean_images.shape).to(clean_images.device)
+            bsz = clean_images.shape[0]
+            # Sample a random timestep for each image
+            timesteps = torch.randint(
+                0, noise_scheduler.config.num_train_timesteps, (bsz,), device=clean_images.device
+            ).long()
+
+            # Add noise to the clean images according to the noise magnitude at each timestep
+            # (this is the forward diffusion process)
+            noisy_images = noise_scheduler.add_noise(clean_images, noise, timesteps)
+
+            with accelerator.accumulate(model):
+                # Predict the noise residual
+                noise_pred = model(noisy_images, timesteps, return_dict=True)[0]
+                loss = F.mse_loss(noise_pred, noise)
+                accelerator.backward(loss)
+
+                accelerator.clip_grad_norm_(model.parameters(), 1.0)
+                optimizer.step()
+                lr_scheduler.step()
+                if args.use_ema:
+                    ema_model.step(model)
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0], "step": global_step}
+            if args.use_ema:
+                logs["ema_decay"] = ema_model.decay
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+        progress_bar.close()
+
+        accelerator.wait_for_everyone()
+
+        # Generate sample images for visual inspection
+        if accelerator.is_main_process:
+            if epoch % args.save_images_epochs == 0 or epoch == args.num_epochs - 1:
+                pipeline = DDPMPipeline(
+                    unet=accelerator.unwrap_model(ema_model.averaged_model if args.use_ema else model),
+                    scheduler=noise_scheduler,
+                )
+
+                generator = torch.manual_seed(0)
+                # run pipeline in inference (sample random noise and denoise)
+                images = pipeline(generator=generator, batch_size=args.eval_batch_size, output_type="numpy").images
+
+                # denormalize the images and save to tensorboard
+                images_processed = (images * 255).round().astype("uint8")
+                accelerator.trackers[0].writer.add_images(
+                    "test_samples", images_processed.transpose(0, 3, 1, 2), epoch
+                )
+
+            if epoch % args.save_model_epochs == 0 or epoch == args.num_epochs - 1:
+                # save the model
+                if args.push_to_hub:
+                    push_to_hub(args, pipeline, repo, commit_message=f"Epoch {epoch}", blocking=False)
+                else:
+                    pipeline.save_pretrained(args.output_dir)
+        accelerator.wait_for_everyone()
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument("--local_rank", type=int, default=-1)
+    parser.add_argument("--dataset_name", type=str, default=None)
+    parser.add_argument("--dataset_config_name", type=str, default=None)
+    parser.add_argument("--train_data_dir", type=str, default=None, help="A folder containing the training data.")
+    parser.add_argument("--output_dir", type=str, default="ddpm-model-64")
+    parser.add_argument("--overwrite_output_dir", action="store_true")
+    parser.add_argument("--cache_dir", type=str, default=None)
+    parser.add_argument("--resolution", type=int, default=64)
+    parser.add_argument("--train_batch_size", type=int, default=16)
+    parser.add_argument("--eval_batch_size", type=int, default=16)
+    parser.add_argument("--num_epochs", type=int, default=100)
+    parser.add_argument("--save_images_epochs", type=int, default=10)
+    parser.add_argument("--save_model_epochs", type=int, default=10)
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
+    parser.add_argument("--learning_rate", type=float, default=1e-4)
+    parser.add_argument("--lr_scheduler", type=str, default="cosine")
+    parser.add_argument("--lr_warmup_steps", type=int, default=500)
+    parser.add_argument("--adam_beta1", type=float, default=0.95)
+    parser.add_argument("--adam_beta2", type=float, default=0.999)
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-6)
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08)
+    parser.add_argument("--use_ema", action="store_true", default=True)
+    parser.add_argument("--ema_inv_gamma", type=float, default=1.0)
+    parser.add_argument("--ema_power", type=float, default=3 / 4)
+    parser.add_argument("--ema_max_decay", type=float, default=0.9999)
+    parser.add_argument("--push_to_hub", action="store_true")
+    parser.add_argument("--use_auth_token", action="store_true")
+    parser.add_argument("--hub_token", type=str, default=None)
+    parser.add_argument("--hub_model_id", type=str, default=None)
+    parser.add_argument("--hub_private_repo", action="store_true")
+    parser.add_argument("--logging_dir", type=str, default="logs")
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose"
+            "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
+            "and an Nvidia Ampere GPU."
+        ),
+    )
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("You must specify either a dataset name from the hub or a train data directory.")
+
+    main(args)