Revert "[LoRA] introduce LoraBaseMixin to promote reusability. (#8774)"

This reverts commit 527430d0a4.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -63,23 +63,27 @@ body:
 
         Please tag a maximum of 2 people.
 
-        Questions on DiffusionPipeline (Saving, Loading, From pretrained, ...):
+        Questions on DiffusionPipeline (Saving, Loading, From pretrained, ...): @sayakpaul @DN6
 
         Questions on pipelines:
-        - Stable Diffusion @yiyixuxu @DN6 @sayakpaul
+        - Stable Diffusion @yiyixuxu @asomoza
         - Stable Diffusion XL @yiyixuxu @sayakpaul @DN6
+        - Stable Diffusion 3: @yiyixuxu @sayakpaul @DN6 @asomoza
         - Kandinsky @yiyixuxu
         - ControlNet @sayakpaul @yiyixuxu @DN6
         - T2I Adapter @sayakpaul @yiyixuxu @DN6
         - IF @DN6
-        - Text-to-Video / Video-to-Video @DN6 @sayakpaul
+        - Text-to-Video / Video-to-Video @DN6 @a-r-r-o-w
         - Wuerstchen @DN6
         - Other: @yiyixuxu @DN6
+        - Improving generation quality: @asomoza
 
         Questions on models:
         - UNet @DN6 @yiyixuxu @sayakpaul
         - VAE @sayakpaul @DN6 @yiyixuxu
-        - Transformers/Attention @DN6 @yiyixuxu @sayakpaul @DN6
+        - Transformers/Attention @DN6 @yiyixuxu @sayakpaul
+
+        Questions on single file checkpoints: @DN6
 
         Questions on Schedulers: @yiyixuxu
 
@@ -99,7 +103,7 @@ body:
 
         Questions on JAX- and MPS-related things: @pcuenca
 
-        Questions on audio pipelines: @DN6
+        Questions on audio pipelines: @sanchit-gandhi
 
 
 

.github/PULL_REQUEST_TEMPLATE.md

@@ -39,7 +39,7 @@ members/contributors who may be interested in your PR.
 Core library:
 
 - Schedulers: @yiyixuxu
-- Pipelines:  @sayakpaul @yiyixuxu @DN6
+- Pipelines and pipeline callbacks: @yiyixuxu and @asomoza
 - Training examples: @sayakpaul
 - Docs: @stevhliu and @sayakpaul
 - JAX and MPS: @pcuenca
@@ -48,7 +48,8 @@ Core library:
 
 Integrations:
 
-- deepspeed: HF Trainer/Accelerate: @pacman100
+- deepspeed: HF Trainer/Accelerate: @SunMarc
+- PEFT: @sayakpaul @BenjaminBossan
 
 HF projects:
 

.github/workflows/benchmark.yml

@@ -13,14 +13,17 @@ env:
 
 jobs:
   torch_pipelines_cuda_benchmark_tests:
+    env: 
+      SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL_BENCHMARK }}
     name: Torch Core Pipelines CUDA Benchmarking Tests
     strategy:
       fail-fast: false
       max-parallel: 1
-    runs-on: [single-gpu, nvidia-gpu, a10, ci]
+    runs-on: 
+      group: aws-g6-4xlarge-plus
     container:
-      image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+      image: diffusers/diffusers-pytorch-compile-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3
@@ -50,4 +53,14 @@ jobs:
         uses: actions/upload-artifact@v2
         with:
           name: benchmark_test_reports
-          path: benchmarks/benchmark_outputs
+          path: benchmarks/benchmark_outputs
+
+      - name: Report success status
+        if: ${{ success() }}
+        run: |
+          pip install requests && python utils/notify_benchmarking_status.py --status=success
+        
+      - name: Report failure status
+        if: ${{ failure() }}
+        run: |
+          pip install requests && python utils/notify_benchmarking_status.py --status=failure

.github/workflows/mirror_community_pipeline.yml

@@ -22,6 +22,9 @@ on:
 
 jobs:
   mirror_community_pipeline:
+    env:
+      SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL_COMMUNITY_MIRROR }}
+    
     runs-on: ubuntu-latest
     steps:
       # Checkout to correct ref
@@ -86,4 +89,14 @@ jobs:
         run: huggingface-cli upload diffusers/community-pipelines-mirror ./examples/community ${PATH_IN_REPO} --repo-type dataset
         env:
             PATH_IN_REPO: ${{ env.PATH_IN_REPO }}
-            HF_TOKEN: ${{ secrets.HF_TOKEN_MIRROR_COMMUNITY_PIPELINES }}
+            HF_TOKEN: ${{ secrets.HF_TOKEN_MIRROR_COMMUNITY_PIPELINES }}
+
+      - name: Report success status
+        if: ${{ success() }}
+        run: |
+          pip install requests && python utils/notify_community_pipelines_mirror.py --status=success
+      
+      - name: Report failure status
+        if: ${{ failure() }}
+        run: |
+          pip install requests && python utils/notify_community_pipelines_mirror.py --status=failure

.github/workflows/nightly_tests.yml

@@ -7,7 +7,7 @@ on:
 
 env:
   DIFFUSERS_IS_CI: yes
-  HF_HOME: /mnt/cache
+  HF_HUB_ENABLE_HF_TRANSFER: 1
   OMP_NUM_THREADS: 8
   MKL_NUM_THREADS: 8
   PYTEST_TIMEOUT: 600
@@ -27,10 +27,6 @@ jobs:
         uses: actions/checkout@v3
         with:
           fetch-depth: 2
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
       - name: Install dependencies
         run: |
           pip install -e .
@@ -50,16 +46,17 @@ jobs:
           path: reports
 
   run_nightly_tests_for_torch_pipelines:
-    name: Torch Pipelines CUDA Nightly Tests
+    name: Nightly Torch Pipelines CUDA Tests
     needs: setup_torch_cuda_pipeline_matrix
     strategy:
       fail-fast: false
+      max-parallel: 8
       matrix:
         module: ${{ fromJson(needs.setup_torch_cuda_pipeline_matrix.outputs.pipeline_test_matrix) }}
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
+      options: --shm-size "16gb" --ipc host --gpus 0
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3
@@ -67,19 +64,16 @@ jobs:
           fetch-depth: 2
       - name: NVIDIA-SMI
         run: nvidia-smi
-
       - name: Install dependencies
         run: |
           python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           python -m uv pip install -e [quality,test]
           python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
           python -m uv pip install pytest-reportlog
-
       - name: Environment
         run: |
           python utils/print_env.py
-
-      - name: Nightly PyTorch CUDA checkpoint (pipelines) tests
+      - name: Pipeline CUDA Test
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
@@ -90,38 +84,36 @@ jobs:
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             --report-log=tests_pipeline_${{ matrix.module }}_cuda.log \
             tests/pipelines/${{ matrix.module }}
-
       - name: Failure short reports
         if: ${{ failure() }}
         run: |
           cat reports/tests_pipeline_${{ matrix.module }}_cuda_stats.txt
           cat reports/tests_pipeline_${{ matrix.module }}_cuda_failures_short.txt
-
       - name: Test suite reports artifacts
         if: ${{ always() }}
         uses: actions/upload-artifact@v2
         with:
           name: pipeline_${{ matrix.module }}_test_reports
           path: reports
-
       - name: Generate Report and Notify Channel
         if: always()
         run: |
           pip install slack_sdk tabulate
-          python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_tests_for_other_torch_modules:
-    name: Torch Non-Pipelines CUDA Nightly Tests
+    name: Nightly Torch CUDA Tests
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+      options: --shm-size "16gb" --ipc host --gpus 0
     defaults:
       run:
         shell: bash
     strategy:
       matrix:
-        module: [models, schedulers, others, examples]
+        max-parallel: 2
+        module: [models, schedulers, lora, others, single_file, examples]
     steps:
     - name: Checkout diffusers
       uses: actions/checkout@v3
@@ -133,8 +125,8 @@ jobs:
         python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m uv pip install -e [quality,test]
         python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
         python -m uv pip install pytest-reportlog
-
     - name: Environment
       run: python utils/print_env.py
 
@@ -158,7 +150,6 @@ jobs:
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
-        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
           -s -v --make-reports=examples_torch_cuda \
           --report-log=examples_torch_cuda.log \
@@ -181,64 +172,7 @@ jobs:
       if: always()
       run: |
         pip install slack_sdk tabulate
-        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
-
-  run_lora_nightly_tests:
-    name: Nightly LoRA Tests with PEFT and TORCH
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
-    container:
-      image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
-    defaults:
-      run:
-        shell: bash
-    steps:
-    - name: Checkout diffusers
-      uses: actions/checkout@v3
-      with:
-        fetch-depth: 2
-
-    - name: Install dependencies
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
-        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
-        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
-        python -m uv pip install pytest-reportlog
-
-    - name: Environment
-      run: python utils/print_env.py
-
-    - name: Run nightly LoRA tests with PEFT and Torch
-      env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
-        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
-        CUBLAS_WORKSPACE_CONFIG: :16:8
-      run: |
-        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
-          --make-reports=tests_torch_lora_cuda \
-          --report-log=tests_torch_lora_cuda.log \
-          tests/lora
-
-    - name: Failure short reports
-      if: ${{ failure() }}
-      run: |
-        cat reports/tests_torch_lora_cuda_stats.txt
-        cat reports/tests_torch_lora_cuda_failures_short.txt
-
-    - name: Test suite reports artifacts
-      if: ${{ always() }}
-      uses: actions/upload-artifact@v2
-      with:
-        name: torch_lora_cuda_test_reports
-        path: reports
-
-    - name: Generate Report and Notify Channel
-      if: always()
-      run: |
-        pip install slack_sdk tabulate
-        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_flax_tpu_tests:
     name: Nightly Flax TPU Tests
@@ -294,14 +228,14 @@ jobs:
       if: always()
       run: |
         pip install slack_sdk tabulate
-        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_onnx_tests:
     name: Nightly ONNXRuntime CUDA tests on Ubuntu
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-onnxruntime-cuda
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
+      options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:
     - name: Checkout diffusers
@@ -318,11 +252,10 @@ jobs:
         python -m uv pip install -e [quality,test]
         python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
         python -m uv pip install pytest-reportlog
-
     - name: Environment
       run: python utils/print_env.py
 
-    - name: Run nightly ONNXRuntime CUDA tests
+    - name: Run Nightly ONNXRuntime CUDA tests
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |
@@ -349,7 +282,7 @@ jobs:
       if: always()
       run: |
         pip install slack_sdk tabulate
-        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_tests_apple_m1:
     name: Nightly PyTorch MPS tests on MacOS
@@ -411,4 +344,4 @@ jobs:
         if: always()
         run: |
           pip install slack_sdk tabulate
-          python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/push_tests.yml

@@ -11,11 +11,9 @@ on:
 
 env:
   DIFFUSERS_IS_CI: yes
-  HF_HOME: /mnt/cache
   OMP_NUM_THREADS: 8
   MKL_NUM_THREADS: 8
   PYTEST_TIMEOUT: 600
-  RUN_SLOW: yes
   PIPELINE_USAGE_CUTOFF: 50000
 
 jobs:
@@ -52,7 +50,7 @@ jobs:
           path: reports
 
   torch_pipelines_cuda_tests:
-    name: Torch Pipelines CUDA Slow Tests
+    name: Torch Pipelines CUDA Tests
     needs: setup_torch_cuda_pipeline_matrix
     strategy:
       fail-fast: false
@@ -62,7 +60,7 @@ jobs:
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
+      options: --shm-size "16gb" --ipc host --gpus 0
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3
@@ -106,7 +104,7 @@ jobs:
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
+      options: --shm-size "16gb" --ipc host --gpus 0
     defaults:
       run:
         shell: bash
@@ -124,12 +122,13 @@ jobs:
         python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m uv pip install -e [quality,test]
         python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
 
     - name: Environment
       run: |
         python utils/print_env.py
 
-    - name: Run slow PyTorch CUDA tests
+    - name: Run PyTorch CUDA tests
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
@@ -153,61 +152,6 @@ jobs:
         name: torch_cuda_test_reports
         path: reports
 
-  peft_cuda_tests:
-    name: PEFT CUDA Tests
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
-    container:
-      image: diffusers/diffusers-pytorch-cuda
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --gpus 0
-    defaults:
-      run:
-        shell: bash
-    steps:
-    - name: Checkout diffusers
-      uses: actions/checkout@v3
-      with:
-        fetch-depth: 2
-
-    - name: Install dependencies
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
-        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
-        python -m pip install -U peft@git+https://github.com/huggingface/peft.git
-
-    - name: Environment
-      run: |
-        python utils/print_env.py
-
-    - name: Run slow PEFT CUDA tests
-      env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
-        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
-        CUBLAS_WORKSPACE_CONFIG: :16:8
-      run: |
-        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx and not PEFTLoRALoading" \
-          --make-reports=tests_peft_cuda \
-          tests/lora/
-        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "lora and not Flax and not Onnx and not PEFTLoRALoading" \
-          --make-reports=tests_peft_cuda_models_lora \
-          tests/models/
-
-    - name: Failure short reports
-      if: ${{ failure() }}
-      run: |
-        cat reports/tests_peft_cuda_stats.txt
-        cat reports/tests_peft_cuda_failures_short.txt
-        cat reports/tests_peft_cuda_models_lora_failures_short.txt
-
-    - name: Test suite reports artifacts
-      if: ${{ always() }}
-      uses: actions/upload-artifact@v2
-      with:
-        name: torch_peft_test_reports
-        path: reports
-
   flax_tpu_tests:
     name: Flax TPU Tests
     runs-on: docker-tpu
@@ -309,7 +253,7 @@ jobs:
 
     container:
       image: diffusers/diffusers-pytorch-compile-cuda
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+      options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:
     - name: Checkout diffusers
@@ -330,6 +274,7 @@ jobs:
     - name: Run example tests on GPU
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        RUN_COMPILE: yes
       run: |
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
     - name: Failure short reports
@@ -350,7 +295,7 @@ jobs:
 
     container:
       image: diffusers/diffusers-pytorch-xformers-cuda
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+      options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:
     - name: Checkout diffusers
@@ -391,7 +336,7 @@ jobs:
 
     container:
       image: diffusers/diffusers-pytorch-cuda
-      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+      options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:
     - name: Checkout diffusers

.github/workflows/ssh-pr-runner.yml

@@ -0,0 +1,39 @@
+name: SSH into PR runners
+
+on:
+  workflow_dispatch:
+    inputs:
+      docker_image:
+        description: 'Name of the Docker image'
+        required: true
+
+env:
+  IS_GITHUB_CI: "1"
+  HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  HF_HOME: /mnt/cache
+  DIFFUSERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  RUN_SLOW: yes
+
+jobs:
+  ssh_runner:
+    name: "SSH"
+    runs-on: [self-hosted, intel-cpu, 32-cpu, 256-ram, ci]
+    container:
+      image: ${{ github.event.inputs.docker_image }}
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface/diffusers:/mnt/cache/ --privileged
+
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+
+      - name: Tailscale # In order to be able to SSH when a test fails
+        uses: huggingface/tailscale-action@main
+        with:
+          authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
+          slackChannel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}
+          slackToken: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+          waitForSSH: true

.github/workflows/ssh-runner.yml

@@ -1,4 +1,4 @@
-name: SSH into runners
+name: SSH into GPU runners
 
 on:
   workflow_dispatch:

.gitignore

@@ -175,4 +175,4 @@ tags
 .ruff_cache
 
 # wandb
-wandb
+wandb

benchmarks/run_all.py

@@ -40,7 +40,7 @@ def main():
         print(f"****** Running file: {file} ******")
 
         # Run with canonical settings.
-        if file != "benchmark_text_to_image.py":
+        if file != "benchmark_text_to_image.py" and file != "benchmark_ip_adapters.py":
             command = f"python {file}"
             run_command(command.split())
 
@@ -49,6 +49,10 @@ def main():
 
     # Run variants.
     for file in python_files:
+        # See: https://github.com/pytorch/pytorch/issues/129637
+        if file == "benchmark_ip_adapters.py":
+            continue
+
         if file == "benchmark_text_to_image.py":
             for ckpt in ALL_T2I_CKPTS:
                 command = f"python {file} --ckpt {ckpt}"

docker/diffusers-onnxruntime-cuda/Dockerfile

@@ -38,6 +38,7 @@ RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
         datasets \
         hf-doc-builder \
         huggingface-hub \
+        hf_transfer \
         Jinja2 \
         librosa \
         numpy==1.26.4 \

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -17,6 +17,7 @@ RUN apt install -y bash \
     libsndfile1-dev \
     libgl1 \
     python3.10 \
+    python3.10-dev \
     python3-pip \
     python3.10-venv && \
     rm -rf /var/lib/apt/lists
@@ -37,6 +38,7 @@ RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     datasets \
     hf-doc-builder \
     huggingface-hub \
+    hf_transfer \
     Jinja2 \
     librosa \
     numpy==1.26.4 \

docker/diffusers-pytorch-cpu/Dockerfile

@@ -16,6 +16,7 @@ RUN apt install -y bash \
                    ca-certificates \
                    libsndfile1-dev \
                    python3.10 \
+                   python3.10-dev \
                    python3-pip \
                    libgl1 \
                    python3.10-venv && \

docker/diffusers-pytorch-cuda/Dockerfile

@@ -17,6 +17,7 @@ RUN apt install -y bash \
     libsndfile1-dev \
     libgl1 \
     python3.10 \
+    python3.10-dev \
     python3-pip \
     python3.10-venv && \
     rm -rf /var/lib/apt/lists
@@ -37,6 +38,7 @@ RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     datasets \
     hf-doc-builder \
     huggingface-hub \
+    hf_transfer \
     Jinja2 \
     librosa \
     numpy==1.26.4 \

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -17,6 +17,7 @@ RUN apt install -y bash \
                    libsndfile1-dev \
                    libgl1 \
                    python3.10 \
+                   python3.10-dev \
                    python3-pip \
                    python3.10-venv && \
     rm -rf /var/lib/apt/lists
@@ -37,6 +38,7 @@ RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
         datasets \
         hf-doc-builder \
         huggingface-hub \
+        hf_transfer \
         Jinja2 \
         librosa \
         numpy==1.26.4 \

docs/source/en/_toctree.yml

@@ -21,6 +21,8 @@
     title: Load LoRAs for inference
   - local: tutorials/fast_diffusion
     title: Accelerate inference of text-to-image diffusion models
+  - local: tutorials/inference_with_big_models
+    title: Working with big models
   title: Tutorials
 - sections:
   - local: using-diffusers/loading
@@ -247,6 +249,12 @@
       title: DiTTransformer2DModel
     - local: api/models/hunyuan_transformer2d
       title: HunyuanDiT2DModel
+    - local: api/models/aura_flow_transformer2d
+      title: AuraFlowTransformer2DModel
+    - local: api/models/latte_transformer3d
+      title: LatteTransformer3DModel
+    - local: api/models/lumina_nextdit2d
+      title: LuminaNextDiT2DModel
     - local: api/models/transformer_temporal
       title: TransformerTemporalModel
     - local: api/models/sd3_transformer2d
@@ -255,6 +263,8 @@
       title: PriorTransformer
     - local: api/models/controlnet
       title: ControlNetModel
+    - local: api/models/controlnet_hunyuandit
+      title: HunyuanDiT2DControlNetModel
     - local: api/models/controlnet_sd3
       title: SD3ControlNetModel
     title: Models
@@ -272,6 +282,8 @@
       title: AudioLDM
     - local: api/pipelines/audioldm2
       title: AudioLDM 2
+    - local: api/pipelines/aura_flow
+      title: AuraFlow
     - local: api/pipelines/auto_pipeline
       title: AutoPipeline
     - local: api/pipelines/blip_diffusion
@@ -280,6 +292,8 @@
       title: Consistency Models
     - local: api/pipelines/controlnet
       title: ControlNet
+    - local: api/pipelines/controlnet_hunyuandit
+      title: ControlNet with Hunyuan-DiT
     - local: api/pipelines/controlnet_sd3
       title: ControlNet with Stable Diffusion 3
     - local: api/pipelines/controlnet_sdxl
@@ -312,12 +326,18 @@
       title: Kandinsky 2.2
     - local: api/pipelines/kandinsky3
       title: Kandinsky 3
+    - local: api/pipelines/kolors
+      title: Kolors
     - local: api/pipelines/latent_consistency_models
       title: Latent Consistency Models
     - local: api/pipelines/latent_diffusion
       title: Latent Diffusion
+    - local: api/pipelines/latte
+      title: Latte
     - local: api/pipelines/ledits_pp
       title: LEDITS++
+    - local: api/pipelines/lumina
+      title: Lumina-T2X
     - local: api/pipelines/marigold
       title: Marigold
     - local: api/pipelines/panorama
@@ -429,6 +449,8 @@
       title: EulerDiscreteScheduler
     - local: api/schedulers/flow_match_euler_discrete
       title: FlowMatchEulerDiscreteScheduler
+    - local: api/schedulers/flow_match_heun_discrete
+      title: FlowMatchHeunDiscreteScheduler
     - local: api/schedulers/heun
       title: HeunDiscreteScheduler
     - local: api/schedulers/ipndm

docs/source/en/api/models/aura_flow_transformer2d.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 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.
+-->
+
+# AuraFlowTransformer2DModel
+
+A Transformer model for image-like data from [AuraFlow](https://blog.fal.ai/auraflow/).
+
+## AuraFlowTransformer2DModel
+
+[[autodoc]] AuraFlowTransformer2DModel

docs/source/en/api/models/autoencoderkl.md

@@ -21,7 +21,7 @@ The abstract from the paper is:
 ## Loading from the original format
 
 By default the [`AutoencoderKL`] should be loaded with [`~ModelMixin.from_pretrained`], but it can also be loaded
-from the original format using [`FromOriginalVAEMixin.from_single_file`] as follows:
+from the original format using [`FromOriginalModelMixin.from_single_file`] as follows:
 
 ```py
 from diffusers import AutoencoderKL

docs/source/en/api/models/controlnet.md

@@ -21,7 +21,7 @@ The abstract from the paper is:
 ## Loading from the original format
 
 By default the [`ControlNetModel`] should be loaded with [`~ModelMixin.from_pretrained`], but it can also be loaded
-from the original format using [`FromOriginalControlnetMixin.from_single_file`] as follows:
+from the original format using [`FromOriginalModelMixin.from_single_file`] as follows:
 
 ```py
 from diffusers import StableDiffusionControlNetPipeline, ControlNetModel

docs/source/en/api/models/controlnet_hunyuandit.md

@@ -0,0 +1,37 @@
+<!--Copyright 2024 The HuggingFace Team and Tencent Hunyuan 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.
+-->
+
+# HunyuanDiT2DControlNetModel
+
+HunyuanDiT2DControlNetModel is an implementation of ControlNet for [Hunyuan-DiT](https://arxiv.org/abs/2405.08748).
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Hunyuan-DiT generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This code is implemented by Tencent Hunyuan Team. You can find pre-trained checkpoints for Hunyuan-DiT ControlNets on [Tencent Hunyuan](https://huggingface.co/Tencent-Hunyuan).
+
+## Example For Loading HunyuanDiT2DControlNetModel
+
+```py
+from diffusers import HunyuanDiT2DControlNetModel
+import torch
+controlnet = HunyuanDiT2DControlNetModel.from_pretrained("Tencent-Hunyuan/HunyuanDiT-v1.1-ControlNet-Diffusers-Pose", torch_dtype=torch.float16)
+```
+
+## HunyuanDiT2DControlNetModel
+
+[[autodoc]] HunyuanDiT2DControlNetModel

docs/source/en/api/models/latte_transformer3d.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 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.
+-->
+
+## LatteTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [Latte](https://github.com/Vchitect/Latte).
+
+## LatteTransformer3DModel
+
+[[autodoc]] LatteTransformer3DModel

docs/source/en/api/models/lumina_nextdit2d.md

@@ -0,0 +1,20 @@
+<!--Copyright 2024 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.
+-->
+
+# LuminaNextDiT2DModel
+
+A Next Version of Diffusion Transformer model for 2D data from [Lumina-T2X](https://github.com/Alpha-VLLM/Lumina-T2X).
+
+## LuminaNextDiT2DModel
+
+[[autodoc]] LuminaNextDiT2DModel
+

docs/source/en/api/models/transformer2d.md

@@ -38,4 +38,4 @@ It is assumed one of the input classes is the masked latent pixel. The predicted
 
 ## Transformer2DModelOutput
 
-[[autodoc]] models.transformers.transformer_2d.Transformer2DModelOutput
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/pipelines/animatediff.md

@@ -560,6 +560,20 @@ export_to_gif(frames, "animatelcm-motion-lora.gif")
 </table>
 
 
+## Using `from_single_file` with the MotionAdapter
+
+`diffusers>=0.30.0` supports loading the AnimateDiff checkpoints into the `MotionAdapter` in their original format via `from_single_file`
+
+```python
+from diffusers import MotionAdapter
+
+ckpt_path = "https://huggingface.co/Lightricks/LongAnimateDiff/blob/main/lt_long_mm_32_frames.ckpt"
+
+adapter = MotionAdapter.from_single_file(ckpt_path, torch_dtype=torch.float16)
+pipe = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapter=adapter)
+
+```
+
 ## AnimateDiffPipeline
 
 [[autodoc]] AnimateDiffPipeline

docs/source/en/api/pipelines/aura_flow.md

@@ -0,0 +1,29 @@
+<!--Copyright 2024 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.
+-->
+
+# AuraFlow
+
+AuraFlow is inspired by [Stable Diffusion 3](../pipelines/stable_diffusion/stable_diffusion_3.md) and is by far the largest text-to-image generation model that comes with an Apache 2.0 license. This model achieves state-of-the-art results on the [GenEval](https://github.com/djghosh13/geneval) benchmark.
+
+It was developed by the Fal team and more details about it can be found in [this blog post](https://blog.fal.ai/auraflow/).
+
+<Tip>
+
+AuraFlow can be quite expensive to run on consumer hardware devices. However, you can perform a suite of optimizations to run it faster and in a more memory-friendly manner. Check out [this section](https://huggingface.co/blog/sd3#memory-optimizations-for-sd3) for more details. 
+
+</Tip>
+
+## AuraFlowPipeline
+
+[[autodoc]] AuraFlowPipeline
+	- all
+	- __call__

docs/source/en/api/pipelines/controlnet_hunyuandit.md

@@ -0,0 +1,36 @@
+<!--Copyright 2024 The HuggingFace Team and Tencent Hunyuan 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.
+-->
+
+# ControlNet with Hunyuan-DiT
+
+HunyuanDiTControlNetPipeline is an implementation of ControlNet for [Hunyuan-DiT](https://arxiv.org/abs/2405.08748).
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Hunyuan-DiT generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This code is implemented by Tencent Hunyuan Team. You can find pre-trained checkpoints for Hunyuan-DiT ControlNets on [Tencent Hunyuan](https://huggingface.co/Tencent-Hunyuan).
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+## HunyuanDiTControlNetPipeline
+[[autodoc]] HunyuanDiTControlNetPipeline
+	- all
+	- __call__

docs/source/en/api/pipelines/hunyuandit.md

@@ -1,4 +1,4 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+<!--Copyright 2024 The HuggingFace Team and Tencent Hunyuan 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
@@ -34,6 +34,12 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 </Tip>
 
+<Tip>
+
+You can further improve generation quality by passing the generated image from [`HungyuanDiTPipeline`] to the [SDXL refiner](../../using-diffusers/sdxl#base-to-refiner-model) model.
+
+</Tip>
+
 ## Optimization
 
 You can optimize the pipeline's runtime and memory consumption with torch.compile and feed-forward chunking. To learn about other optimization methods, check out the [Speed up inference](../../optimization/fp16) and [Reduce memory usage](../../optimization/memory) guides.

docs/source/en/api/pipelines/kolors.md

@@ -0,0 +1,49 @@
+<!--Copyright 2024 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.
+-->
+
+# Kolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis
+
+![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/kolors_header_collage.png)
+
+Kolors is a large-scale text-to-image generation model based on latent diffusion, developed by [the Kuaishou Kolors team](kwai-kolors@kuaishou.com). Trained on billions of text-image pairs, Kolors exhibits significant advantages over both open-source and closed-source models in visual quality, complex semantic accuracy, and text rendering for both Chinese and English characters. Furthermore, Kolors supports both Chinese and English inputs, demonstrating strong performance in understanding and generating Chinese-specific content. For more details, please refer to this [technical report](https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf).
+
+The abstract from the technical report is:
+
+*We present Kolors, a latent diffusion model for text-to-image synthesis, characterized by its profound understanding of both English and Chinese, as well as an impressive degree of photorealism. There are three key insights contributing to the development of Kolors. Firstly, unlike large language model T5 used in Imagen and Stable Diffusion 3, Kolors is built upon the General Language Model (GLM), which enhances its comprehension capabilities in both English and Chinese. Moreover, we employ a multimodal large language model to recaption the extensive training dataset for fine-grained text understanding. These strategies significantly improve Kolors’ ability to comprehend intricate semantics, particularly those involving multiple entities, and enable its advanced text rendering capabilities. Secondly, we divide the training of Kolors into two phases: the concept learning phase with broad knowledge and the quality improvement phase with specifically curated high-aesthetic data. Furthermore, we investigate the critical role of the noise schedule and introduce a novel schedule to optimize high-resolution image generation. These strategies collectively enhance the visual appeal of the generated high-resolution images. Lastly, we propose a category-balanced benchmark KolorsPrompts, which serves as a guide for the training and evaluation of Kolors. Consequently, even when employing the commonly used U-Net backbone, Kolors has demonstrated remarkable performance in human evaluations, surpassing the existing open-source models and achieving Midjourney-v6 level performance, especially in terms of visual appeal. We will release the code and weights of Kolors at <https://github.com/Kwai-Kolors/Kolors>, and hope that it will benefit future research and applications in the visual generation community.*
+
+## Usage Example
+
+```python
+import torch
+
+from diffusers import DPMSolverMultistepScheduler, KolorsPipeline
+
+pipe = KolorsPipeline.from_pretrained("Kwai-Kolors/Kolors-diffusers", torch_dtype=torch.float16, variant="fp16")
+pipe.to("cuda")
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
+
+image = pipe(
+    prompt='一张瓢虫的照片，微距，变焦，高质量，电影，拿着一个牌子，写着"可图"',
+    negative_prompt="",
+    guidance_scale=6.5,
+    num_inference_steps=25,
+).images[0]
+
+image.save("kolors_sample.png")
+```
+
+## KolorsPipeline
+
+[[autodoc]] KolorsPipeline
+
+- all
+- __call__

docs/source/en/api/pipelines/latte.md

@@ -0,0 +1,75 @@
+<!-- # Copyright 2024 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. -->
+
+# Latte
+
+![latte text-to-video](https://github.com/Vchitect/Latte/blob/52bc0029899babbd6e9250384c83d8ed2670ff7a/visuals/latte.gif?raw=true)
+
+[Latte: Latent Diffusion Transformer for Video Generation](https://arxiv.org/abs/2401.03048) from Monash University, Shanghai AI Lab, Nanjing University, and Nanyang Technological University.
+
+The abstract from the paper is:
+
+*We propose a novel Latent Diffusion Transformer, namely Latte, for video generation. Latte first extracts spatio-temporal tokens from input videos and then adopts a series of Transformer blocks to model video distribution in the latent space. In order to model a substantial number of tokens extracted from videos, four efficient variants are introduced from the perspective of decomposing the spatial and temporal dimensions of input videos. To improve the quality of generated videos, we determine the best practices of Latte through rigorous experimental analysis, including video clip patch embedding, model variants, timestep-class information injection, temporal positional embedding, and learning strategies. Our comprehensive evaluation demonstrates that Latte achieves state-of-the-art performance across four standard video generation datasets, i.e., FaceForensics, SkyTimelapse, UCF101, and Taichi-HD. In addition, we extend Latte to text-to-video generation (T2V) task, where Latte achieves comparable results compared to recent T2V models. We strongly believe that Latte provides valuable insights for future research on incorporating Transformers into diffusion models for video generation.*
+
+**Highlights**: Latte is a latent diffusion transformer proposed as a backbone for modeling different modalities (trained for text-to-video generation here). It achieves state-of-the-art performance across four standard video benchmarks - [FaceForensics](https://arxiv.org/abs/1803.09179), [SkyTimelapse](https://arxiv.org/abs/1709.07592), [UCF101](https://arxiv.org/abs/1212.0402) and [Taichi-HD](https://arxiv.org/abs/2003.00196). To prepare and download the datasets for evaluation, please refer to [this https URL](https://github.com/Vchitect/Latte/blob/main/docs/datasets_evaluation.md).
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+### Inference
+
+Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
+
+First, load the pipeline:
+
+```python
+import torch
+from diffusers import LattePipeline
+
+pipeline = LattePipeline.from_pretrained(
+	"maxin-cn/Latte-1", torch_dtype=torch.float16
+).to("cuda")
+```
+
+Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
+
+```python
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.vae.to(memory_format=torch.channels_last)
+```
+
+Finally, compile the components and run inference:
+
+```python
+pipeline.transformer = torch.compile(pipeline.transformer)
+pipeline.vae.decode = torch.compile(pipeline.vae.decode)
+
+video = pipeline(prompt="A dog wearing sunglasses floating in space, surreal, nebulae in background").frames[0]
+```
+
+The [benchmark](https://gist.github.com/a-r-r-o-w/4e1694ca46374793c0361d740a99ff19) results on an 80GB A100 machine are:
+
+```
+Without torch.compile(): Average inference time: 16.246 seconds.
+With torch.compile(): Average inference time: 14.573 seconds.
+```
+
+## LattePipeline
+
+[[autodoc]] LattePipeline
+  - all
+  - __call__

docs/source/en/api/pipelines/lumina.md

@@ -0,0 +1,88 @@
+<!--Copyright 2024 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.
+-->
+
+# Lumina-T2X
+![concepts](https://github.com/Alpha-VLLM/Lumina-T2X/assets/54879512/9f52eabb-07dc-4881-8257-6d8a5f2a0a5a)
+
+[Lumina-Next : Making Lumina-T2X Stronger and Faster with Next-DiT](https://github.com/Alpha-VLLM/Lumina-T2X/blob/main/assets/lumina-next.pdf) from Alpha-VLLM, OpenGVLab, Shanghai AI Laboratory.
+
+The abstract from the paper is:
+
+*Lumina-T2X is a nascent family of Flow-based Large Diffusion Transformers (Flag-DiT) that establishes a unified framework for transforming noise into various modalities, such as images and videos, conditioned on text instructions. Despite its promising capabilities, Lumina-T2X still encounters challenges including training instability, slow inference, and extrapolation artifacts. In this paper, we present Lumina-Next, an improved version of Lumina-T2X, showcasing stronger generation performance with increased training and inference efficiency. We begin with a comprehensive analysis of the Flag-DiT architecture and identify several suboptimal components, which we address by introducing the Next-DiT architecture with 3D RoPE and sandwich normalizations. To enable better resolution extrapolation, we thoroughly compare different context extrapolation methods applied to text-to-image generation with 3D RoPE, and propose Frequency- and Time-Aware Scaled RoPE tailored for diffusion transformers. Additionally, we introduce a sigmoid time discretization schedule to reduce sampling steps in solving the Flow ODE and the Context Drop method to merge redundant visual tokens for faster network evaluation, effectively boosting the overall sampling speed. Thanks to these improvements, Lumina-Next not only improves the quality and efficiency of basic text-to-image generation but also demonstrates superior resolution extrapolation capabilities and multilingual generation using decoder-based LLMs as the text encoder, all in a zero-shot manner. To further validate Lumina-Next as a versatile generative framework, we instantiate it on diverse tasks including visual recognition, multi-view, audio, music, and point cloud generation, showcasing strong performance across these domains. By releasing all codes and model weights at https://github.com/Alpha-VLLM/Lumina-T2X, we aim to advance the development of next-generation generative AI capable of universal modeling.*
+
+**Highlights**: Lumina-Next is a next-generation Diffusion Transformer that significantly enhances text-to-image generation, multilingual generation, and multitask performance by introducing the Next-DiT architecture, 3D RoPE, and frequency- and time-aware RoPE, among other improvements.
+
+Lumina-Next has the following components:
+* It improves sampling efficiency with fewer and faster Steps.
+* It uses a Next-DiT as a transformer backbone with Sandwichnorm 3D RoPE, and Grouped-Query Attention.
+* It uses a Frequency- and Time-Aware Scaled RoPE.
+
+---
+
+[Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers](https://arxiv.org/abs/2405.05945) from Alpha-VLLM, OpenGVLab, Shanghai AI Laboratory.
+
+The abstract from the paper is:
+
+*Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.*
+
+
+You can find the original codebase at [Alpha-VLLM](https://github.com/Alpha-VLLM/Lumina-T2X) and all the available checkpoints at [Alpha-VLLM Lumina Family](https://huggingface.co/collections/Alpha-VLLM/lumina-family-66423205bedb81171fd0644b).
+
+**Highlights**: Lumina-T2X supports Any Modality, Resolution, and Duration.
+
+Lumina-T2X has the following components:
+* It uses a Flow-based Large Diffusion Transformer as the backbone
+* It supports different any modalities with one backbone and corresponding encoder, decoder.
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+### Inference (Text-to-Image)
+
+Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
+
+First, load the pipeline:
+
+```python
+from diffusers import LuminaText2ImgPipeline
+import torch 
+
+pipeline = LuminaText2ImgPipeline.from_pretrained(
+	"Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16
+).to("cuda")
+```
+
+Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
+
+```python
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.vae.to(memory_format=torch.channels_last)
+```
+
+Finally, compile the components and run inference:
+
+```python
+pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
+pipeline.vae.decode = torch.compile(pipeline.vae.decode, mode="max-autotune", fullgraph=True)
+
+image = pipeline(prompt="Upper body of a young woman in a Victorian-era outfit with brass goggles and leather straps. Background shows an industrial revolution cityscape with smoky skies and tall, metal structures").images[0]
+```
+
+## LuminaText2ImgPipeline
+
+[[autodoc]] LuminaText2ImgPipeline
+	- all
+	- __call__
+	

docs/source/en/api/pipelines/pag.md

@@ -20,6 +20,16 @@ The abstract from the paper is:
 
 *Recent studies have demonstrated that diffusion models are capable of generating high-quality samples, but their quality heavily depends on sampling guidance techniques, such as classifier guidance (CG) and classifier-free guidance (CFG). These techniques are often not applicable in unconditional generation or in various downstream tasks such as image restoration. In this paper, we propose a novel sampling guidance, called Perturbed-Attention Guidance (PAG), which improves diffusion sample quality across both unconditional and conditional settings, achieving this without requiring additional training or the integration of external modules. PAG is designed to progressively enhance the structure of samples throughout the denoising process. It involves generating intermediate samples with degraded structure by substituting selected self-attention maps in diffusion U-Net with an identity matrix, by considering the self-attention mechanisms' ability to capture structural information, and guiding the denoising process away from these degraded samples. In both ADM and Stable Diffusion, PAG surprisingly improves sample quality in conditional and even unconditional scenarios. Moreover, PAG significantly improves the baseline performance in various downstream tasks where existing guidances such as CG or CFG cannot be fully utilized, including ControlNet with empty prompts and image restoration such as inpainting and deblurring.*
 
+## StableDiffusionPAGPipeline
+[[autodoc]] StableDiffusionPAGPipeline
+	- all
+	- __call__
+
+## StableDiffusionControlNetPAGPipeline
+[[autodoc]] StableDiffusionControlNetPAGPipeline
+	- all
+	- __call__
+
 ## StableDiffusionXLPAGPipeline
 [[autodoc]] StableDiffusionXLPAGPipeline
 	- all

docs/source/en/api/pipelines/pixart_sigma.md

@@ -37,6 +37,12 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers)
 
 </Tip>
 
+<Tip>
+
+You can further improve generation quality by passing the generated image from [`PixArtSigmaPipeline`] to the [SDXL refiner](../../using-diffusers/sdxl#base-to-refiner-model) model.
+
+</Tip>
+
 ## Inference with under 8GB GPU VRAM
 
 Run the [`PixArtSigmaPipeline`] with under 8GB GPU VRAM by loading the text encoder in 8-bit precision. Let's walk through a full-fledged example.

docs/source/en/api/schedulers/flow_match_heun_discrete.md

@@ -0,0 +1,18 @@
+<!--Copyright 2024 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.
+-->
+
+# FlowMatchHeunDiscreteScheduler
+
+`FlowMatchHeunDiscreteScheduler` is based on the flow-matching sampling introduced in [EDM](https://arxiv.org/abs/2403.03206).
+
+## FlowMatchHeunDiscreteScheduler
+[[autodoc]] FlowMatchHeunDiscreteScheduler

docs/source/en/training/distributed_inference.md

@@ -52,76 +52,6 @@ To learn more, take a look at the [Distributed Inference with 🤗 Accelerate](h
 
 </Tip>
 
-### Device placement
-
-> [!WARNING]
-> This feature is experimental and its APIs might change in the future.
-
-With Accelerate, you can use the `device_map` to determine how to distribute the models of a pipeline across multiple devices. This is useful in situations where you have more than one GPU.
-
-For example, if you have two 8GB GPUs, then using [`~DiffusionPipeline.enable_model_cpu_offload`] may not work so well because:
-
-* it only works on a single GPU
-* a single model might not fit on a single GPU ([`~DiffusionPipeline.enable_sequential_cpu_offload`] might work but it will be extremely slow and it is also limited to a single GPU)
-
-To make use of both GPUs, you can use the "balanced" device placement strategy which splits the models across all available GPUs.
-
-> [!WARNING]
-> Only the "balanced" strategy is supported at the moment, and we plan to support additional mapping strategies in the future.
-
-```diff
-from diffusers import DiffusionPipeline
-import torch
-
-pipeline = DiffusionPipeline.from_pretrained(
--    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True,
-+    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True, device_map="balanced"
-)
-image = pipeline("a dog").images[0]
-image
-```
-
-You can also pass a dictionary to enforce the maximum GPU memory that can be used on each device:
-
-```diff
-from diffusers import DiffusionPipeline
-import torch
-
-max_memory = {0:"1GB", 1:"1GB"}
-pipeline = DiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5",
-    torch_dtype=torch.float16,
-    use_safetensors=True,
-    device_map="balanced",
-+   max_memory=max_memory
-)
-image = pipeline("a dog").images[0]
-image
-```
-
-If a device is not present in `max_memory`, then it will be completely ignored and will not participate in the device placement.
-
-By default, Diffusers uses the maximum memory of all devices. If the models don't fit on the GPUs, they are offloaded to the CPU. If the CPU doesn't have enough memory, then you might see an error. In that case, you could defer to using [`~DiffusionPipeline.enable_sequential_cpu_offload`] and [`~DiffusionPipeline.enable_model_cpu_offload`].
-
-Call [`~DiffusionPipeline.reset_device_map`] to reset the `device_map` of a pipeline. This is also necessary if you want to use methods like `to()`, [`~DiffusionPipeline.enable_sequential_cpu_offload`], and [`~DiffusionPipeline.enable_model_cpu_offload`] on a pipeline that was device-mapped.
-
-```py
-pipeline.reset_device_map()
-```
-
-Once a pipeline has been device-mapped, you can also access its device map via `hf_device_map`:
-
-```py
-print(pipeline.hf_device_map)
-```
-
-An example device map would look like so:
-
-
-```bash
-{'unet': 1, 'vae': 1, 'safety_checker': 0, 'text_encoder': 0}
-```
-
 ## PyTorch Distributed
 
 PyTorch supports [`DistributedDataParallel`](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html) which enables data parallelism.
@@ -176,3 +106,6 @@ Once you've completed the inference script, use the `--nproc_per_node` argument
 ```bash
 torchrun run_distributed.py --nproc_per_node=2
 ```
+
+> [!TIP]
+> You can use `device_map` within a [`DiffusionPipeline`] to distribute its model-level components on multiple devices. Refer to the [Device placement](../tutorials/inference_with_big_models#device-placement) guide to learn more.

docs/source/en/tutorials/basic_training.md

@@ -340,6 +340,7 @@ Now you can wrap all these components together in a training loop with 🤗 Acce
 ...                 loss = F.mse_loss(noise_pred, noise)
 ...                 accelerator.backward(loss)
 
+...             if (step + 1) % config.gradient_accumulation_steps == 0:
 ...                 accelerator.clip_grad_norm_(model.parameters(), 1.0)
 ...                 optimizer.step()
 ...                 lr_scheduler.step()

docs/source/en/tutorials/inference_with_big_models.md

@@ -0,0 +1,139 @@
+<!--Copyright 2024 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.
+-->
+
+# Working with big models
+
+A modern diffusion model, like [Stable Diffusion XL (SDXL)](../using-diffusers/sdxl), is not just a single model, but a collection of multiple models. SDXL has four different model-level components:
+
+* A variational autoencoder (VAE)
+* Two text encoders
+* A UNet for denoising
+
+Usually, the text encoders and the denoiser are much larger compared to the VAE. 
+
+As models get bigger and better, it’s possible your model is so big that even a single copy won’t fit in memory. But that doesn’t mean it can’t be loaded. If you have more than one GPU, there is more memory available to store your model. In this case, it’s better to split your model checkpoint into several smaller *checkpoint shards*.
+
+When a text encoder checkpoint has multiple shards, like [T5-xxl for SD3](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers/tree/main/text_encoder_3), it is automatically handled by the [Transformers](https://huggingface.co/docs/transformers/index) library as it is a required dependency of Diffusers when using the [`StableDiffusion3Pipeline`]. More specifically, Transformers will automatically handle the loading of multiple shards within the requested model class and get it ready so that inference can be performed.
+
+The denoiser checkpoint can also have multiple shards and supports inference thanks to the [Accelerate](https://huggingface.co/docs/accelerate/index) library. 
+
+> [!TIP]
+> Refer to the [Handling big models for inference](https://huggingface.co/docs/accelerate/main/en/concept_guides/big_model_inference) guide for general guidance when working with big models that are hard to fit into memory.
+
+For example, let's save a sharded checkpoint for the [SDXL UNet](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/unet):
+
+```python
+from diffusers import UNet2DConditionModel
+
+unet = UNet2DConditionModel.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet"
+)
+unet.save_pretrained("sdxl-unet-sharded", max_shard_size="5GB")
+```
+
+The size of the fp32 variant of the SDXL UNet checkpoint is ~10.4GB. Set the `max_shard_size` parameter to 5GB to create 3 shards. After saving, you can load them in [`StableDiffusionXLPipeline`]:
+
+```python
+from diffusers import UNet2DConditionModel, StableDiffusionXLPipeline 
+import torch
+
+unet = UNet2DConditionModel.from_pretrained(
+    "sayakpaul/sdxl-unet-sharded", torch_dtype=torch.float16
+)
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", unet=unet, torch_dtype=torch.float16
+).to("cuda")
+
+image = pipeline("a cute dog running on the grass", num_inference_steps=30).images[0]
+image.save("dog.png")
+```
+
+If placing all the model-level components on the GPU at once is not feasible, use [`~DiffusionPipeline.enable_model_cpu_offload`] to help you: 
+
+```diff
+- pipeline.to("cuda")
++ pipeline.enable_model_cpu_offload()
+```
+
+In general, we recommend sharding when a checkpoint is more than 5GB (in fp32). 
+
+## Device placement
+
+On distributed setups, you can run inference across multiple GPUs with Accelerate.
+
+> [!WARNING]
+> This feature is experimental and its APIs might change in the future.
+
+With Accelerate, you can use the `device_map` to determine how to distribute the models of a pipeline across multiple devices. This is useful in situations where you have more than one GPU.
+
+For example, if you have two 8GB GPUs, then using [`~DiffusionPipeline.enable_model_cpu_offload`] may not work so well because:
+
+* it only works on a single GPU
+* a single model might not fit on a single GPU ([`~DiffusionPipeline.enable_sequential_cpu_offload`] might work but it will be extremely slow and it is also limited to a single GPU)
+
+To make use of both GPUs, you can use the "balanced" device placement strategy which splits the models across all available GPUs.
+
+> [!WARNING]
+> Only the "balanced" strategy is supported at the moment, and we plan to support additional mapping strategies in the future.
+
+```diff
+from diffusers import DiffusionPipeline
+import torch
+
+pipeline = DiffusionPipeline.from_pretrained(
+-    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True,
++    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True, device_map="balanced"
+)
+image = pipeline("a dog").images[0]
+image
+```
+
+You can also pass a dictionary to enforce the maximum GPU memory that can be used on each device:
+
+```diff
+from diffusers import DiffusionPipeline
+import torch
+
+max_memory = {0:"1GB", 1:"1GB"}
+pipeline = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    device_map="balanced",
++   max_memory=max_memory
+)
+image = pipeline("a dog").images[0]
+image
+```
+
+If a device is not present in `max_memory`, then it will be completely ignored and will not participate in the device placement.
+
+By default, Diffusers uses the maximum memory of all devices. If the models don't fit on the GPUs, they are offloaded to the CPU. If the CPU doesn't have enough memory, then you might see an error. In that case, you could defer to using [`~DiffusionPipeline.enable_sequential_cpu_offload`] and [`~DiffusionPipeline.enable_model_cpu_offload`].
+
+Call [`~DiffusionPipeline.reset_device_map`] to reset the `device_map` of a pipeline. This is also necessary if you want to use methods like `to()`, [`~DiffusionPipeline.enable_sequential_cpu_offload`], and [`~DiffusionPipeline.enable_model_cpu_offload`] on a pipeline that was device-mapped.
+
+```py
+pipeline.reset_device_map()
+```
+
+Once a pipeline has been device-mapped, you can also access its device map via `hf_device_map`:
+
+```py
+print(pipeline.hf_device_map)
+```
+
+An example device map would look like so:
+
+
+```bash
+{'unet': 1, 'vae': 1, 'safety_checker': 0, 'text_encoder': 0}
+```

docs/source/en/using-diffusers/other-formats.md

@@ -418,7 +418,7 @@ my_local_checkpoint_path = hf_hub_download(
 
 my_local_config_path = snapshot_download(
     repo_id="segmind/SSD-1B",
-    allowed_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
+    allow_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
 )
 
 pipeline = StableDiffusionXLPipeline.from_single_file(my_local_checkpoint_path, config=my_local_config_path, local_files_only=True)
@@ -438,7 +438,7 @@ my_local_checkpoint_path = hf_hub_download(
 
 my_local_config_path = snapshot_download(
     repo_id="segmind/SSD-1B",
-    allowed_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
+    allow_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
     local_dir="my_local_config"
 )
 
@@ -468,7 +468,7 @@ print("My local checkpoint: ", my_local_checkpoint_path)
 
 my_local_config_path = snapshot_download(
     repo_id="segmind/SSD-1B",
-    allowed_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
+    allow_patterns=["*.json", "**/*.json", "*.txt", "**/*.txt"]
     local_dir_use_symlinks=False,
 )
 print("My local config: ", my_local_config_path)

docs/source/en/using-diffusers/pag.md

@@ -44,6 +44,13 @@ pipeline.enable_model_cpu_offload()
 > [!TIP]
 > The `pag_applied_layers` argument allows you to specify which layers PAG is applied to. Additionally, you can use `set_pag_applied_layers` method to update these layers after the pipeline has been created. Check out the [pag_applied_layers](#pag_applied_layers) section to learn more about applying PAG to other layers.
 
+If you already have a pipeline created and loaded, you can enable PAG on it using the `from_pipe` API with the `enable_pag` flag. Internally, a PAG pipeline is created based on the pipeline and task you specified. In the example below, since we used `AutoPipelineForText2Image` and passed a `StableDiffusionXLPipeline`, a `StableDiffusionXLPAGPipeline` is created accordingly. Note that this does not require additional memory, and you will have both `StableDiffusionXLPipeline` and  `StableDiffusionXLPAGPipeline` loaded and ready to use. You can read more about the `from_pipe` API and how to reuse pipelines in diffuser [here](https://huggingface.co/docs/diffusers/using-diffusers/loading#reuse-a-pipeline).
+
+```py
+pipeline_sdxl = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
+pipeline = AutoPipelineForText2Image.from_pipe(pipeline_sdxl, enable_pag=True)
+```
+
 To generate an image, you will also need to pass a `pag_scale`. When `pag_scale` increases, images gain more semantically coherent structures and exhibit fewer artifacts. However overly large guidance scale can lead to smoother textures and slight saturation in the images, similarly to CFG. `pag_scale=3.0` is used in the official demo and works well in most of the use cases, but feel free to experiment and select the appropriate value according to your needs! PAG is disabled when `pag_scale=0`.
 
 ```py
@@ -74,7 +81,7 @@ for pag_scale in [0.0, 3.0]:
 </hfoption>
 <hfoption id="Image-to-image">
 
-Similary, you can use PAG with image-to-image pipelines.
+You can use PAG with image-to-image pipelines.
 
 ```py
 from diffusers import AutoPipelineForImage2Image
@@ -88,7 +95,32 @@ pipeline = AutoPipelineForImage2Image.from_pretrained(
     torch_dtype=torch.float16
 )
 pipeline.enable_model_cpu_offload()
+```
 
+If you already have a image-to-image pipeline and would like enable PAG on it, you can run this
+
+```py
+pipeline_t2i = AutoPipelineForImage2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
+pipeline = AutoPipelineForImage2Image.from_pipe(pipeline_t2i, enable_pag=True)
+```
+
+It is also very easy to directly switch from a text-to-image pipeline to PAG enabled image-to-image pipeline
+
+```py
+pipeline_pag = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
+pipeline = AutoPipelineForImage2Image.from_pipe(pipeline_t2i, enable_pag=True)
+```
+
+If you have a PAG enabled text-to-image pipeline, you can directly switch to a image-to-image pipeline with PAG still enabled
+
+```py
+pipeline_pag = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", enable_pag=True, torch_dtype=torch.float16)
+pipeline = AutoPipelineForImage2Image.from_pipe(pipeline_t2i)
+```
+
+Now let's generate an image!
+
+```py
 pag_scales =  4.0
 guidance_scales = 7.0
 
@@ -120,7 +152,25 @@ pipeline = AutoPipelineForInpainting.from_pretrained(
     torch_dtype=torch.float16
 )
 pipeline.enable_model_cpu_offload()
+```
 
+You can enable PAG on an exisiting inpainting pipeline like this
+
+```py
+pipeline_inpaint = AutoPipelineForInpaiting.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
+pipeline = AutoPipelineForInpaiting.from_pipe(pipeline_inpaint, enable_pag=True)
+```
+
+This still works when your pipeline has a different task: 
+
+```py
+pipeline_t2i = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)
+pipeline = AutoPipelineForInpaiting.from_pipe(pipeline_t2i, enable_pag=True)
+```
+
+Let's generate an image! 
+
+```py
 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 = load_image(img_url).convert("RGB")
@@ -169,6 +219,12 @@ pipeline = AutoPipelineForText2Image.from_pretrained(
 pipeline.enable_model_cpu_offload()
 ```
 
+<Tip>
+
+If you already have a controlnet pipeline and want to enable PAG, you can use the `from_pipe` API: `AutoPipelineForText2Image.from_pipe(pipeline_controlnet, enable_pag=True)`
+
+</Tip>
+
 You can use the pipeline in the same way you normally use ControlNet pipelines, with the added option to specify a `pag_scale` parameter. Note that PAG works well for unconditional generation. In this example, we will generate an image without a prompt.
 
 ```py

docs/source/en/using-diffusers/sdxl.md

@@ -285,6 +285,12 @@ refiner = DiffusionPipeline.from_pretrained(
 ).to("cuda")
 ```
 
+<Tip warning={true}>
+
+You can use SDXL refiner with a different base model. For example, you can use the [Hunyuan-DiT](../../api/pipelines/hunyuandit) or [PixArt-Sigma](../../api/pipelines/pixart_sigma) pipelines to generate images with better prompt adherence. Once you have generated an image, you can pass it to the SDXL refiner model to enhance final generation quality.
+
+</Tip>
+
 Generate an image from the base model, and set the model output to **latent** space:
 
 ```py

docs/source/en/using-diffusers/shap-e.md

@@ -52,7 +52,7 @@ images = pipe(
 ).images
 ```
 
-Now use the [`~utils.export_to_gif`] function to turn the list of image frames into a gif of the 3D object.
+이제 [`~utils.export_to_gif`] 함수를 사용해 이미지 프레임 리스트를 3D 오브젝트의 gif로 변환합니다.
 
 ```py
 from diffusers.utils import export_to_gif

docs/source/en/using-diffusers/svd.md

@@ -21,6 +21,7 @@ This guide will show you how to use SVD to generate short videos from images.
 Before you begin, make sure you have the following libraries installed:
 
 ```py
+# Colab에서 필요한 라이브러리를 설치하기 위해 주석을 제외하세요
 !pip install -q -U diffusers transformers accelerate
 ```
 

docs/source/ko/_toctree.yml

@@ -1,121 +1,183 @@
 - sections:
   - local: index
-    title: "🧨 Diffusers"
+    title: 🧨 Diffusers
   - local: quicktour
     title: "훑어보기"
   - local: stable_diffusion
     title: Stable Diffusion
   - local: installation
-    title: "설치"
-  title: "시작하기"
+    title: 설치
+  title: 시작하기
 - sections:
   - local: tutorials/tutorial_overview
     title: 개요
   - local: using-diffusers/write_own_pipeline
     title: 모델과 스케줄러 이해하기
-  - local: in_translation
-    title: AutoPipeline
+  - local: in_translation # tutorials/autopipeline
+    title: (번역중) AutoPipeline
   - local: tutorials/basic_training
     title: Diffusion 모델 학습하기
-  title: Tutorials
+  - local: in_translation # tutorials/using_peft_for_inference
+    title: (번역중) 추론을 위한 LoRAs 불러오기
+  - local: in_translation # tutorials/fast_diffusion
+    title: (번역중) Text-to-image diffusion 모델 추론 가속화하기
+  - local: in_translation  # tutorials/inference_with_big_models
+    title: (번역중) 큰 모델로 작업하기
+  title: 튜토리얼
 - sections:
-  - sections:
-    - local: using-diffusers/loading_overview
-      title: 개요
-    - local: using-diffusers/loading
-      title: 파이프라인, 모델, 스케줄러 불러오기
-    - local: using-diffusers/schedulers
-      title: 다른 스케줄러들을 가져오고 비교하기
-    - local: using-diffusers/custom_pipeline_overview
-      title: 커뮤니티 파이프라인 불러오기
-    - local: using-diffusers/using_safetensors
-      title: 세이프텐서 불러오기
-    - local: using-diffusers/other-formats
-      title: 다른 형식의 Stable Diffusion 불러오기
-    - local: in_translation
-      title: Hub에 파일 push하기
-    title: 불러오기 & 허브
-  - sections:
-    - local: using-diffusers/pipeline_overview
-      title: 개요
-    - local: using-diffusers/unconditional_image_generation
-      title: Unconditional 이미지 생성
-    - local: using-diffusers/conditional_image_generation
-      title: Text-to-image 생성
-    - local: using-diffusers/img2img
-      title: Text-guided image-to-image
-    - local: using-diffusers/inpaint
-      title: Text-guided 이미지 인페인팅
-    - local: using-diffusers/depth2img
-      title: Text-guided depth-to-image
-    - local: using-diffusers/textual_inversion_inference
-      title: Textual inversion
-    - local: training/distributed_inference
-      title: 여러 GPU를 사용한 분산 추론
-    - local: in_translation
-      title: Distilled Stable Diffusion 추론
-    - local: using-diffusers/reusing_seeds
-      title: Deterministic 생성으로 이미지 퀄리티 높이기
-    - local: using-diffusers/control_brightness
-      title: 이미지 밝기 조정하기
-    - local: using-diffusers/reproducibility
-      title: 재현 가능한 파이프라인 생성하기
-    - local: using-diffusers/custom_pipeline_examples
-      title: 커뮤니티 파이프라인들
-    - local: using-diffusers/contribute_pipeline
-      title: 커뮤티니 파이프라인에 기여하는 방법
-    - local: using-diffusers/stable_diffusion_jax_how_to
-      title: JAX/Flax에서의 Stable Diffusion
-    - local: using-diffusers/weighted_prompts
-      title: Weighting Prompts
-    title: 추론을 위한 파이프라인
-  - sections:
-    - local: training/overview
-      title: 개요
-    - local: training/create_dataset
-      title: 학습을 위한 데이터셋 생성하기
-    - local: training/adapt_a_model
-      title: 새로운 태스크에 모델 적용하기
+  - local: using-diffusers/loading
+    title: 파이프라인 불러오기
+  - local: using-diffusers/custom_pipeline_overview
+    title: 커뮤니티 파이프라인과 컴포넌트 불러오기
+  - local: using-diffusers/schedulers
+    title: 스케줄러와 모델 불러오기
+  - local: using-diffusers/other-formats
+    title: 모델 파일과 레이아웃
+  - local: using-diffusers/loading_adapters
+    title: 어댑터 불러오기
+  - local: using-diffusers/push_to_hub
+    title: 파일들을 Hub로 푸시하기
+  title: 파이프라인과 어댑터 불러오기
+- sections:
+  - local: using-diffusers/unconditional_image_generation
+    title: Unconditional 이미지 생성
+  - local: using-diffusers/conditional_image_generation
+    title: Text-to-image
+  - local: using-diffusers/img2img
+    title: Image-to-image
+  - local: using-diffusers/inpaint
+    title: 인페인팅
+  - local: in_translation # using-diffusers/text-img2vid
+    title: (번역중) Text 또는 image-to-video
+  - local: using-diffusers/depth2img
+    title: Depth-to-image
+  title: 생성 태스크
+- sections:
+  - local: in_translation # using-diffusers/overview_techniques
+    title: (번역중) 개요
+  - local: training/distributed_inference
+    title: 여러 GPU를 사용한 분산 추론
+  - local: in_translation # using-diffusers/merge_loras
+    title: (번역중) LoRA 병합
+  - local: in_translation # using-diffusers/scheduler_features
+    title: (번역중) 스케줄러 기능
+  - local: in_translation # using-diffusers/callback
+    title: (번역중) 파이프라인 콜백
+  - local: in_translation # using-diffusers/reusing_seeds
+    title: (번역중) 재현 가능한 파이프라인
+  - local: in_translation # using-diffusers/image_quality
+    title: (번역중) 이미지 퀄리티 조절하기
+  - local: using-diffusers/weighted_prompts
+    title: 프롬프트 기술
+  title: 추론 테크닉
+- sections:
+  - local: in_translation # advanced_inference/outpaint
+    title: (번역중) Outpainting
+  title: 추론 심화
+- sections:
+  - local: in_translation # using-diffusers/sdxl
+    title: (번역중) Stable Diffusion XL
+  - local: using-diffusers/sdxl_turbo
+    title: SDXL Turbo
+  - local: using-diffusers/kandinsky
+    title: Kandinsky
+  - local: in_translation # using-diffusers/ip_adapter
+    title: (번역중) IP-Adapter
+  - local: in_translation # using-diffusers/pag
+    title: (번역중) PAG
+  - local: in_translation # using-diffusers/controlnet
+    title: (번역중) ControlNet
+  - local: in_translation # using-diffusers/t2i_adapter
+    title: (번역중) T2I-Adapter
+  - local: in_translation # using-diffusers/inference_with_lcm
+    title: (번역중) Latent Consistency Model
+  - local: using-diffusers/textual_inversion_inference
+    title: Textual inversion
+  - local: using-diffusers/shap-e
+    title: Shap-E
+  - local: using-diffusers/diffedit
+    title: DiffEdit
+  - local: in_translation # using-diffusers/inference_with_tcd_lora
+    title: (번역중) Trajectory Consistency Distillation-LoRA
+  - local: using-diffusers/svd
+    title: Stable Video Diffusion
+  - local: in_translation # using-diffusers/marigold_usage
+    title: (번역중) Marigold 컴퓨터 비전
+  title: 특정 파이프라인 예시
+- sections:
+  - local: training/overview
+    title: 개요
+  - local: training/create_dataset
+    title: 학습을 위한 데이터셋 생성하기
+  - local: training/adapt_a_model
+    title: 새로운 태스크에 모델 적용하기
+  - isExpanded: false
+    sections:
     - local: training/unconditional_training
       title: Unconditional 이미지 생성
+    - local: training/text2image
+      title: Text-to-image
+    - local: in_translation # training/sdxl
+      title: (번역중) Stable Diffusion XL
+    - local: in_translation # training/kandinsky
+      title: (번역중) Kandinsky 2.2
+    - local: in_translation # training/wuerstchen
+      title: (번역중) Wuerstchen
+    - local: training/controlnet
+      title: ControlNet
+    - local: in_translation # training/t2i_adapters
+      title: (번역중) T2I-Adapters
+    - local: training/instructpix2pix
+      title: InstructPix2Pix
+    title: 모델
+  - isExpanded: false
+    sections:
     - local: training/text_inversion
       title: Textual Inversion
     - local: training/dreambooth
       title: DreamBooth
-    - local: training/text2image
-      title: Text-to-image
     - local: training/lora
-      title: Low-Rank Adaptation of Large Language Models (LoRA)
-    - local: training/controlnet
-      title: ControlNet
-    - local: training/instructpix2pix
-      title: InstructPix2Pix 학습
+      title: LoRA
     - local: training/custom_diffusion
       title: Custom Diffusion
-    title: Training
-  title: Diffusers 사용하기
+    - local: in_translation # training/lcm_distill
+      title: (번역중) Latent Consistency Distillation
+    - local: in_translation # training/ddpo
+      title: (번역중) DDPO 강화학습 훈련
+    title: 메서드
+  title: 학습
 - sections:
-  - local: optimization/opt_overview
-    title: 개요
   - local: optimization/fp16
-    title: 메모리와 속도
+    title: 추론 스피드업
+  - local: in_translation # optimization/memory
+    title: (번역중) 메모리 사용량 줄이기
   - local: optimization/torch2.0
-    title: Torch2.0 지원
+    title: PyTorch 2.0
   - local: optimization/xformers
     title: xFormers
-  - local: optimization/onnx
-    title: ONNX
-  - local: optimization/open_vino
-    title: OpenVINO
-  - local: optimization/coreml
-    title: Core ML
-  - local: optimization/mps
-    title: MPS
-  - local: optimization/habana
-    title: Habana Gaudi
   - local: optimization/tome
-    title: Token Merging
-  title: 최적화/특수 하드웨어
+    title: Token merging
+  - local: in_translation # optimization/deepcache
+    title: (번역중) DeepCache
+  - local: in_translation # optimization/tgate
+    title: (번역중) TGATE
+  - sections:
+    - local: using-diffusers/stable_diffusion_jax_how_to
+      title: JAX/Flax
+    - local: optimization/onnx
+      title: ONNX
+    - local: optimization/open_vino
+      title: OpenVINO
+    - local: optimization/coreml
+      title: Core ML
+    title: 최적화된 모델 형식
+  - sections:
+    - local: optimization/mps
+      title: Metal Performance Shaders (MPS)
+    - local: optimization/habana
+      title: Habana Gaudi
+    title: 최적화된 하드웨어
+  title: 추론 가속화와 메모리 줄이기
 - sections:
   - local: conceptual/philosophy
     title: 철학

docs/source/ko/conceptual/ethical_guidelines.md

@@ -10,26 +10,27 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# 🧨 Diffusers의 윤리 지침
+# 🧨 Diffusers의 윤리 지침 [[-diffusers-ethical-guidelines]]
 
-## 서문
+## 서문 [[preamble]]
 
 [Diffusers](https://huggingface.co/docs/diffusers/index)는 사전 훈련된 diffusion 모델을 제공하며 추론 및 훈련을 위한 모듈식 툴박스로 사용됩니다.
 
 이 기술의 실제 적용과 사회에 미칠 수 있는 부정적인 영향을 고려하여 Diffusers 라이브러리의 개발, 사용자 기여 및 사용에 윤리 지침을 제공하는 것이 중요하다고 생각합니다.
 
-이 기술을 사용하는 데 연관된 위험은 아직 조사 중이지만, 몇 가지 예를 들면: 예술가들에 대한 저작권 문제; 딥 페이크의 악용; 부적절한 맥락에서의 성적 콘텐츠 생성; 동의 없는 impersonation; 사회적인 편견으로 인해 억압되는 그룹들에 대한 해로운 영향입니다.
+이이 기술을 사용함에 따른 위험은 여전히 검토 중이지만, 몇 가지 예를 들면: 예술가들에 대한 저작권 문제; 딥 페이크의 악용; 부적절한 맥락에서의 성적 콘텐츠 생성; 동의 없는 사칭; 소수자 집단의 억압을 영속화하는 유해한 사회적 편견 등이 있습니다.
+
 우리는 위험을 지속적으로 추적하고 커뮤니티의 응답과 소중한 피드백에 따라 다음 지침을 조정할 것입니다.
 
 
-## 범위
+## 범위 [[scope]]
 
 Diffusers 커뮤니티는 프로젝트의 개발에 다음과 같은 윤리 지침을 적용하며, 특히 윤리적 문제와 관련된 민감한 주제에 대한 커뮤니티의 기여를 조정하는 데 도움을 줄 것입니다.
 
 
-## 윤리 지침
+## 윤리 지침 [[ethical-guidelines]]
 
-다음 윤리 지침은 일반적으로 적용되지만, 기술적 선택을 할 때 윤리적으로 민감한 문제를 다룰 때 주로 적용할 것입니다. 또한, 해당 기술의 최신 동향과 관련된 신규 위험에 따라 시간이 지남에 따라 이러한 윤리 원칙을 조정할 것을 약속합니다.
+다음 윤리 지침은 일반적으로 적용되지만, 민감한 윤리적 문제와 관련하여 기술적 선택을 할 때 이를 우선적으로 적용할 것입니다. 나아가, 해당 기술의 최신 동향과 관련된 새로운 위험이 발생함에 따라 이러한 윤리 원칙을 조정할 것을 약속드립니다.
 
 - **투명성**: 우리는 PR을 관리하고, 사용자에게 우리의 선택을 설명하며, 기술적 의사결정을 내릴 때 투명성을 유지할 것을 약속합니다.
 
@@ -44,7 +45,7 @@ Diffusers 커뮤니티는 프로젝트의 개발에 다음과 같은 윤리 지
 - **책임**: 우리는 커뮤니티와 팀워크를 통해, 이 기술의 잠재적인 위험과 위험을 예측하고 완화하는 데 대한 공동 책임을 가지고 있습니다.
 
 
-## 구현 사례: 안전 기능과 메커니즘
+## 구현 사례: 안전 기능과 메커니즘 [[examples-of-implementations-safety-features-and-mechanisms]]
 
 팀은 diffusion 기술과 관련된 잠재적인 윤리 및 사회적 위험에 대처하기 위한 기술적 및 비기술적 도구를 제공하고자 하고 있습니다. 또한, 커뮤니티의 참여는 이러한 기능의 구현하고 우리와 함께 인식을 높이는 데 매우 중요합니다.
 

docs/source/ko/conceptual/philosophy.md

@@ -10,30 +10,30 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# 철학
+# 철학 [[philosophy]] 
 
 🧨 Diffusers는 다양한 모달리티에서 **최신의** 사전 훈련된 diffusion 모델을 제공합니다.
 그 목적은 추론과 훈련을 위한 **모듈식 툴박스**로 사용되는 것입니다.
 
-우리는 오랜 시간에 견딜 수 있는 라이브러리를 구축하는 것을 목표로 하고, 따라서 API 설계를 매우 중요시합니다.
+저희는 시간이 지나도 변치 않는 라이브러리를 구축하는 것을 목표로 하기에 API 설계를 매우 중요하게 생각합니다.
 
-간단히 말해서, Diffusers는 PyTorch의 자연스러운 확장이 되도록 구축되었습니다. 따라서 대부분의 설계 선택은 [PyTorch의 설계 원칙](https://pytorch.org/docs/stable/community/design.html#pytorch-design-philosophy)에 기반합니다. 이제 가장 중요한 것들을 살펴보겠습니다:
+간단히 말해서, Diffusers는 PyTorch를 자연스럽게 확장할 수 있도록 만들어졌습니다. 따라서 대부분의 설계 선택은 [PyTorch의 설계 원칙](https://pytorch.org/docs/stable/community/design.html#pytorch-design-philosophy)에 기반합니다. 이제 가장 중요한 것들을 살펴보겠습니다:
 
-## 성능보다는 사용성을
+## 성능보다는 사용성을 [[usability-over-performance]]
 
-- Diffusers는 많은 내장 성능 향상 기능을 갖고 있지만 (자세한 내용은 [메모리와 속도](https://huggingface.co/docs/diffusers/optimization/fp16) 참조), 모델은 항상 가장 높은 정밀도와 최소한의 최적화로 로드됩니다. 따라서 기본적인 diffusion 파이프라인은 따로 정의하지 않는다면 CPU에서 float32 정밀도로 인스턴스화됩니다. 이는 다양한 플랫폼과 가속기에서의 사용성을 보장하며, 라이브러리를 실행하기 위해 복잡한 설치가 필요하지 않음을 의미합니다.
+- Diffusers는 다양한 성능 향상 기능이 내장되어 있지만 (자세한 내용은 [메모리와 속도](https://huggingface.co/docs/diffusers/optimization/fp16) 참조), 모델은 항상 가장 높은 정밀도와 최소한의 최적화로 로드됩니다. 따라서 사용자가 별도로 정의하지 않는 한 기본적으로 diffusion 파이프라인은 항상 float32 정밀도로 CPU에 인스턴스화됩니다. 이는 다양한 플랫폼과 가속기에서의 사용성을 보장하며, 라이브러리를 실행하기 위해 복잡한 설치가 필요하지 않다는 것을 의미합니다.
 - Diffusers는 **가벼운** 패키지를 지향하기 때문에 필수 종속성은 거의 없지만 성능을 향상시킬 수 있는 많은 선택적 종속성이 있습니다 (`accelerate`, `safetensors`, `onnx` 등). 저희는 라이브러리를 가능한 한 가볍게 유지하여 다른 패키지에 대한 종속성 걱정이 없도록 노력하고 있습니다.
 - Diffusers는 간결하고 이해하기 쉬운 코드를 선호합니다. 이는 람다 함수나 고급 PyTorch 연산자와 같은 압축된 코드 구문을 자주 사용하지 않는 것을 의미합니다.
 
-## 쉬움보다는 간단함을
+## 쉬움보다는 간단함을 [[simple-over-easy]]
 
 PyTorch에서는 **명시적인 것이 암시적인 것보다 낫다**와 **단순한 것이 복잡한 것보다 낫다**라고 말합니다. 이 설계 철학은 라이브러리의 여러 부분에 반영되어 있습니다:
-- [`DiffusionPipeline.to`](https://huggingface.co/docs/diffusers/main/en/api/diffusion_pipeline#diffusers.DiffusionPipeline.to)와 같은 메서드를 사용하여 사용자가 장치 관리를 할 수 있도록 PyTorch의 API를 따릅니다.
+- [`DiffusionPipeline.to`](https://huggingface.co/docs/diffusers/main/en/api/diffusion_pipeline#diffusers.DiffusionPipeline.to)와 같은 메소드를 사용하여 사용자가 장치 관리를 할 수 있도록 PyTorch의 API를 따릅니다.
 - 잘못된 입력을 조용히 수정하는 대신 간결한 오류 메시지를 발생시키는 것이 우선입니다. Diffusers는 라이브러리를 가능한 한 쉽게 사용할 수 있도록 하는 것보다 사용자를 가르치는 것을 목표로 합니다.
 - 복잡한 모델과 스케줄러 로직이 내부에서 마법처럼 처리하는 대신 노출됩니다. 스케줄러/샘플러는 서로에게 최소한의 종속성을 가지고 분리되어 있습니다. 이로써 사용자는 언롤된 노이즈 제거 루프를 작성해야 합니다. 그러나 이 분리는 디버깅을 더 쉽게하고 노이즈 제거 과정을 조정하거나 diffusers 모델이나 스케줄러를 교체하는 데 사용자에게 더 많은 제어권을 제공합니다.
 - diffusers 파이프라인의 따로 훈련된 구성 요소인 text encoder, unet 및 variational autoencoder는 각각 자체 모델 클래스를 갖습니다. 이로써 사용자는 서로 다른 모델의 구성 요소 간의 상호 작용을 처리해야 하며, 직렬화 형식은 모델 구성 요소를 다른 파일로 분리합니다. 그러나 이는 디버깅과 커스터마이징을 더 쉽게합니다. DreamBooth나 Textual Inversion 훈련은 Diffusers의 'diffusion 파이프라인의 단일 구성 요소들을 분리할 수 있는 능력' 덕분에 매우 간단합니다.
 
-## 추상화보다는 수정 가능하고 기여하기 쉬움을
+## 추상화보다는 수정 가능하고 기여하기 쉬움을 [[tweakable-contributor-friendly-over-abstraction]]
 
 라이브러리의 대부분에 대해 Diffusers는 [Transformers 라이브러리](https://github.com/huggingface/transformers)의 중요한 설계 원칙을 채택합니다, 바로 성급한 추상화보다는 copy-pasted 코드를 선호한다는 것입니다. 이 설계 원칙은 [Don't repeat yourself (DRY)](https://en.wikipedia.org/wiki/Don%27t_repeat_yourself)와 같은 인기 있는 설계 원칙과는 대조적으로 매우 의견이 분분한데요.
 간단히 말해서, Transformers가 모델링 파일에 대해 수행하는 것처럼, Diffusers는 매우 낮은 수준의 추상화와 매우 독립적인 코드를 유지하는 것을 선호합니다. 함수, 긴 코드 블록, 심지어 클래스도 여러 파일에 복사할 수 있으며, 이는 처음에는 라이브러리를 유지할 수 없게 만드는 나쁜, 서투른 설계 선택으로 보일 수 있습니다. 하지만 이러한 설계는 매우 성공적이며, 커뮤니티 기반의 오픈 소스 기계 학습 라이브러리에 매우 적합합니다. 그 이유는 다음과 같습니다:
@@ -48,11 +48,11 @@ Diffusers에서는 이러한 철학을 파이프라인과 스케줄러에 모두
 좋아요, 이제 🧨 Diffusers가 설계된 방식을 대략적으로 이해했을 것입니다 🤗.
 우리는 이러한 설계 원칙을 일관되게 라이브러리 전체에 적용하려고 노력하고 있습니다. 그럼에도 불구하고 철학에 대한 일부 예외 사항이나 불행한 설계 선택이 있을 수 있습니다. 디자인에 대한 피드백이 있다면 [GitHub에서 직접](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feedback.md&title=) 알려주시면 감사하겠습니다.
 
-## 디자인 철학 자세히 알아보기
+## 디자인 철학 자세히 알아보기 [[design-philosophy-in-details]]
 
 이제 디자인 철학의 세부 사항을 좀 더 자세히 살펴보겠습니다. Diffusers는 주로 세 가지 주요 클래스로 구성됩니다: [파이프라인](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines), [모델](https://github.com/huggingface/diffusers/tree/main/src/diffusers/models), 그리고 [스케줄러](https://github.com/huggingface/diffusers/tree/main/src/diffusers/schedulers). 각 클래스에 대한 더 자세한 설계 결정 사항을 살펴보겠습니다.
 
-### 파이프라인
+### 파이프라인 [[pipelines]]
 
 파이프라인은 사용하기 쉽도록 설계되었으며 (따라서 [*쉬움보다는 간단함을*](#쉬움보다는-간단함을)을 100% 따르지는 않음), feature-complete하지 않으며, 추론을 위한 [모델](#모델)과 [스케줄러](#스케줄러)를 사용하는 방법의 예시로 간주될 수 있습니다.
 
@@ -65,11 +65,11 @@ Diffusers에서는 이러한 철학을 파이프라인과 스케줄러에 모두
 - 파이프라인은 매우 가독성이 좋고, 이해하기 쉽고, 쉽게 조정할 수 있도록 설계되어야 합니다.
 - 파이프라인은 서로 상호작용하고, 상위 수준 API에 쉽게 통합할 수 있도록 설계되어야 합니다.
 - 파이프라인은 사용자 인터페이스가 feature-complete하지 않게 하는 것을 목표로 합니다. future-complete한 사용자 인터페이스를 원한다면 [InvokeAI](https://github.com/invoke-ai/InvokeAI), [Diffuzers](https://github.com/abhishekkrthakur/diffuzers), [lama-cleaner](https://github.com/Sanster/lama-cleaner)를 참조해야 합니다.
-- 모든 파이프라인은 오로지 `__call__` 메서드를 통해 실행할 수 있어야 합니다. `__call__` 인자의 이름은 모든 파이프라인에서 공유되어야 합니다.
+- 모든 파이프라인은 오로지 `__call__` 메소드를 통해 실행할 수 있어야 합니다. `__call__` 인자의 이름은 모든 파이프라인에서 공유되어야 합니다.
 - 파이프라인은 해결하고자 하는 작업의 이름으로 지정되어야 합니다.
 - 대부분의 경우에 새로운 diffusion 파이프라인은 새로운 파이프라인 폴더/파일에 구현되어야 합니다.
 
-### 모델
+### 모델 [[models]]
 
 모델은 [PyTorch의 Module 클래스](https://pytorch.org/docs/stable/generated/torch.nn.Module.html)의 자연스러운 확장이 되도록, 구성 가능한 툴박스로 설계되었습니다. 그리고 모델은 **단일 파일 정책**을 일부만 따릅니다.
 
@@ -85,7 +85,7 @@ Diffusers에서는 이러한 철학을 파이프라인과 스케줄러에 모두
 - 모델은 미래의 변경 사항을 쉽게 확장할 수 있도록 설계되어야 합니다. 이는 공개 함수 인수들과 구성 인수들을 제한하고,미래의 변경 사항을 "예상"하는 것을 통해 달성할 수 있습니다. 예를 들어, 불리언 `is_..._type` 인수보다는 새로운 미래 유형에 쉽게 확장할 수 있는 문자열 "...type" 인수를 추가하는 것이 일반적으로 더 좋습니다. 새로운 모델 체크포인트가 작동하도록 하기 위해 기존 아키텍처에 최소한의 변경만을 가해야 합니다.
 - 모델 디자인은 코드의 가독성과 간결성을 유지하는 것과 많은 모델 체크포인트를 지원하는 것 사이의 어려운 균형 조절입니다. 모델링 코드의 대부분은 새로운 모델 체크포인트를 위해 클래스를 수정하는 것이 좋지만, [UNet 블록](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py) 및 [Attention 프로세서](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)와 같이 코드를 장기적으로 간결하고 읽기 쉽게 유지하기 위해 새로운 클래스를 추가하는 예외도 있습니다.
 
-### 스케줄러
+### 스케줄러 [[schedulers]]
 
 스케줄러는 추론을 위한 노이즈 제거 과정을 안내하고 훈련을 위한 노이즈 스케줄을 정의하는 역할을 합니다. 스케줄러는 개별 클래스로 설계되어 있으며, 로드 가능한 구성 파일과 **단일 파일 정책**을 엄격히 따릅니다.
 
@@ -95,7 +95,7 @@ Diffusers에서는 이러한 철학을 파이프라인과 스케줄러에 모두
 - 하나의 스케줄러 Python 파일은 하나의 스케줄러 알고리즘(논문에서 정의된 것과 같은)에 해당합니다.
 - 스케줄러가 유사한 기능을 공유하는 경우, `# Copied from` 메커니즘을 사용할 수 있습니다.
 - 모든 스케줄러는 `SchedulerMixin`과 `ConfigMixin`을 상속합니다.
-- [`ConfigMixin.from_config`](https://huggingface.co/docs/diffusers/main/en/api/configuration#diffusers.ConfigMixin.from_config) 메서드를 사용하여 스케줄러를 쉽게 교체할 수 있습니다. 자세한 내용은 [여기](../using-diffusers/schedulers.md)에서 설명합니다.
+- [`ConfigMixin.from_config`](https://huggingface.co/docs/diffusers/main/en/api/configuration#diffusers.ConfigMixin.from_config) 메소드를 사용하여 스케줄러를 쉽게 교체할 수 있습니다. 자세한 내용은 [여기](../using-diffusers/schedulers.md)에서 설명합니다.
 - 모든 스케줄러는 `set_num_inference_steps`와 `step` 함수를 가져야 합니다. `set_num_inference_steps(...)`는 각 노이즈 제거 과정(즉, `step(...)`이 호출되기 전) 이전에 호출되어야 합니다.
 - 각 스케줄러는 모델이 호출될 타임스텝의 배열인 `timesteps` 속성을 통해 루프를 돌 수 있는 타임스텝을 노출합니다.
 - `step(...)` 함수는 예측된 모델 출력과 "현재" 샘플(x_t)을 입력으로 받고, "이전" 약간 더 노이즈가 제거된 샘플(x_t-1)을 반환합니다.

docs/source/ko/index.md

@@ -46,52 +46,4 @@ specific language governing permissions and limitations under the License.
       <p class="text-gray-700">🤗 Diffusers 클래스 및 메서드의 작동 방식에 대한 기술 설명.</p>
     </a>
   </div>
-</div>
-
-## Supported pipelines
-
-| Pipeline | Paper/Repository | Tasks |
-|---|---|:---:|
-| [alt_diffusion](./api/pipelines/alt_diffusion) | [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) | Image-to-Image Text-Guided Generation |
-| [audio_diffusion](./api/pipelines/audio_diffusion) | [Audio Diffusion](https://github.com/teticio/audio-diffusion.git) | Unconditional Audio Generation |
-| [controlnet](./api/pipelines/stable_diffusion/controlnet) | [Adding Conditional Control to Text-to-Image Diffusion Models](https://arxiv.org/abs/2302.05543) | Image-to-Image Text-Guided Generation |
-| [cycle_diffusion](./api/pipelines/cycle_diffusion) | [Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance](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 |
-| [if](./if) | [**IF**](./api/pipelines/if) | Image Generation |
-| [if_img2img](./if) | [**IF**](./api/pipelines/if) | Image-to-Image Generation |
-| [if_inpainting](./if) | [**IF**](./api/pipelines/if) | Image-to-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 |
-| [paint_by_example](./api/pipelines/paint_by_example) | [Paint by Example: Exemplar-based Image Editing with Diffusion Models](https://arxiv.org/abs/2211.13227) | Image-Guided Image Inpainting |
-| [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 |
-| [semantic_stable_diffusion](./api/pipelines/semantic_stable_diffusion) | [Semantic Guidance](https://arxiv.org/abs/2301.12247) | Text-Guided Generation |
-| [stable_diffusion_text2img](./api/pipelines/stable_diffusion/text2img) | [Stable Diffusion](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation |
-| [stable_diffusion_img2img](./api/pipelines/stable_diffusion/img2img) | [Stable Diffusion](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation |
-| [stable_diffusion_inpaint](./api/pipelines/stable_diffusion/inpaint) | [Stable Diffusion](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting |
-| [stable_diffusion_panorama](./api/pipelines/stable_diffusion/panorama) | [MultiDiffusion](https://multidiffusion.github.io/) | Text-to-Panorama Generation |
-| [stable_diffusion_pix2pix](./api/pipelines/stable_diffusion/pix2pix) | [InstructPix2Pix: Learning to Follow Image Editing Instructions](https://arxiv.org/abs/2211.09800)  | Text-Guided Image Editing|
-| [stable_diffusion_pix2pix_zero](./api/pipelines/stable_diffusion/pix2pix_zero) | [Zero-shot Image-to-Image Translation](https://pix2pixzero.github.io/) | Text-Guided Image Editing |
-| [stable_diffusion_attend_and_excite](./api/pipelines/stable_diffusion/attend_and_excite) | [Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models](https://arxiv.org/abs/2301.13826) | Text-to-Image Generation |
-| [stable_diffusion_self_attention_guidance](./api/pipelines/stable_diffusion/self_attention_guidance) | [Improving Sample Quality of Diffusion Models Using Self-Attention Guidance](https://arxiv.org/abs/2210.00939) | Text-to-Image Generation Unconditional Image Generation |
-| [stable_diffusion_image_variation](./stable_diffusion/image_variation) | [Stable Diffusion Image Variations](https://github.com/LambdaLabsML/lambda-diffusers#stable-diffusion-image-variations) | Image-to-Image Generation |
-| [stable_diffusion_latent_upscale](./stable_diffusion/latent_upscale) | [Stable Diffusion Latent Upscaler](https://twitter.com/StabilityAI/status/1590531958815064065) | Text-Guided Super Resolution Image-to-Image |
-| [stable_diffusion_model_editing](./api/pipelines/stable_diffusion/model_editing) | [Editing Implicit Assumptions in Text-to-Image Diffusion Models](https://time-diffusion.github.io/) | Text-to-Image Model Editing |
-| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [Stable Diffusion 2](https://stability.ai/blog/stable-diffusion-v2-release) | Text-to-Image Generation |
-| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [Stable Diffusion 2](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Image Inpainting |
-| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [Depth-Conditional Stable Diffusion](https://github.com/Stability-AI/stablediffusion#depth-conditional-stable-diffusion) | Depth-to-Image Generation |
-| [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [Stable Diffusion 2](https://stability.ai/blog/stable-diffusion-v2-release) | Text-Guided Super Resolution Image-to-Image |
-| [stable_diffusion_safe](./api/pipelines/stable_diffusion_safe) | [Safe Stable Diffusion](https://arxiv.org/abs/2211.05105) | Text-Guided Generation |
-| [stable_unclip](./stable_unclip) | Stable unCLIP | Text-to-Image Generation |
-| [stable_unclip](./stable_unclip) | Stable unCLIP | Image-to-Image Text-Guided Generation |
-| [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 |
-| [text_to_video_sd](./api/pipelines/text_to_video) | [Modelscope's Text-to-video-synthesis Model in Open Domain](https://modelscope.cn/models/damo/text-to-video-synthesis/summary) | Text-to-Video Generation |
-| [unclip](./api/pipelines/unclip) | [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://arxiv.org/abs/2204.06125)(implementation by [kakaobrain](https://github.com/kakaobrain/karlo)) | Text-to-Image Generation |
-| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation |
-| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Image Variations Generation |
-| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Dual Image and Text Guided 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 |
+</div>

docs/source/ko/optimization/opt_overview.md

@@ -1,17 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# 개요
-
-노이즈가 많은 출력에서 적은 출력으로 만드는 과정으로 고품질 생성 모델의 출력을 만드는 각각의 반복되는 스텝은 많은 계산이 필요합니다. 🧨 Diffuser의 목표 중 하나는 모든 사람이 이 기술을 널리 이용할 수 있도록 하는 것이며, 여기에는 소비자 및 특수 하드웨어에서 빠른 추론을 가능하게 하는 것을 포함합니다.
-
-이 섹션에서는 추론 속도를 최적화하고 메모리 소비를 줄이기 위한 반정밀(half-precision) 가중치 및 sliced attention과 같은 팁과 요령을 다룹니다. 또한 [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) 또는 [ONNX Runtime](https://onnxruntime.ai/docs/)을 사용하여 PyTorch 코드의 속도를 높이고, [xFormers](https://facebookresearch.github.io/xformers/)를 사용하여 memory-efficient attention을 활성화하는 방법을 배울 수 있습니다. Apple Silicon, Intel 또는 Habana 프로세서와 같은 특정 하드웨어에서 추론을 실행하기 위한 가이드도 있습니다.

docs/source/ko/quicktour.md

@@ -15,7 +15,7 @@ specific language governing permissions and limitations under the License.
 
 Diffusion 모델은 이미지나 오디오와 같은 관심 샘플들을 생성하기 위해 랜덤 가우시안 노이즈를 단계별로 제거하도록 학습됩니다. 이로 인해 생성 AI에 대한 관심이 매우 높아졌으며, 인터넷에서 diffusion 생성 이미지의 예를 본 적이 있을 것입니다. 🧨 Diffusers는 누구나 diffusion 모델들을 널리 이용할 수 있도록 하기 위한 라이브러리입니다.
 
-개발자든 일반 사용자든 이 훑어보기를 통해 🧨 diffusers를 소개하고 빠르게 생성할 수 있도록 도와드립니다! 알아야 할 라이브러리의 주요 구성 요소는 크게 세 가지입니다:
+개발자든 일반 사용자든 이 훑어보기를 통해 🧨 Diffusers를 소개하고 빠르게 생성할 수 있도록 도와드립니다! 알아야 할 라이브러리의 주요 구성 요소는 크게 세 가지입니다:
 
 * [`DiffusionPipeline`]은 추론을 위해 사전 학습된 diffusion 모델에서 샘플을 빠르게 생성하도록 설계된 높은 수준의 엔드투엔드 클래스입니다.
 * Diffusion 시스템 생성을 위한 빌딩 블록으로 사용할 수 있는 널리 사용되는 사전 학습된 [model](./api/models) 아키텍처 및 모듈.

docs/source/ko/using-diffusers/contribute_pipeline.md

@@ -1,182 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# 커뮤니티 파이프라인에 기여하는 방법
-
-<Tip>
-
-💡 모든 사람이 속도 저하 없이 쉽게 작업을 공유할 수 있도록 커뮤니티 파이프라인을 추가하는 이유에 대한 자세한 내용은 GitHub 이슈 [#841](https://github.com/huggingface/diffusers/issues/841)를 참조하세요.
-
-</Tip>
-
-커뮤니티 파이프라인을 사용하면 [`DiffusionPipeline`] 위에 원하는 추가 기능을 추가할 수 있습니다. `DiffusionPipeline` 위에 구축할 때의 가장 큰 장점은 누구나 인수를 하나만 추가하면 파이프라인을 로드하고 사용할 수 있어 커뮤니티가 매우 쉽게 접근할 수 있다는 것입니다.
-
-이번 가이드에서는 커뮤니티 파이프라인을 생성하는 방법과 작동 원리를 설명합니다.
-간단하게 설명하기 위해 `UNet`이 단일 forward pass를 수행하고 스케줄러를 한 번 호출하는 "one-step" 파이프라인을 만들겠습니다.
-
-## 파이프라인 초기화
-
-커뮤니티 파이프라인을 위한 `one_step_unet.py` 파일을 생성하는 것으로 시작합니다. 이 파일에서, Hub에서 모델 가중치와 스케줄러 구성을 로드할 수 있도록 [`DiffusionPipeline`]을 상속하는 파이프라인 클래스를 생성합니다. one-step 파이프라인에는 `UNet`과 스케줄러가 필요하므로 이를 `__init__` 함수에 인수로 추가해야합니다:
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-
-class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
-    def __init__(self, unet, scheduler):
-        super().__init__()
-```
-
-파이프라인과 그 구성요소(`unet` and `scheduler`)를 [`~DiffusionPipeline.save_pretrained`]으로 저장할 수 있도록 하려면 `register_modules` 함수에 추가하세요:
-
-```diff
-  from diffusers import DiffusionPipeline
-  import torch
-
-  class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
-      def __init__(self, unet, scheduler):
-          super().__init__()
-
-+         self.register_modules(unet=unet, scheduler=scheduler)
-```
-
-이제 '초기화' 단계가 완료되었으니 forward pass로 이동할 수 있습니다! 🔥
-
-## Forward pass 정의
-
-Forward pass 에서는(`__call__`로 정의하는 것이 좋습니다) 원하는 기능을 추가할 수 있는 완전한 창작 자유가 있습니다. 우리의 놀라운 one-step 파이프라인의 경우, 임의의 이미지를 생성하고 `timestep=1`을 설정하여 `unet`과 `scheduler`를 한 번만 호출합니다:
-
-```diff
-  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.config.in_channels, self.unet.config.sample_size, self.unet.config.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
-```
-
-끝났습니다! 🚀 이제 이 파이프라인에 `unet`과 `scheduler`를 전달하여 실행할 수 있습니다:
-
-```python
-from diffusers import DDPMScheduler, UNet2DModel
-
-scheduler = DDPMScheduler()
-unet = UNet2DModel()
-
-pipeline = UnetSchedulerOneForwardPipeline(unet=unet, scheduler=scheduler)
-
-output = pipeline()
-```
-
-하지만 파이프라인 구조가 동일한 경우 기존 가중치를 파이프라인에 로드할 수 있다는 장점이 있습니다. 예를 들어 one-step 파이프라인에 [`google/ddpm-cifar10-32`](https://huggingface.co/google/ddpm-cifar10-32) 가중치를 로드할 수 있습니다:
-
-```python
-pipeline = UnetSchedulerOneForwardPipeline.from_pretrained("google/ddpm-cifar10-32")
-
-output = pipeline()
-```
-
-## 파이프라인 공유
-
-🧨Diffusers [리포지토리](https://github.com/huggingface/diffusers)에서 Pull Request를 열어 [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) 하위 폴더에 `one_step_unet.py`의 멋진 파이프라인을 추가하세요.
-
-병합이 되면, `diffusers >= 0.4.0`이 설치된 사용자라면 누구나 `custom_pipeline` 인수에 지정하여 이 파이프라인을 마술처럼 🪄 사용할 수 있습니다:
-
-```python
-from diffusers import DiffusionPipeline
-
-pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
-pipe()
-```
-
-커뮤니티 파이프라인을 공유하는 또 다른 방법은 Hub 에서 선호하는 [모델 리포지토리](https://huggingface.co/docs/hub/models-uploading)에 직접  `one_step_unet.py` 파일을 업로드하는 것입니다. `one_step_unet.py` 파일을 지정하는 대신 모델 저장소 id를 `custom_pipeline` 인수에 전달하세요:
-
-```python
-from diffusers import DiffusionPipeline
-
-pipeline = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="stevhliu/one_step_unet")
-```
-
-다음 표에서 두 가지 공유 워크플로우를 비교하여 자신에게 가장 적합한 옵션을 결정하는 데 도움이 되는 정보를 확인하세요:
-
-|                | GitHub 커뮤니티 파이프라인                                                                                        | HF Hub 커뮤니티 파이프라인                                                                 |
-|----------------|------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------|
-| 사용법          | 동일                                                                                                             | 동일                                                                                      |
-| 리뷰 과정 | 병합하기 전에 GitHub에서 Pull Request를 열고 Diffusers 팀의 검토 과정을 거칩니다. 속도가 느릴 수 있습니다. | 검토 없이 Hub 저장소에 바로 업로드합니다. 가장 빠른 워크플로우 입니다. |
-| 가시성     | 공식 Diffusers 저장소 및 문서에 포함되어 있습니다.                                                  | HF 허브 프로필에 포함되며 가시성을 확보하기 위해 자신의 사용량/프로모션에 의존합니다. |
-
-<Tip>
-
-💡 커뮤니티 파이프라인 파일에 원하는 패키지를 사용할 수 있습니다. 사용자가 패키지를 설치하기만 하면 모든 것이 정상적으로 작동합니다. 파이프라인이 자동으로 감지되므로 `DiffusionPipeline`에서 상속하는 파이프라인 클래스가 하나만 있는지 확인하세요.
-
-</Tip>
-
-## 커뮤니티 파이프라인은 어떻게 작동하나요?
-
-커뮤니티 파이프라인은 [`DiffusionPipeline`]을 상속하는 클래스입니다:
-
-- [`custom_pipeline`] 인수로 로드할 수 있습니다.
-- 모델 가중치 및 스케줄러 구성은 [`pretrained_model_name_or_path`]에서 로드됩니다.
-- 커뮤니티 파이프라인에서 기능을 구현하는 코드는 `pipeline.py` 파일에 정의되어 있습니다.
-
-공식 저장소에서 모든 파이프라인 구성 요소 가중치를 로드할 수 없는 경우가 있습니다. 이 경우 다른 구성 요소는 파이프라인에 직접 전달해야 합니다:
-
-```python
-from diffusers import DiffusionPipeline
-from transformers import CLIPFeatureExtractor, CLIPModel
-
-model_id = "CompVis/stable-diffusion-v1-4"
-clip_model_id = "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
-
-feature_extractor = CLIPFeatureExtractor.from_pretrained(clip_model_id)
-clip_model = CLIPModel.from_pretrained(clip_model_id, torch_dtype=torch.float16)
-
-pipeline = DiffusionPipeline.from_pretrained(
-    model_id,
-    custom_pipeline="clip_guided_stable_diffusion",
-    clip_model=clip_model,
-    feature_extractor=feature_extractor,
-    scheduler=scheduler,
-    torch_dtype=torch.float16,
-)
-```
-
-커뮤니티 파이프라인의 마법은 다음 코드에 담겨 있습니다. 이 코드를 통해 커뮤니티 파이프라인을 GitHub 또는 Hub에서 로드할 수 있으며, 모든 🧨 Diffusers 패키지에서 사용할 수 있습니다.
-
-```python
-# 2. 파이프라인 클래스를 로드합니다. 사용자 지정 모듈을 사용하는 경우 Hub에서 로드합니다
-# 명시적 클래스에서 로드하는 경우, 이를 사용해 보겠습니다.
-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"])
-```

docs/source/ko/using-diffusers/control_brightness.md

@@ -1,45 +0,0 @@
-# 이미지 밝기 조절하기
-
-Stable Diffusion 파이프라인은 [일반적인 디퓨전 노이즈 스케줄과 샘플 단계에 결함이 있음](https://huggingface.co/papers/2305.08891) 논문에서 설명한 것처럼 매우 밝거나 어두운 이미지를 생성하는 데는 성능이 평범합니다. 이 논문에서 제안한 솔루션은 현재 [`DDIMScheduler`]에 구현되어 있으며 이미지의 밝기를 개선하는 데 사용할 수 있습니다.
-
-<Tip>
-
-💡 제안된 솔루션에 대한 자세한 내용은 위에 링크된 논문을 참고하세요!
-
-</Tip>
-
-해결책 중 하나는 *v 예측값*과 *v 로스*로 모델을 훈련하는 것입니다. 다음 flag를 [`train_text_to_image.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) 또는 [`train_text_to_image_lora.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py) 스크립트에 추가하여 `v_prediction`을 활성화합니다:
-
-```bash
---prediction_type="v_prediction"
-```
-
-예를 들어, `v_prediction`으로 미세 조정된 [`ptx0/pseudo-journey-v2`](https://huggingface.co/ptx0/pseudo-journey-v2) 체크포인트를 사용해 보겠습니다.
-
-다음으로 [`DDIMScheduler`]에서 다음 파라미터를 설정합니다:
-
-1. rescale_betas_zero_snr=True`, 노이즈 스케줄을 제로 터미널 신호 대 잡음비(SNR)로 재조정합니다.
-2. `timestep_spacing="trailing"`, 마지막 타임스텝부터 샘플링 시작
-
-```py
->>> from diffusers import DiffusionPipeline, DDIMScheduler
-
->>> pipeline = DiffusionPipeline.from_pretrained("ptx0/pseudo-journey-v2")
-# switch the scheduler in the pipeline to use the DDIMScheduler
-
->>> pipeline.scheduler = DDIMScheduler.from_config(
-...     pipeline.scheduler.config, rescale_betas_zero_snr=True, timestep_spacing="trailing"
-... )
->>> pipeline.to("cuda")
-```
-
-마지막으로 파이프라인에 대한 호출에서 `guidance_rescale`을 설정하여 과다 노출을 방지합니다:
-
-```py
-prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k"
-image = pipeline(prompt, guidance_rescale=0.7).images[0]
-```
-
-<div class="flex justify-center">
-    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/zero_snr.png"/>
-</div>

docs/source/ko/using-diffusers/custom_pipeline_examples.md

@@ -1,275 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# 커뮤니티 파이프라인
-
-> **커뮤니티 파이프라인에 대한 자세한 내용은 [이 이슈](https://github.com/huggingface/diffusers/issues/841)를 참조하세요.
-
-**커뮤니티** 예제는 커뮤니티에서 추가한 추론 및 훈련 예제로 구성되어 있습니다.
-다음 표를 참조하여 모든 커뮤니티 예제에 대한 개요를 확인하시기 바랍니다. **코드 예제**를 클릭하면 복사하여 붙여넣기할 수 있는 코드 예제를 확인할 수 있습니다.
-커뮤니티가 예상대로 작동하지 않는 경우 이슈를 개설하고 작성자에게 핑을 보내주세요.
-
-| 예                                     | 설명                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             | 코드 예제                                                                         | 콜랩                                                                                                                                                                                                              |저자                                                         |
-|:---------------------------------------|:------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------:|
-| CLIP Guided Stable Diffusion           | CLIP 가이드 기반의 Stable Diffusion으로 텍스트에서 이미지로 생성하기                                                                                                                                                                                                                                                                                                                                                                                                                                | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)                    | [![콜랩에서 열기](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)                 | 커뮤니티 파이프라인을 어떻게 사용해야 하는지에 대한 예시(참고 https://github.com/huggingface/diffusers/issues/841)                                                                                                                                                                                                                                                                                                                                                                                   | [One Step U-Net](#one-step-unet)                                                 | -                                                                                                                                                                                                                  | [Patrick von Platen](https://github.com/patrickvonplaten/) |
-| Stable Diffusion Interpolation         | 서로 다른 프롬프트/시드 간 Stable Diffusion의 latent space 보간                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Stable Diffusion Interpolation](#stable-diffusion-interpolation)                | -                                                                                                                                                                                                                  | [Nate Raw](https://github.com/nateraw/)                    |
-| Stable Diffusion Mega                  | 모든 기능을 갖춘 **하나의** Stable Diffusion 파이프라인 [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 | 토큰 길이 제한이 없고 프롬프트에서 파싱 가중치 지원을 하는 **하나의** Stable Diffusion 파이프라인,                                                                                                                                                                                                                                                                                                                                                                                                    | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion) |-                                                                                                                                                                                                                  | [SkyTNT](https://github.com/SkyTNT)                        |
-| Speech to Image                        | 자동 음성 인식을 사용하여 텍스트를 작성하고 Stable Diffusion을 사용하여 이미지를 생성합니다.                                                                                                                                                                                                                                                                                                                                                                                                          | [Speech to Image](#speech-to-image)                                               | -                                                                                                                                                                                                                 | [Mikail Duzenli](https://github.com/MikailINTech)          |
-
-커스텀 파이프라인을 불러오려면 `diffusers/examples/community`에 있는 파일 중 하나로서 `custom_pipeline` 인수를 `DiffusionPipeline`에 전달하기만 하면 됩니다. 자신만의 파이프라인이 있는 PR을 보내주시면 빠르게 병합해드리겠습니다.
-```py
-pipe = DiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4", custom_pipeline="filename_in_the_community_folder"
-)
-```
-
-## 사용 예시
-
-### CLIP 가이드 기반의 Stable Diffusion
-
-모든 노이즈 제거 단계에서 추가 CLIP 모델을 통해 Stable Diffusion을 가이드함으로써 CLIP 모델 기반의 Stable Diffusion은 보다 더 사실적인 이미지를 생성을 할 수 있습니다.
-
-다음 코드는 약 12GB의 GPU RAM이 필요합니다.
-
-```python
-from diffusers import DiffusionPipeline
-from transformers import CLIPImageProcessor, CLIPModel
-import torch
-
-
-feature_extractor = CLIPImageProcessor.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,
-    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)
-
-# 이미지 로컬에 저장하기
-for i, img in enumerate(images):
-    img.save(f"./clip_guided_sd/image_{i}.png")
-```
-
-이미지` 목록에는 로컬에 저장하거나 구글 콜랩에 직접 표시할 수 있는 PIL 이미지 목록이 포함되어 있습니다. 생성된 이미지는 기본적으로 안정적인 확산을 사용하는 것보다 품질이 높은 경향이 있습니다. 예를 들어 위의 스크립트는 다음과 같은 이미지를 생성합니다:
-
-![clip_guidance](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/clip_guidance/merged_clip_guidance.jpg).
-
-### One Step Unet
-
-예시 "one-step-unet"는 다음과 같이 실행할 수 있습니다.
-
-```python
-from diffusers import DiffusionPipeline
-
-pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
-pipe()
-```
-
-**참고**: 이 커뮤니티 파이프라인은 기능으로 유용하지 않으며 커뮤니티 파이프라인을 추가할 수 있는 방법의 예시일 뿐입니다(https://github.com/huggingface/diffusers/issues/841 참조).
-
-### Stable Diffusion Interpolation
-
-다음 코드는 최소 8GB VRAM의 GPU에서 실행할 수 있으며 약 5분 정도 소요됩니다.
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-pipe = DiffusionPipeline.from_pretrained(
-    "CompVis/stable-diffusion-v1-4",
-    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,
-)
-```
-
-walk(...)` 함수의 출력은 `output_dir`에 정의된 대로 폴더에 저장된 이미지 목록을 반환합니다. 이 이미지를 사용하여 안정적으로 확산되는 동영상을 만들 수 있습니다.
-
-> 안정된 확산을 이용한 동영상 제작 방법과 더 많은 기능에 대한 자세한 내용은 https://github.com/nateraw/stable-diffusion-videos 에서 확인하시기 바랍니다.
-
-### Stable Diffusion Mega
-
-The Stable Diffusion Mega 파이프라인을 사용하면 Stable Diffusion 파이프라인의 주요 사용 사례를 단일 클래스에서 사용할 수 있습니다.
-```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,
-)
-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, 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, image=init_image, mask_image=mask_image, strength=0.75).images
-```
-
-위에 표시된 것처럼 하나의 파이프라인에서 '텍스트-이미지 변환', '이미지-이미지 변환', '인페인팅'을 모두 실행할 수 있습니다.
-
-### Long Prompt Weighting Stable Diffusion
-
-파이프라인을 사용하면 77개의 토큰 길이 제한 없이 프롬프트를 입력할 수 있습니다. 또한 "()"를 사용하여 단어 가중치를 높이거나 "[]"를 사용하여 단어 가중치를 낮출 수 있습니다.
-또한 파이프라인을 사용하면 단일 클래스에서 Stable Diffusion 파이프라인의 주요 사용 사례를 사용할 수 있습니다.
-
-#### pytorch
-
-```python
-from diffusers import DiffusionPipeline
-import torch
-
-pipe = DiffusionPipeline.from_pretrained(
-    "hakurei/waifu-diffusion", custom_pipeline="lpw_stable_diffusion", 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]
-```
-
-토큰 인덱스 시퀀스 길이가 이 모델에 지정된 최대 시퀀스 길이보다 길면(*** > 77). 이 시퀀스를 모델에서 실행하면 인덱싱 오류가 발생합니다`. 정상적인 현상이니 걱정하지 마세요.
-### Speech to Image
-
-다음 코드는 사전학습된 OpenAI whisper-small과 Stable Diffusion을 사용하여 오디오 샘플에서 이미지를 생성할 수 있습니다.
-```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,
-
-    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])
-```
-위 예시는 다음의 결과 이미지를 보입니다.
-
-![image](https://user-images.githubusercontent.com/45072645/196901736-77d9c6fc-63ee-4072-90b0-dc8b903d63e3.png)

docs/source/ko/using-diffusers/diffedit.md

@@ -0,0 +1,285 @@
+<!--Copyright 2023 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.
+-->
+
+# DiffEdit
+
+[[open-in-colab]]
+
+이미지 편집을 하려면 일반적으로 편집할 영역의 마스크를 제공해야 합니다. DiffEdit는 텍스트 쿼리를 기반으로 마스크를 자동으로 생성하므로 이미지 편집 소프트웨어 없이도 마스크를 만들기가 전반적으로 더 쉬워집니다. DiffEdit 알고리즘은 세 단계로 작동합니다:
+
+1. Diffusion 모델이 일부 쿼리 텍스트와 참조 텍스트를 조건부로 이미지의 노이즈를 제거하여 이미지의 여러 영역에 대해 서로 다른 노이즈 추정치를 생성하고, 그 차이를 사용하여 쿼리 텍스트와 일치하도록 이미지의 어느 영역을 변경해야 하는지 식별하기 위한 마스크를 추론합니다.
+2. 입력 이미지가 DDIM을 사용하여 잠재 공간으로 인코딩됩니다.
+3. 마스크 외부의 픽셀이 입력 이미지와 동일하게 유지되도록 마스크를 가이드로 사용하여 텍스트 쿼리에 조건이 지정된 diffusion 모델로 latents를 디코딩합니다.
+
+이 가이드에서는 마스크를 수동으로 만들지 않고 DiffEdit를 사용하여 이미지를 편집하는 방법을 설명합니다.
+
+시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:
+
+```py
+# Colab에서 필요한 라이브러리를 설치하기 위해 주석을 제외하세요
+#!pip install -q diffusers transformers accelerate
+```
+
+[`StableDiffusionDiffEditPipeline`]에는 이미지 마스크와 부분적으로 반전된 latents 집합이 필요합니다. 이미지 마스크는 [`~StableDiffusionDiffEditPipeline.generate_mask`] 함수에서 생성되며, 두 개의 파라미터인 `source_prompt`와 `target_prompt`가 포함됩니다. 이 매개변수는 이미지에서 무엇을 편집할지 결정합니다. 예를 들어, *과일* 한 그릇을 *배* 한 그릇으로 변경하려면 다음과 같이 하세요:
+
+```py
+source_prompt = "a bowl of fruits"
+target_prompt = "a bowl of pears"
+```
+
+부분적으로 반전된 latents는 [`~StableDiffusionDiffEditPipeline.invert`] 함수에서 생성되며, 일반적으로 이미지를 설명하는 `prompt` 또는 *캡션*을 포함하는 것이 inverse latent sampling 프로세스를 가이드하는 데 도움이 됩니다. 캡션은 종종 `source_prompt`가 될 수 있지만, 다른 텍스트 설명으로 자유롭게 실험해 보세요!
+
+파이프라인, 스케줄러, 역 스케줄러를 불러오고 메모리 사용량을 줄이기 위해 몇 가지 최적화를 활성화해 보겠습니다:
+
+```py
+import torch
+from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionDiffEditPipeline
+
+pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-2-1",
+    torch_dtype=torch.float16,
+    safety_checker=None,
+    use_safetensors=True,
+)
+pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
+pipeline.enable_model_cpu_offload()
+pipeline.enable_vae_slicing()
+```
+
+수정하기 위한 이미지를 불러옵니다:
+
+```py
+from diffusers.utils import load_image, make_image_grid
+
+img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
+raw_image = load_image(img_url).resize((768, 768))
+raw_image
+```
+
+이미지 마스크를 생성하기 위해 [`~StableDiffusionDiffEditPipeline.generate_mask`] 함수를 사용합니다. 이미지에서 편집할 내용을 지정하기 위해 `source_prompt`와 `target_prompt`를 전달해야 합니다:
+
+```py
+from PIL import Image
+
+source_prompt = "a bowl of fruits"
+target_prompt = "a basket of pears"
+mask_image = pipeline.generate_mask(
+    image=raw_image,
+    source_prompt=source_prompt,
+    target_prompt=target_prompt,
+)
+Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
+```
+
+다음으로, 반전된 latents를 생성하고 이미지를 묘사하는 캡션에 전달합니다:
+
+```py
+inv_latents = pipeline.invert(prompt=source_prompt, image=raw_image).latents
+```
+
+마지막으로, 이미지 마스크와 반전된 latents를 파이프라인에 전달합니다. `target_prompt`는 이제 `prompt`가 되며, `source_prompt`는 `negative_prompt`로 사용됩니다.
+
+```py
+output_image = pipeline(
+    prompt=target_prompt,
+    mask_image=mask_image,
+    image_latents=inv_latents,
+    negative_prompt=source_prompt,
+).images[0]
+mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
+make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">original image</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/blob/main/assets/target.png?raw=true"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">edited image</figcaption>
+  </div>
+</div>
+
+## Source와 target 임베딩 생성하기
+
+Source와 target 임베딩은 수동으로 생성하는 대신 [Flan-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5) 모델을 사용하여 자동으로 생성할 수 있습니다.
+
+Flan-T5 모델과 토크나이저를 🤗 Transformers 라이브러리에서 불러옵니다:
+
+```py
+import torch
+from transformers import AutoTokenizer, T5ForConditionalGeneration
+
+tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")
+model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large", device_map="auto", torch_dtype=torch.float16)
+```
+
+모델에 프롬프트할 source와 target 프롬프트를 생성하기 위해 초기 텍스트들을 제공합니다.
+
+```py
+source_concept = "bowl"
+target_concept = "basket"
+
+source_text = f"Provide a caption for images containing a {source_concept}. "
+"The captions should be in English and should be no longer than 150 characters."
+
+target_text = f"Provide a caption for images containing a {target_concept}. "
+"The captions should be in English and should be no longer than 150 characters."
+```
+
+다음으로, 프롬프트들을 생성하기 위해 유틸리티 함수를 생성합니다.
+
+```py
+@torch.no_grad()
+def generate_prompts(input_prompt):
+    input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids.to("cuda")
+
+    outputs = model.generate(
+        input_ids, temperature=0.8, num_return_sequences=16, do_sample=True, max_new_tokens=128, top_k=10
+    )
+    return tokenizer.batch_decode(outputs, skip_special_tokens=True)
+
+source_prompts = generate_prompts(source_text)
+target_prompts = generate_prompts(target_text)
+print(source_prompts)
+print(target_prompts)
+```
+
+<Tip>
+
+다양한 품질의 텍스트를 생성하는 전략에 대해 자세히 알아보려면 [생성 전략](https://huggingface.co/docs/transformers/main/en/generation_strategies) 가이드를 참조하세요.
+
+</Tip>
+
+텍스트 인코딩을 위해 [`StableDiffusionDiffEditPipeline`]에서 사용하는 텍스트 인코더 모델을 불러옵니다. 텍스트 인코더를 사용하여 텍스트 임베딩을 계산합니다:
+
+```py
+import torch
+from diffusers import StableDiffusionDiffEditPipeline
+
+pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16, use_safetensors=True
+)
+pipeline.enable_model_cpu_offload()
+pipeline.enable_vae_slicing()
+
+@torch.no_grad()
+def embed_prompts(sentences, tokenizer, text_encoder, device="cuda"):
+    embeddings = []
+    for sent in sentences:
+        text_inputs = tokenizer(
+            sent,
+            padding="max_length",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_embeds = text_encoder(text_input_ids.to(device), attention_mask=None)[0]
+        embeddings.append(prompt_embeds)
+    return torch.concatenate(embeddings, dim=0).mean(dim=0).unsqueeze(0)
+
+source_embeds = embed_prompts(source_prompts, pipeline.tokenizer, pipeline.text_encoder)
+target_embeds = embed_prompts(target_prompts, pipeline.tokenizer, pipeline.text_encoder)
+```
+
+마지막으로, 임베딩을 [`~StableDiffusionDiffEditPipeline.generate_mask`] 및 [`~StableDiffusionDiffEditPipeline.invert`] 함수와 파이프라인에 전달하여 이미지를 생성합니다:
+
+```diff
+  from diffusers import DDIMInverseScheduler, DDIMScheduler
+  from diffusers.utils import load_image, make_image_grid
+  from PIL import Image
+
+  pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+  pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
+
+  img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
+  raw_image = load_image(img_url).resize((768, 768))
+
+  mask_image = pipeline.generate_mask(
+      image=raw_image,
+-     source_prompt=source_prompt,
+-     target_prompt=target_prompt,
++     source_prompt_embeds=source_embeds,
++     target_prompt_embeds=target_embeds,
+  )
+
+  inv_latents = pipeline.invert(
+-     prompt=source_prompt,
++     prompt_embeds=source_embeds,
+      image=raw_image,
+  ).latents
+
+  output_image = pipeline(
+      mask_image=mask_image,
+      image_latents=inv_latents,
+-     prompt=target_prompt,
+-     negative_prompt=source_prompt,
++     prompt_embeds=target_embeds,
++     negative_prompt_embeds=source_embeds,
+  ).images[0]
+  mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L")
+  make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
+```
+
+## 반전을 위한 캡션 생성하기
+
+`source_prompt`를 캡션으로 사용하여 부분적으로 반전된 latents를 생성할 수 있지만, [BLIP](https://huggingface.co/docs/transformers/model_doc/blip) 모델을 사용하여 캡션을 자동으로 생성할 수도 있습니다.
+
+🤗 Transformers 라이브러리에서 BLIP 모델과 프로세서를 불러옵니다:
+
+```py
+import torch
+from transformers import BlipForConditionalGeneration, BlipProcessor
+
+processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base", torch_dtype=torch.float16, low_cpu_mem_usage=True)
+```
+
+입력 이미지에서 캡션을 생성하는 유틸리티 함수를 만듭니다:
+
+```py
+@torch.no_grad()
+def generate_caption(images, caption_generator, caption_processor):
+    text = "a photograph of"
+
+    inputs = caption_processor(images, text, return_tensors="pt").to(device="cuda", dtype=caption_generator.dtype)
+    caption_generator.to("cuda")
+    outputs = caption_generator.generate(**inputs, max_new_tokens=128)
+
+    # 캡션 generator 오프로드
+    caption_generator.to("cpu")
+
+    caption = caption_processor.batch_decode(outputs, skip_special_tokens=True)[0]
+    return caption
+```
+
+입력 이미지를 불러오고 `generate_caption` 함수를 사용하여 해당 이미지에 대한 캡션을 생성합니다:
+
+```py
+from diffusers.utils import load_image
+
+img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
+raw_image = load_image(img_url).resize((768, 768))
+caption = generate_caption(raw_image, model, processor)
+```
+
+<div class="flex justify-center">
+    <figure>
+        <img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
+        <figcaption class="text-center">generated caption: "a photograph of a bowl of fruit on a table"</figcaption>
+    </figure>
+</div>
+
+이제 캡션을 [`~StableDiffusionDiffEditPipeline.invert`] 함수에 놓아 부분적으로 반전된 latents를 생성할 수 있습니다!

docs/source/ko/using-diffusers/kandinsky.md

@@ -0,0 +1,768 @@
+<!--Copyright 2023 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.
+-->
+
+# Kandinsky
+
+[[open-in-colab]]
+
+Kandinsky 모델은 일련의 다국어 text-to-image 생성 모델입니다. Kandinsky 2.0 모델은 두 개의 다국어 텍스트 인코더를 사용하고 그 결과를 연결해 UNet에 사용됩니다.
+
+[Kandinsky 2.1](../api/pipelines/kandinsky)은 텍스트와 이미지 임베딩 간의 매핑을 생성하는 image prior 모델([`CLIP`](https://huggingface.co/docs/transformers/model_doc/clip))을 포함하도록 아키텍처를 변경했습니다. 이 매핑은 더 나은 text-image alignment를 제공하며, 학습 중에 텍스트 임베딩과 함께 사용되어 더 높은 품질의 결과를 가져옵니다. 마지막으로, Kandinsky 2.1은 spatial conditional 정규화 레이어를 추가하여 사실감을 높여주는 [Modulating Quantized Vectors (MoVQ)](https://huggingface.co/papers/2209.09002) 디코더를 사용하여 latents를 이미지로 디코딩합니다.
+
+[Kandinsky 2.2](../api/pipelines/kandinsky_v22)는 image prior 모델의 이미지 인코더를 더 큰 CLIP-ViT-G 모델로 교체하여 품질을 개선함으로써 이전 모델을 개선했습니다. 또한 image prior 모델은 해상도와 종횡비가 다른 이미지로 재훈련되어 더 높은 해상도의 이미지와 다양한 이미지 크기를 생성합니다.
+
+[Kandinsky 3](../api/pipelines/kandinsky3)는 아키텍처를 단순화하고 prior 모델과 diffusion 모델을 포함하는 2단계 생성 프로세스에서 벗어나고 있습니다. 대신, Kandinsky 3는 [Flan-UL2](https://huggingface.co/google/flan-ul2)를 사용하여 텍스트를 인코딩하고, [BigGan-deep](https://hf.co/papers/1809.11096) 블록이 포함된 UNet을 사용하며, [Sber-MoVQGAN](https://github.com/ai-forever/MoVQGAN)을 사용하여 latents를 이미지로 디코딩합니다. 텍스트 이해와 생성된 이미지 품질은 주로 더 큰 텍스트 인코더와 UNet을 사용함으로써 달성됩니다.
+
+이 가이드에서는 text-to-image, image-to-image, 인페인팅, 보간 등을 위해 Kandinsky 모델을 사용하는 방법을 설명합니다.
+
+시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:
+
+```py
+# Colab에서 필요한 라이브러리를 설치하기 위해 주석을 제외하세요
+#!pip install -q diffusers transformers accelerate
+```
+
+<Tip warning={true}>
+
+Kandinsky 2.1과 2.2의 사용법은 매우 유사합니다! 유일한 차이점은 Kandinsky 2.2는 latents를 디코딩할 때 `프롬프트`를 입력으로 받지 않는다는 것입니다. 대신, Kandinsky 2.2는 디코딩 중에는 `image_embeds`만 받아들입니다.
+
+<br>
+
+Kandinsky 3는 더 간결한 아키텍처를 가지고 있으며 prior 모델이 필요하지 않습니다. 즉, [Stable Diffusion XL](sdxl)과 같은 다른 diffusion 모델과 사용법이 동일합니다.
+
+</Tip>
+
+## Text-to-image
+
+모든 작업에 Kandinsky 모델을 사용하려면 항상 프롬프트를 인코딩하고 이미지 임베딩을 생성하는 prior 파이프라인을 설정하는 것부터 시작해야 합니다. 이전 파이프라인은 negative 프롬프트 `""`에 해당하는 `negative_image_embeds`도 생성합니다. 더 나은 결과를 얻으려면 이전 파이프라인에 실제 `negative_prompt`를 전달할 수 있지만, 이렇게 하면 prior 파이프라인의 유효 배치 크기가 2배로 증가합니다.
+
+<hfoptions id="text-to-image">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers import KandinskyPriorPipeline, KandinskyPipeline
+import torch
+
+prior_pipeline = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16).to("cuda")
+pipeline = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16).to("cuda")
+
+prompt = "A alien cheeseburger creature eating itself, claymation, cinematic, moody lighting"
+negative_prompt = "low quality, bad quality" # negative 프롬프트 포함은 선택적이지만, 보통 결과는 더 좋습니다
+image_embeds, negative_image_embeds = prior_pipeline(prompt, negative_prompt, guidance_scale=1.0).to_tuple()
+```
+
+이제 모든 프롬프트와 임베딩을 [`KandinskyPipeline`]에 전달하여 이미지를 생성합니다:
+
+```py
+image = pipeline(prompt, image_embeds=image_embeds, negative_prompt=negative_prompt, negative_image_embeds=negative_image_embeds, height=768, width=768).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-docs/cheeseburger.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers import KandinskyV22PriorPipeline, KandinskyV22Pipeline
+import torch
+
+prior_pipeline = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16).to("cuda")
+pipeline = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16).to("cuda")
+
+prompt = "A alien cheeseburger creature eating itself, claymation, cinematic, moody lighting"
+negative_prompt = "low quality, bad quality" # negative 프롬프트 포함은 선택적이지만, 보통 결과는 더 좋습니다
+image_embeds, negative_image_embeds = prior_pipeline(prompt, guidance_scale=1.0).to_tuple()
+```
+
+이미지 생성을 위해 `image_embeds`와 `negative_image_embeds`를 [`KandinskyV22Pipeline`]에 전달합니다:
+
+```py
+image = pipeline(image_embeds=image_embeds, negative_image_embeds=negative_image_embeds, height=768, width=768).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-text-to-image.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 3">
+
+Kandinsky 3는 prior 모델이 필요하지 않으므로 [`Kandinsky3Pipeline`]을 직접 불러오고 이미지 생성 프롬프트를 전달할 수 있습니다:
+
+```py
+from diffusers import Kandinsky3Pipeline
+import torch
+
+pipeline = Kandinsky3Pipeline.from_pretrained("kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16)
+pipeline.enable_model_cpu_offload()
+
+prompt = "A alien cheeseburger creature eating itself, claymation, cinematic, moody lighting"
+image = pipeline(prompt).images[0]
+image
+```
+
+</hfoption>
+</hfoptions>
+
+🤗 Diffusers는 또한 [`KandinskyCombinedPipeline`] 및 [`KandinskyV22CombinedPipeline`]이 포함된 end-to-end API를 제공하므로 prior 파이프라인과 text-to-image 변환 파이프라인을 별도로 불러올 필요가 없습니다. 결합된 파이프라인은 prior 모델과 디코더를 모두 자동으로 불러옵니다. 원하는 경우 `prior_guidance_scale` 및 `prior_num_inference_steps` 매개 변수를 사용하여 prior 파이프라인에 대해 다른 값을 설정할 수 있습니다.
+
+내부에서 결합된 파이프라인을 자동으로 호출하려면 [`AutoPipelineForText2Image`]를 사용합니다:
+
+<hfoptions id="text-to-image">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16)
+pipeline.enable_model_cpu_offload()
+
+prompt = "A alien cheeseburger creature eating itself, claymation, cinematic, moody lighting"
+negative_prompt = "low quality, bad quality"
+
+image = pipeline(prompt=prompt, negative_prompt=negative_prompt, prior_guidance_scale=1.0, guidance_scale=4.0, height=768, width=768).images[0]
+image
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16)
+pipeline.enable_model_cpu_offload()
+
+prompt = "A alien cheeseburger creature eating itself, claymation, cinematic, moody lighting"
+negative_prompt = "low quality, bad quality"
+
+image = pipeline(prompt=prompt, negative_prompt=negative_prompt, prior_guidance_scale=1.0, guidance_scale=4.0, height=768, width=768).images[0]
+image
+```
+
+</hfoption>
+</hfoptions>
+
+## Image-to-image
+
+Image-to-image 경우, 초기 이미지와 텍스트 프롬프트를 전달하여 파이프라인에 이미지를 conditioning합니다. Prior 파이프라인을 불러오는 것으로 시작합니다:
+
+<hfoptions id="image-to-image">
+<hfoption id="Kandinsky 2.1">
+
+```py
+import torch
+from diffusers import KandinskyImg2ImgPipeline, KandinskyPriorPipeline
+
+prior_pipeline = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipeline = KandinskyImg2ImgPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+import torch
+from diffusers import KandinskyV22Img2ImgPipeline, KandinskyPriorPipeline
+
+prior_pipeline = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipeline = KandinskyV22Img2ImgPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+```
+
+</hfoption>
+<hfoption id="Kandinsky 3">
+
+Kandinsky 3는 prior 모델이 필요하지 않으므로 image-to-image 파이프라인을 직접 불러올 수 있습니다:
+
+```py
+from diffusers import Kandinsky3Img2ImgPipeline
+from diffusers.utils import load_image
+import torch
+
+pipeline = Kandinsky3Img2ImgPipeline.from_pretrained("kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16)
+pipeline.enable_model_cpu_offload()
+```
+
+</hfoption>
+</hfoptions>
+
+Conditioning할 이미지를 다운로드합니다:
+
+```py
+from diffusers.utils import load_image
+
+# 이미지 다운로드
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+original_image = load_image(url)
+original_image = original_image.resize((768, 512))
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"/>
+</div>
+
+Prior 파이프라인으로 `image_embeds`와 `negative_image_embeds`를 생성합니다:
+
+```py
+prompt = "A fantasy landscape, Cinematic lighting"
+negative_prompt = "low quality, bad quality"
+
+image_embeds, negative_image_embeds = prior_pipeline(prompt, negative_prompt).to_tuple()
+```
+
+이제 원본 이미지와 모든 프롬프트 및 임베딩을 파이프라인으로 전달하여 이미지를 생성합니다:
+
+<hfoptions id="image-to-image">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers.utils import make_image_grid
+
+image = pipeline(prompt, negative_prompt=negative_prompt, image=original_image, image_embeds=image_embeds, negative_image_embeds=negative_image_embeds, height=768, width=768, strength=0.3).images[0]
+make_image_grid([original_image.resize((512, 512)), image.resize((512, 512))], rows=1, cols=2)
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-docs/img2img_fantasyland.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers.utils import make_image_grid
+
+image = pipeline(image=original_image, image_embeds=image_embeds, negative_image_embeds=negative_image_embeds, height=768, width=768, strength=0.3).images[0]
+make_image_grid([original_image.resize((512, 512)), image.resize((512, 512))], rows=1, cols=2)
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-image-to-image.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 3">
+
+```py
+image = pipeline(prompt, negative_prompt=negative_prompt, image=image, strength=0.75, num_inference_steps=25).images[0]
+image
+```
+
+</hfoption>
+</hfoptions>
+
+또한 🤗 Diffusers에서는 [`KandinskyImg2ImgCombinedPipeline`] 및 [`KandinskyV22Img2ImgCombinedPipeline`]이 포함된 end-to-end API를 제공하므로 prior 파이프라인과 image-to-image 파이프라인을 별도로 불러올 필요가 없습니다. 결합된 파이프라인은 prior 모델과 디코더를 모두 자동으로 불러옵니다. 원하는 경우 `prior_guidance_scale` 및 `prior_num_inference_steps` 매개 변수를 사용하여 이전 파이프라인에 대해 다른 값을 설정할 수 있습니다.
+
+내부에서 결합된 파이프라인을 자동으로 호출하려면 [`AutoPipelineForImage2Image`]를 사용합니다:
+
+<hfoptions id="image-to-image">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers import AutoPipelineForImage2Image
+from diffusers.utils import make_image_grid, load_image
+import torch
+
+pipeline = AutoPipelineForImage2Image.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16, use_safetensors=True)
+pipeline.enable_model_cpu_offload()
+
+prompt = "A fantasy landscape, Cinematic lighting"
+negative_prompt = "low quality, bad quality"
+
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+original_image = load_image(url)
+
+original_image.thumbnail((768, 768))
+
+image = pipeline(prompt=prompt, negative_prompt=negative_prompt, image=original_image, strength=0.3).images[0]
+make_image_grid([original_image.resize((512, 512)), image.resize((512, 512))], rows=1, cols=2)
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers import AutoPipelineForImage2Image
+from diffusers.utils import make_image_grid, load_image
+import torch
+
+pipeline = AutoPipelineForImage2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16)
+pipeline.enable_model_cpu_offload()
+
+prompt = "A fantasy landscape, Cinematic lighting"
+negative_prompt = "low quality, bad quality"
+
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+original_image = load_image(url)
+
+original_image.thumbnail((768, 768))
+
+image = pipeline(prompt=prompt, negative_prompt=negative_prompt, image=original_image, strength=0.3).images[0]
+make_image_grid([original_image.resize((512, 512)), image.resize((512, 512))], rows=1, cols=2)
+```
+
+</hfoption>
+</hfoptions>
+
+## Inpainting
+
+<Tip warning={true}>
+
+⚠️ Kandinsky 모델은 이제 검은색 픽셀 대신 ⬜️ **흰색 픽셀**을 사용하여 마스크 영역을 표현합니다. 프로덕션에서 [`KandinskyInpaintPipeline`]을 사용하는 경우 흰색 픽셀을 사용하도록 마스크를 변경해야 합니다:
+
+```py
+# PIL 입력에 대해
+import PIL.ImageOps
+mask = PIL.ImageOps.invert(mask)
+
+# PyTorch와 NumPy 입력에 대해
+mask = 1 - mask
+```
+
+</Tip>
+
+인페인팅에서는 원본 이미지, 원본 이미지에서 대체할 영역의 마스크, 인페인팅할 내용에 대한 텍스트 프롬프트가 필요합니다. Prior 파이프라인을 불러옵니다:
+
+<hfoptions id="inpaint">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers import KandinskyInpaintPipeline, KandinskyPriorPipeline
+from diffusers.utils import load_image, make_image_grid
+import torch
+import numpy as np
+from PIL import Image
+
+prior_pipeline = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipeline = KandinskyInpaintPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers import KandinskyV22InpaintPipeline, KandinskyV22PriorPipeline
+from diffusers.utils import load_image, make_image_grid
+import torch
+import numpy as np
+from PIL import Image
+
+prior_pipeline = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipeline = KandinskyV22InpaintPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+```
+
+</hfoption>
+</hfoptions>
+
+초기 이미지를 불러오고 마스크를 생성합니다:
+
+```py
+init_image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png")
+mask = np.zeros((768, 768), dtype=np.float32)
+# mask area above cat's head
+mask[:250, 250:-250] = 1
+```
+
+Prior 파이프라인으로 임베딩을 생성합니다:
+
+```py
+prompt = "a hat"
+prior_output = prior_pipeline(prompt)
+```
+
+이제 이미지 생성을 위해 초기 이미지, 마스크, 프롬프트와 임베딩을 파이프라인에 전달합니다:
+
+<hfoptions id="inpaint">
+<hfoption id="Kandinsky 2.1">
+
+```py
+output_image = pipeline(prompt, image=init_image, mask_image=mask, **prior_output, height=768, width=768, num_inference_steps=150).images[0]
+mask = Image.fromarray((mask*255).astype('uint8'), 'L')
+make_image_grid([init_image, mask, output_image], rows=1, cols=3)
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-docs/inpaint_cat_hat.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+output_image = pipeline(image=init_image, mask_image=mask, **prior_output, height=768, width=768, num_inference_steps=150).images[0]
+mask = Image.fromarray((mask*255).astype('uint8'), 'L')
+make_image_grid([init_image, mask, output_image], rows=1, cols=3)
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinskyv22-inpaint.png"/>
+</div>
+
+</hfoption>
+</hfoptions>
+
+[`KandinskyInpaintCombinedPipeline`] 및 [`KandinskyV22InpaintCombinedPipeline`]을 사용하여 내부에서 prior 및 디코더 파이프라인을 함께 호출할 수 있습니다. 이를 위해 [`AutoPipelineForInpainting`]을 사용합니다:
+
+<hfoptions id="inpaint">
+<hfoption id="Kandinsky 2.1">
+
+```py
+import torch
+import numpy as np
+from PIL import Image
+from diffusers import AutoPipelineForInpainting
+from diffusers.utils import load_image, make_image_grid
+
+pipe = AutoPipelineForInpainting.from_pretrained("kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16)
+pipe.enable_model_cpu_offload()
+
+init_image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png")
+mask = np.zeros((768, 768), dtype=np.float32)
+# 고양이 머리 위 마스크 지역
+mask[:250, 250:-250] = 1
+prompt = "a hat"
+
+output_image = pipe(prompt=prompt, image=init_image, mask_image=mask).images[0]
+mask = Image.fromarray((mask*255).astype('uint8'), 'L')
+make_image_grid([init_image, mask, output_image], rows=1, cols=3)
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+import torch
+import numpy as np
+from PIL import Image
+from diffusers import AutoPipelineForInpainting
+from diffusers.utils import load_image, make_image_grid
+
+pipe = AutoPipelineForInpainting.from_pretrained("kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16)
+pipe.enable_model_cpu_offload()
+
+init_image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png")
+mask = np.zeros((768, 768), dtype=np.float32)
+# 고양이 머리 위 마스크 영역
+mask[:250, 250:-250] = 1
+prompt = "a hat"
+
+output_image = pipe(prompt=prompt, image=original_image, mask_image=mask).images[0]
+mask = Image.fromarray((mask*255).astype('uint8'), 'L')
+make_image_grid([init_image, mask, output_image], rows=1, cols=3)
+```
+
+</hfoption>
+</hfoptions>
+
+## Interpolation (보간)
+
+Interpolation(보간)을 사용하면 이미지와 텍스트 임베딩 사이의 latent space를 탐색할 수 있어 prior 모델의 중간 결과물을 볼 수 있는 멋진 방법입니다. Prior 파이프라인과 보간하려는 두 개의 이미지를 불러옵니다:
+
+<hfoptions id="interpolate">
+<hfoption id="Kandinsky 2.1">
+
+```py
+from diffusers import KandinskyPriorPipeline, KandinskyPipeline
+from diffusers.utils import load_image, make_image_grid
+import torch
+
+prior_pipeline = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+img_1 = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png")
+img_2 = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/starry_night.jpeg")
+make_image_grid([img_1.resize((512,512)), img_2.resize((512,512))], rows=1, cols=2)
+```
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+from diffusers import KandinskyV22PriorPipeline, KandinskyV22Pipeline
+from diffusers.utils import load_image, make_image_grid
+import torch
+
+prior_pipeline = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+img_1 = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png")
+img_2 = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/starry_night.jpeg")
+make_image_grid([img_1.resize((512,512)), img_2.resize((512,512))], rows=1, cols=2)
+```
+
+</hfoption>
+</hfoptions>
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">a cat</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/starry_night.jpeg"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">Van Gogh's Starry Night painting</figcaption>
+  </div>
+</div>
+
+보간할 텍스트 또는 이미지를 지정하고 각 텍스트 또는 이미지에 대한 가중치를 설정합니다. 가중치를 실험하여 보간에 어떤 영향을 미치는지 확인하세요!
+
+```py
+images_texts = ["a cat", img_1, img_2]
+weights = [0.3, 0.3, 0.4]
+```
+
+`interpolate` 함수를 호출하여 임베딩을 생성한 다음, 파이프라인으로 전달하여 이미지를 생성합니다:
+
+<hfoptions id="interpolate">
+<hfoption id="Kandinsky 2.1">
+
+```py
+# 프롬프트는 빈칸으로 남겨도 됩니다
+prompt = ""
+prior_out = prior_pipeline.interpolate(images_texts, weights)
+
+pipeline = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+
+image = pipeline(prompt, **prior_out, height=768, width=768).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinsky-docs/starry_cat.png"/>
+</div>
+
+</hfoption>
+<hfoption id="Kandinsky 2.2">
+
+```py
+# 프롬프트는 빈칸으로 남겨도 됩니다
+prompt = ""
+prior_out = prior_pipeline.interpolate(images_texts, weights)
+
+pipeline = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+
+image = pipeline(prompt, **prior_out, height=768, width=768).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/kandinskyv22-interpolate.png"/>
+</div>
+
+</hfoption>
+</hfoptions>
+
+## ControlNet
+
+<Tip warning={true}>
+
+⚠️ ControlNet은 Kandinsky 2.2에서만 지원됩니다!
+
+</Tip>
+
+ControlNet을 사용하면 depth map이나 edge detection와 같은 추가 입력을 통해 사전학습된 large diffusion 모델을 conditioning할 수 있습니다. 예를 들어, 모델이 depth map의 구조를 이해하고 보존할 수 있도록 깊이 맵으로 Kandinsky 2.2를 conditioning할 수 있습니다.
+
+이미지를 불러오고 depth map을 추출해 보겠습니다:
+
+```py
+from diffusers.utils import load_image
+
+img = load_image(
+    "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/cat.png"
+).resize((768, 768))
+img
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/cat.png"/>
+</div>
+
+그런 다음 🤗 Transformers의 `depth-estimation` [`~transformers.Pipeline`]을 사용하여 이미지를 처리해 depth map을 구할 수 있습니다:
+
+```py
+import torch
+import numpy as np
+
+from transformers import pipeline
+
+def make_hint(image, depth_estimator):
+    image = depth_estimator(image)["depth"]
+    image = np.array(image)
+    image = image[:, :, None]
+    image = np.concatenate([image, image, image], axis=2)
+    detected_map = torch.from_numpy(image).float() / 255.0
+    hint = detected_map.permute(2, 0, 1)
+    return hint
+
+depth_estimator = pipeline("depth-estimation")
+hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda")
+```
+
+### Text-to-image [[controlnet-text-to-image]]
+
+Prior 파이프라인과 [`KandinskyV22ControlnetPipeline`]를 불러옵니다:
+
+```py
+from diffusers import KandinskyV22PriorPipeline, KandinskyV22ControlnetPipeline
+
+prior_pipeline = KandinskyV22PriorPipeline.from_pretrained(
+    "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16, use_safetensors=True
+).to("cuda")
+
+pipeline = KandinskyV22ControlnetPipeline.from_pretrained(
+    "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16
+).to("cuda")
+```
+
+프롬프트와 negative 프롬프트로 이미지 임베딩을 생성합니다:
+
+```py
+prompt = "A robot, 4k photo"
+negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature"
+
+generator = torch.Generator(device="cuda").manual_seed(43)
+
+image_emb, zero_image_emb = prior_pipeline(
+    prompt=prompt, negative_prompt=negative_prior_prompt, generator=generator
+).to_tuple()
+```
+
+마지막으로 이미지 임베딩과 depth 이미지를 [`KandinskyV22ControlnetPipeline`]에 전달하여 이미지를 생성합니다:
+
+```py
+image = pipeline(image_embeds=image_emb, negative_image_embeds=zero_image_emb, hint=hint, num_inference_steps=50, generator=generator, height=768, width=768).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/robot_cat_text2img.png"/>
+</div>
+
+### Image-to-image [[controlnet-image-to-image]]
+
+ControlNet을 사용한 image-to-image의 경우, 다음을 사용할 필요가 있습니다:
+
+- [`KandinskyV22PriorEmb2EmbPipeline`]로 텍스트 프롬프트와 이미지에서 이미지 임베딩을 생성합니다.
+- [`KandinskyV22ControlnetImg2ImgPipeline`]로 초기 이미지와 이미지 임베딩에서 이미지를 생성합니다.
+
+🤗 Transformers에서 `depth-estimation` [`~transformers.Pipeline`]을 사용하여 고양이의 초기 이미지의 depth map을 처리해 추출합니다:
+
+```py
+import torch
+import numpy as np
+
+from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22ControlnetImg2ImgPipeline
+from diffusers.utils import load_image
+from transformers import pipeline
+
+img = load_image(
+    "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/cat.png"
+).resize((768, 768))
+
+def make_hint(image, depth_estimator):
+    image = depth_estimator(image)["depth"]
+    image = np.array(image)
+    image = image[:, :, None]
+    image = np.concatenate([image, image, image], axis=2)
+    detected_map = torch.from_numpy(image).float() / 255.0
+    hint = detected_map.permute(2, 0, 1)
+    return hint
+
+depth_estimator = pipeline("depth-estimation")
+hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda")
+```
+
+Prior 파이프라인과 [`KandinskyV22ControlnetImg2ImgPipeline`]을 불러옵니다:
+
+```py
+prior_pipeline = KandinskyV22PriorEmb2EmbPipeline.from_pretrained(
+    "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16, use_safetensors=True
+).to("cuda")
+
+pipeline = KandinskyV22ControlnetImg2ImgPipeline.from_pretrained(
+    "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16
+).to("cuda")
+```
+
+텍스트 프롬프트와 초기 이미지를 이전 파이프라인에 전달하여 이미지 임베딩을 생성합니다:
+
+```py
+prompt = "A robot, 4k photo"
+negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature"
+
+generator = torch.Generator(device="cuda").manual_seed(43)
+
+img_emb = prior_pipeline(prompt=prompt, image=img, strength=0.85, generator=generator)
+negative_emb = prior_pipeline(prompt=negative_prior_prompt, image=img, strength=1, generator=generator)
+```
+
+이제 [`KandinskyV22ControlnetImg2ImgPipeline`]을 실행하여 초기 이미지와 이미지 임베딩으로부터 이미지를 생성할 수 있습니다:
+
+```py
+image = pipeline(image=img, strength=0.5, image_embeds=img_emb.image_embeds, negative_image_embeds=negative_emb.image_embeds, hint=hint, num_inference_steps=50, generator=generator, height=768, width=768).images[0]
+make_image_grid([img.resize((512, 512)), image.resize((512, 512))], rows=1, cols=2)
+```
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinskyv22/robot_cat.png"/>
+</div>
+
+## 최적화
+
+Kandinsky는 mapping을 생성하기 위한 prior 파이프라인과 latents를 이미지로 디코딩하기 위한 두 번째 파이프라인이 필요하다는 점에서 독특합니다. 대부분의 계산이 두 번째 파이프라인에서 이루어지므로 최적화의 노력은 두 번째 파이프라인에 집중되어야 합니다. 다음은 추론 중 Kandinsky키를 개선하기 위한 몇 가지 팁입니다.
+
+1. PyTorch < 2.0을 사용할 경우 [xFormers](../optimization/xformers)을 활성화합니다.
+
+```diff
+  from diffusers import DiffusionPipeline
+  import torch
+
+  pipe = DiffusionPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16)
++ pipe.enable_xformers_memory_efficient_attention()
+```
+
+2. PyTorch >= 2.0을 사용할 경우 `torch.compile`을 활성화하여 scaled dot-product attention (SDPA)를 자동으로 사용하도록 합니다:
+
+```diff
+  pipe.unet.to(memory_format=torch.channels_last)
++ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+```
+
+이는 attention processor를 명시적으로 [`~models.attention_processor.AttnAddedKVProcessor2_0`]을 사용하도록 설정하는 것과 동일합니다:
+
+```py
+from diffusers.models.attention_processor import AttnAddedKVProcessor2_0
+
+pipe.unet.set_attn_processor(AttnAddedKVProcessor2_0())
+```
+
+3. 메모리 부족 오류를 방지하기 위해 [`~KandinskyPriorPipeline.enable_model_cpu_offload`]를 사용하여 모델을 CPU로 오프로드합니다:
+
+```diff
+  from diffusers import DiffusionPipeline
+  import torch
+
+  pipe = DiffusionPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16)
++ pipe.enable_model_cpu_offload()
+```
+
+4. 기본적으로 text-to-image 파이프라인은 [`DDIMScheduler`]를 사용하지만, [`DDPMScheduler`]와 같은 다른 스케줄러로 대체하여 추론 속도와 이미지 품질 간의 균형에 어떤 영향을 미치는지 확인할 수 있습니다:
+
+```py
+from diffusers import DDPMScheduler
+from diffusers import DiffusionPipeline
+
+scheduler = DDPMScheduler.from_pretrained("kandinsky-community/kandinsky-2-1", subfolder="ddpm_scheduler")
+pipe = DiffusionPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", scheduler=scheduler, torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+```

docs/source/ko/using-diffusers/loading_adapters.md

@@ -0,0 +1,359 @@
+<!--Copyright 2024 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.
+-->
+
+# 어댑터 불러오기
+
+[[open-in-colab]]
+
+특정 물체의 이미지 또는 특정 스타일의 이미지를 생성하도록 diffusion 모델을 개인화하기 위한 몇 가지 [학습](../training/overview) 기법이 있습니다. 이러한 학습 방법은 각각 다른 유형의 어댑터를 생성합니다. 일부 어댑터는 완전히 새로운 모델을 생성하는 반면, 다른 어댑터는 임베딩 또는 가중치의 작은 부분만 수정합니다. 이는 각 어댑터의 로딩 프로세스도 다르다는 것을 의미합니다.
+
+이 가이드에서는 DreamBooth, textual inversion 및 LoRA 가중치를 불러오는 방법을 설명합니다.
+
+<Tip>
+
+사용할 체크포인트와 임베딩은 [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer), [LoRA the Explorer](https://huggingface.co/spaces/multimodalart/LoraTheExplorer), [Diffusers Models Gallery](https://huggingface.co/spaces/huggingface-projects/diffusers-gallery)에서 찾아보시기 바랍니다.
+
+</Tip>
+
+## DreamBooth
+
+[DreamBooth](https://dreambooth.github.io/)는 물체의 여러 이미지에 대한 *diffusion 모델 전체*를 미세 조정하여 새로운 스타일과 설정으로 해당 물체의 이미지를 생성합니다. 이 방법은 모델이 물체 이미지와 연관시키는 방법을 학습하는 프롬프트에 특수 단어를 사용하는 방식으로 작동합니다. 모든 학습 방법 중에서 드림부스는 전체 체크포인트 모델이기 때문에 파일 크기가 가장 큽니다(보통 몇 GB).
+
+Hergé가 그린 단 10개의 이미지로 학습된 [herge_style](https://huggingface.co/sd-dreambooth-library/herge-style) 체크포인트를 불러와 해당 스타일의 이미지를 생성해 보겠습니다. 이 모델이 작동하려면 체크포인트를 트리거하는 프롬프트에 특수 단어 `herge_style`을 포함시켜야 합니다:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("sd-dreambooth-library/herge-style", torch_dtype=torch.float16).to("cuda")
+prompt = "A cute herge_style brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_dreambooth.png" />
+</div>
+
+## Textual inversion
+
+[Textual inversion](https://textual-inversion.github.io/)은 DreamBooth와 매우 유사하며 몇 개의 이미지만으로 특정 개념(스타일, 개체)을 생성하는 diffusion 모델을 개인화할 수도 있습니다. 이 방법은 프롬프트에 특정 단어를 입력하면 해당 이미지를 나타내는 새로운 임베딩을 학습하고 찾아내는 방식으로 작동합니다. 결과적으로 diffusion 모델 가중치는 동일하게 유지되고 훈련 프로세스는 비교적 작은(수 KB) 파일을 생성합니다.
+
+Textual inversion은 임베딩을 생성하기 때문에 DreamBooth처럼 단독으로 사용할 수 없으며 또 다른 모델이 필요합니다.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
+```
+
+이제 [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] 메서드를 사용하여 textual inversion 임베딩을 불러와 이미지를 생성할 수 있습니다. [sd-concepts-library/gta5-artwork](https://huggingface.co/sd-concepts-library/gta5-artwork) 임베딩을 불러와 보겠습니다. 이를 트리거하려면 프롬프트에 특수 단어 `<gta5-artwork>`를 포함시켜야 합니다:
+
+```py
+pipeline.load_textual_inversion("sd-concepts-library/gta5-artwork")
+prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, <gta5-artwork> style"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_txt_embed.png" />
+</div>
+
+Textual inversion은 또한 바람직하지 않은 사물에 대해 *네거티브 임베딩*을 생성하여 모델이 흐릿한 이미지나 손의 추가 손가락과 같은 바람직하지 않은 사물이 포함된 이미지를 생성하지 못하도록 학습할 수도 있습니다. 이는 프롬프트를 빠르게 개선하는 것이 쉬운 방법이 될 수 있습니다. 이는 이전과 같이 임베딩을 [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`]으로 불러오지만 이번에는 두 개의 매개변수가 더 필요합니다:
+
+- `weight_name`: 파일이 특정 이름의 🤗 Diffusers 형식으로 저장된 경우이거나 파일이 A1111 형식으로 저장된 경우, 불러올 가중치 파일을 지정합니다.
+- `token`: 임베딩을 트리거하기 위해 프롬프트에서 사용할 특수 단어를 지정합니다.
+
+[sayakpaul/EasyNegative-test](https://huggingface.co/sayakpaul/EasyNegative-test) 임베딩을 불러와 보겠습니다:
+
+```py
+pipeline.load_textual_inversion(
+    "sayakpaul/EasyNegative-test", weight_name="EasyNegative.safetensors", token="EasyNegative"
+)
+```
+
+이제 `token`을 사용해 네거티브 임베딩이 있는 이미지를 생성할 수 있습니다:
+
+```py
+prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, EasyNegative"
+negative_prompt = "EasyNegative"
+
+image = pipeline(prompt, negative_prompt=negative_prompt, num_inference_steps=50).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png" />
+</div>
+
+## LoRA
+
+[Low-Rank Adaptation (LoRA)](https://huggingface.co/papers/2106.09685)은 속도가 빠르고 파일 크기가 (수백 MB로) 작기 때문에 널리 사용되는 학습 기법입니다. 이 가이드의 다른 방법과 마찬가지로, LoRA는 몇 장의 이미지만으로 새로운 스타일을 학습하도록 모델을 학습시킬 수 있습니다. 이는 diffusion 모델에 새로운 가중치를 삽입한 다음 전체 모델 대신 새로운 가중치만 학습시키는 방식으로 작동합니다. 따라서 LoRA를 더 빠르게 학습시키고 더 쉽게 저장할 수 있습니다.
+
+<Tip>
+
+LoRA는 다른 학습 방법과 함께 사용할 수 있는 매우 일반적인 학습 기법입니다. 예를 들어, DreamBooth와 LoRA로 모델을 학습하는 것이 일반적입니다. 또한 새롭고 고유한 이미지를 생성하기 위해 여러 개의 LoRA를 불러오고 병합하는 것이 점점 더 일반화되고 있습니다. 병합은 이 불러오기 가이드의 범위를 벗어나므로 자세한 내용은 심층적인 [LoRA 병합](merge_loras) 가이드에서 확인할 수 있습니다.
+
+</Tip>
+
+LoRA는 다른 모델과 함께 사용해야 합니다:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+```
+
+그리고 [`~loaders.LoraLoaderMixin.load_lora_weights`] 메서드를 사용하여 [ostris/super-cereal-sdxl-lora](https://huggingface.co/ostris/super-cereal-sdxl-lora) 가중치를 불러오고 리포지토리에서 가중치 파일명을 지정합니다:
+
+```py
+pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora", weight_name="cereal_box_sdxl_v1.safetensors")
+prompt = "bears, pizza bites"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_lora.png" />
+</div>
+
+[`~loaders.LoraLoaderMixin.load_lora_weights`] 메서드는 LoRA 가중치를 UNet과 텍스트 인코더에 모두 불러옵니다. 이 메서드는 해당 케이스에서 LoRA를 불러오는 데 선호되는 방식입니다:
+
+- LoRA 가중치에 UNet 및 텍스트 인코더에 대한 별도의 식별자가 없는 경우
+- LoRA 가중치에 UNet과 텍스트 인코더에 대한 별도의 식별자가 있는 경우
+
+하지만 LoRA 가중치만 UNet에 로드해야 하는 경우에는 [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] 메서드를 사용할 수 있습니다. [jbilcke-hf/sdxl-cinematic-1](https://huggingface.co/jbilcke-hf/sdxl-cinematic-1) LoRA를 불러와 보겠습니다:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.unet.load_attn_procs("jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors")
+
+# 프롬프트에서 cnmt를 사용하여 LoRA를 트리거합니다.
+prompt = "A cute cnmt eating a slice of pizza, stunning color scheme, masterpiece, illustration"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
+</div>
+
+LoRA 가중치를 언로드하려면 [`~loaders.LoraLoaderMixin.unload_lora_weights`] 메서드를 사용하여 LoRA 가중치를 삭제하고 모델을 원래 가중치로 복원합니다:
+
+```py
+pipeline.unload_lora_weights()
+```
+
+### LoRA 가중치 스케일 조정하기
+
+[`~loaders.LoraLoaderMixin.load_lora_weights`] 및 [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] 모두 `cross_attention_kwargs={"scale": 0.5}` 파라미터를 전달하여 얼마나 LoRA 가중치를 사용할지 조정할 수 있습니다. 값이 `0`이면 기본 모델 가중치만 사용하는 것과 같고, 값이 `1`이면 완전히 미세 조정된 LoRA를 사용하는 것과 같습니다.
+
+레이어당 사용되는 LoRA 가중치의 양을 보다 세밀하게 제어하려면 [`~loaders.LoraLoaderMixin.set_adapters`]를 사용하여 각 레이어의 가중치를 얼마만큼 조정할지 지정하는 딕셔너리를 전달할 수 있습니다.
+```python
+pipe = ... # 파이프라인 생성
+pipe.load_lora_weights(..., adapter_name="my_adapter")
+scales = {
+    "text_encoder": 0.5,
+    "text_encoder_2": 0.5,  # 파이프에 두 번째 텍스트 인코더가 있는 경우에만 사용 가능
+    "unet": {
+        "down": 0.9,  # down 부분의 모든 트랜스포머는 스케일 0.9를 사용
+        # "mid"  # 이 예제에서는 "mid"가 지정되지 않았으므로 중간 부분의 모든 트랜스포머는 기본 스케일 1.0을 사용
+        "up": {
+            "block_0": 0.6,  # # up의 0번째 블록에 있는 3개의 트랜스포머는 모두 스케일 0.6을 사용
+            "block_1": [0.4, 0.8, 1.0],  # up의 첫 번째 블록에 있는 3개의 트랜스포머는 각각 스케일 0.4, 0.8, 1.0을 사용
+        }
+    }
+}
+pipe.set_adapters("my_adapter", scales)
+```
+
+이는 여러 어댑터에서도 작동합니다. 방법은 [이 가이드](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference#customize-adapters-strength)를 참조하세요.
+
+<Tip warning={true}>
+
+현재 [`~loaders.LoraLoaderMixin.set_adapters`]는 어텐션 가중치의 스케일링만 지원합니다. LoRA에 다른 부분(예: resnets or down-/upsamplers)이 있는 경우 1.0의 스케일을 유지합니다.
+
+</Tip>
+
+### Kohya와 TheLastBen
+
+커뮤니티에서 인기 있는 다른 LoRA trainer로는 [Kohya](https://github.com/kohya-ss/sd-scripts/)와 [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion)의 trainer가 있습니다. 이 trainer들은 🤗 Diffusers가 훈련한 것과는 다른 LoRA 체크포인트를 생성하지만, 같은 방식으로 불러올 수 있습니다.
+
+<hfoptions id="other-trainers">
+<hfoption id="Kohya">
+
+Kohya LoRA를 불러오기 위해, 예시로 [Civitai](https://civitai.com/)에서 [Blueprintify SD XL 1.0](https://civitai.com/models/150986/blueprintify-sd-xl-10) 체크포인트를 다운로드합니다:
+
+```sh
+!wget https://civitai.com/api/download/models/168776 -O blueprintify-sd-xl-10.safetensors
+```
+
+LoRA 체크포인트를 [`~loaders.LoraLoaderMixin.load_lora_weights`] 메서드로 불러오고 `weight_name` 파라미터에 파일명을 지정합니다:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("path/to/weights", weight_name="blueprintify-sd-xl-10.safetensors")
+```
+
+이미지를 생성합니다:
+
+```py
+# LoRA를 트리거하기 위해 bl3uprint를 프롬프트에 사용
+prompt = "bl3uprint, a highly detailed blueprint of the eiffel tower, explaining how to build all parts, many txt, blueprint grid backdrop"
+image = pipeline(prompt).images[0]
+image
+```
+
+<Tip warning={true}>
+
+Kohya LoRA를 🤗 Diffusers와 함께 사용할 때 몇 가지 제한 사항이 있습니다:
+
+- [여기](https://github.com/huggingface/diffusers/pull/4287/#issuecomment-1655110736)에 설명된 여러 가지 이유로 인해 이미지가 ComfyUI와 같은 UI에서 생성된 이미지와 다르게 보일 수 있습니다.
+- [LyCORIS 체크포인트](https://github.com/KohakuBlueleaf/LyCORIS)가 완전히 지원되지 않습니다. [`~loaders.LoraLoaderMixin.load_lora_weights`] 메서드는 LoRA 및 LoCon 모듈로 LyCORIS 체크포인트를 불러올 수 있지만, Hada 및 LoKR은 지원되지 않습니다.
+
+</Tip>
+
+</hfoption>
+<hfoption id="TheLastBen">
+
+TheLastBen에서 체크포인트를 불러오는 방법은 매우 유사합니다. 예를 들어, [TheLastBen/William_Eggleston_Style_SDXL](https://huggingface.co/TheLastBen/William_Eggleston_Style_SDXL) 체크포인트를 불러오려면:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("TheLastBen/William_Eggleston_Style_SDXL", weight_name="wegg.safetensors")
+
+# LoRA를 트리거하기 위해 william eggleston를 프롬프트에 사용
+prompt = "a house by william eggleston, sunrays, beautiful, sunlight, sunrays, beautiful"
+image = pipeline(prompt=prompt).images[0]
+image
+```
+
+</hfoption>
+</hfoptions>
+
+## IP-Adapter
+
+[IP-Adapter](https://ip-adapter.github.io/)는 모든 diffusion 모델에 이미지 프롬프트를 사용할 수 있는 경량 어댑터입니다. 이 어댑터는 이미지와 텍스트 feature의 cross-attention 레이어를 분리하여 작동합니다. 다른 모든 모델 컴포넌트튼 freeze되고 UNet의 embedded 이미지 features만 학습됩니다. 따라서 IP-Adapter 파일은 일반적으로 최대 100MB에 불과합니다.
+
+다양한 작업과 구체적인 사용 사례에 IP-Adapter를 사용하는 방법에 대한 자세한 내용은 [IP-Adapter](../using-diffusers/ip_adapter) 가이드에서 확인할 수 있습니다.
+
+> [!TIP]
+> Diffusers는 현재 가장 많이 사용되는 일부 파이프라인에 대해서만 IP-Adapter를 지원합니다. 멋진 사용 사례가 있는 지원되지 않는 파이프라인에 IP-Adapter를 통합하고 싶다면 언제든지 기능 요청을 여세요!
+> 공식 IP-Adapter 체크포인트는 [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter)에서 확인할 수 있습니다.
+
+시작하려면 Stable Diffusion 체크포인트를 불러오세요.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+from diffusers.utils import load_image
+
+pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
+```
+
+그런 다음 IP-Adapter 가중치를 불러와 [`~loaders.IPAdapterMixin.load_ip_adapter`] 메서드를 사용하여 파이프라인에 추가합니다.
+
+```py
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
+```
+
+불러온 뒤, 이미지 및 텍스트 프롬프트가 있는 파이프라인을 사용하여 이미지 생성 프로세스를 가이드할 수 있습니다.
+
+```py
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png")
+generator = torch.Generator(device="cpu").manual_seed(33)
+images = pipeline(
+    prompt='best quality, high quality, wearing sunglasses',
+    ip_adapter_image=image,
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+    num_inference_steps=50,
+    generator=generator,
+).images[0]
+images
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ip-bear.png" />
+</div>
+
+### IP-Adapter Plus
+
+IP-Adapter는 이미지 인코더를 사용하여 이미지 feature를 생성합니다. IP-Adapter 리포지토리에 `image_encoder` 하위 폴더가 있는 경우, 이미지 인코더가 자동으로 불러와 파이프라인에 등록됩니다. 그렇지 않은 경우, [`~transformers.CLIPVisionModelWithProjection`] 모델을 사용하여 이미지 인코더를 명시적으로 불러와 파이프라인에 전달해야 합니다.
+
+이는 ViT-H 이미지 인코더를 사용하는 *IP-Adapter Plus* 체크포인트에 해당하는 케이스입니다.
+
+```py
+from transformers import CLIPVisionModelWithProjection
+
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+    "h94/IP-Adapter",
+    subfolder="models/image_encoder",
+    torch_dtype=torch.float16
+)
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    image_encoder=image_encoder,
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter-plus_sdxl_vit-h.safetensors")
+```
+
+### IP-Adapter Face ID 모델
+
+IP-Adapter FaceID 모델은 CLIP 이미지 임베딩 대신 `insightface`에서 생성한 이미지 임베딩을 사용하는 실험적인 IP Adapter입니다. 이러한 모델 중 일부는 LoRA를 사용하여 ID 일관성을 개선하기도 합니다.
+이러한 모델을 사용하려면 `insightface`와 해당 요구 사항을 모두 설치해야 합니다.
+
+<Tip warning={true}>
+InsightFace 사전학습된 모델은 비상업적 연구 목적으로만 사용할 수 있으므로, IP-Adapter-FaceID 모델은 연구 목적으로만 릴리즈되었으며 상업적 용도로는 사용할 수 없습니다.
+</Tip>
+
+```py
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter-FaceID", subfolder=None, weight_name="ip-adapter-faceid_sdxl.bin", image_encoder_folder=None)
+```
+
+두 가지 IP 어댑터 FaceID Plus 모델 중 하나를 사용하려는 경우, 이 모델들은 더 나은 사실감을 얻기 위해 `insightface`와 CLIP 이미지 임베딩을 모두 사용하므로, CLIP 이미지 인코더도 불러와야 합니다.
+
+```py
+from transformers import CLIPVisionModelWithProjection
+
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+    "laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
+    torch_dtype=torch.float16,
+)
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    image_encoder=image_encoder,
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter-FaceID", subfolder=None, weight_name="ip-adapter-faceid-plus_sd15.bin")
+```

docs/source/ko/using-diffusers/loading_overview.md

@@ -1,18 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# Overview
-
-🧨 Diffusers는 생성 작업을 위한 다양한 파이프라인, 모델, 스케줄러를 제공합니다. 이러한 컴포넌트를 최대한 간단하게 로드할 수 있도록 단일 통합 메서드인 `from_pretrained()`를 제공하여 Hugging Face [Hub](https://huggingface.co/models?library=diffusers&sort=downloads) 또는 로컬 머신에서 이러한 컴포넌트를 불러올 수 있습니다. 파이프라인이나 모델을 로드할 때마다, 최신 파일이 자동으로 다운로드되고 캐시되므로, 다음에 파일을 다시 다운로드하지 않고도 빠르게 재사용할 수 있습니다.
-
-이 섹션은 파이프라인 로딩, 파이프라인에서 다양한 컴포넌트를 로드하는 방법, 체크포인트 variants를 불러오는 방법, 그리고 커뮤니티 파이프라인을 불러오는 방법에 대해 알아야 할 모든 것들을 다룹니다. 또한 스케줄러를 불러오는 방법과 서로 다른 스케줄러를 사용할 때 발생하는 속도와 품질간의 트레이드 오프를 비교하는 방법 역시 다룹니다. 그리고 마지막으로 🧨 Diffusers와 함께 파이토치에서 사용할 수 있도록 KerasCV 체크포인트를 변환하고 불러오는 방법을 살펴봅니다.
-

docs/source/ko/using-diffusers/pipeline_overview.md

@@ -1,17 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# Overview
-
-파이프라인은 독립적으로 훈련된 모델과 스케줄러를 함께 모아서 추론을 위해 diffusion 시스템을 빠르고 쉽게 사용할 수 있는 방법을 제공하는 end-to-end 클래스입니다. 모델과 스케줄러의 특정 조합은 특수한 기능과 함께 [`StableDiffusionPipeline`] 또는 [`StableDiffusionControlNetPipeline`]과 같은 특정 파이프라인 유형을 정의합니다. 모든 파이프라인 유형은 기본 [`DiffusionPipeline`] 클래스에서 상속됩니다. 어느 체크포인트를 전달하면, 파이프라인 유형을 자동으로 감지하고 필요한 구성 요소들을 불러옵니다.
-
-이 섹션에서는 unconditional 이미지 생성, text-to-image 생성의 다양한 테크닉과 변화를 파이프라인에서 지원하는 작업들을 소개합니다. 프롬프트에 있는 특정 단어가 출력에 영향을 미치는 것을 조정하기 위해 재현성을 위한 시드 설정과 프롬프트에 가중치를 부여하는 것으로 생성 프로세스를 더 잘 제어하는 방법에 대해 배울 수 있습니다. 마지막으로 음성에서부터 이미지 생성과 같은 커스텀 작업을 위한 커뮤니티 파이프라인을 만드는 방법을 알 수 있습니다.

docs/source/ko/using-diffusers/push_to_hub.md

@@ -0,0 +1,177 @@
+<!--Copyright 2024 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.
+-->
+
+# 파일들을 Hub로 푸시하기
+
+[[open-in-colab]]
+
+🤗 Diffusers는 모델, 스케줄러 또는 파이프라인을 Hub에 업로드할 수 있는 [`~diffusers.utils.PushToHubMixin`]을 제공합니다. 이는 Hub에 당신의 파일을 저장하는 쉬운 방법이며, 다른 사람들과 작업을 공유할 수도 있습니다. 실제적으로 [`~diffusers.utils.PushToHubMixin`]가 동작하는 방식은 다음과 같습니다:
+
+1. Hub에 리포지토리를 생성합니다.
+2. 나중에 다시 불러올 수 있도록 모델, 스케줄러 또는 파이프라인 파일을 저장합니다.
+3. 이러한 파일이 포함된 폴더를 Hub에 업로드합니다.
+
+이 가이드는 [`~diffusers.utils.PushToHubMixin`]을 사용하여 Hub에 파일을 업로드하는 방법을 보여줍니다.
+
+먼저 액세스 [토큰](https://huggingface.co/settings/tokens)으로 Hub 계정에 로그인해야 합니다:
+
+```py
+from huggingface_hub import notebook_login
+
+notebook_login()
+```
+
+## 모델
+
+모델을 허브에 푸시하려면 [`~diffusers.utils.PushToHubMixin.push_to_hub`]를 호출하고 Hub에 저장할 모델의 리포지토리 id를 지정합니다:
+
+```py
+from diffusers import ControlNetModel
+
+controlnet = ControlNetModel(
+    block_out_channels=(32, 64),
+    layers_per_block=2,
+    in_channels=4,
+    down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+    cross_attention_dim=32,
+    conditioning_embedding_out_channels=(16, 32),
+)
+controlnet.push_to_hub("my-controlnet-model")
+```
+
+모델의 경우 Hub에 푸시할 가중치의 [*변형*](loading#checkpoint-variants)을 지정할 수도 있습니다. 예를 들어, `fp16` 가중치를 푸시하려면 다음과 같이 하세요:
+
+```py
+controlnet.push_to_hub("my-controlnet-model", variant="fp16")
+```
+
+[`~diffusers.utils.PushToHubMixin.push_to_hub`] 함수는 모델의 `config.json` 파일을 저장하고 가중치는 `safetensors` 형식으로 자동으로 저장됩니다.
+
+이제 Hub의 리포지토리에서 모델을 다시 불러올 수 있습니다:
+
+```py
+model = ControlNetModel.from_pretrained("your-namespace/my-controlnet-model")
+```
+
+## 스케줄러
+
+스케줄러를 허브에 푸시하려면 [`~diffusers.utils.PushToHubMixin.push_to_hub`]를 호출하고 Hub에 저장할 스케줄러의 리포지토리 id를 지정합니다:
+
+```py
+from diffusers import DDIMScheduler
+
+scheduler = DDIMScheduler(
+    beta_start=0.00085,
+    beta_end=0.012,
+    beta_schedule="scaled_linear",
+    clip_sample=False,
+    set_alpha_to_one=False,
+)
+scheduler.push_to_hub("my-controlnet-scheduler")
+```
+
+[`~diffusers.utils.PushToHubMixin.push_to_hub`] 함수는 스케줄러의 `scheduler_config.json` 파일을 지정된 리포지토리에 저장합니다.
+
+이제 허브의 리포지토리에서 스케줄러를 다시 불러올 수 있습니다:
+
+```py
+scheduler = DDIMScheduler.from_pretrained("your-namepsace/my-controlnet-scheduler")
+```
+
+## 파이프라인
+
+모든 컴포넌트가 포함된 전체 파이프라인을 Hub로 푸시할 수도 있습니다. 예를 들어, 원하는 파라미터로 [`StableDiffusionPipeline`]의 컴포넌트들을 초기화합니다:
+
+```py
+from diffusers import (
+    UNet2DConditionModel,
+    AutoencoderKL,
+    DDIMScheduler,
+    StableDiffusionPipeline,
+)
+from transformers import CLIPTextModel, CLIPTextConfig, CLIPTokenizer
+
+unet = UNet2DConditionModel(
+    block_out_channels=(32, 64),
+    layers_per_block=2,
+    sample_size=32,
+    in_channels=4,
+    out_channels=4,
+    down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+    up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+    cross_attention_dim=32,
+)
+
+scheduler = DDIMScheduler(
+    beta_start=0.00085,
+    beta_end=0.012,
+    beta_schedule="scaled_linear",
+    clip_sample=False,
+    set_alpha_to_one=False,
+)
+
+vae = AutoencoderKL(
+    block_out_channels=[32, 64],
+    in_channels=3,
+    out_channels=3,
+    down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+    up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+    latent_channels=4,
+)
+
+text_encoder_config = CLIPTextConfig(
+    bos_token_id=0,
+    eos_token_id=2,
+    hidden_size=32,
+    intermediate_size=37,
+    layer_norm_eps=1e-05,
+    num_attention_heads=4,
+    num_hidden_layers=5,
+    pad_token_id=1,
+    vocab_size=1000,
+)
+text_encoder = CLIPTextModel(text_encoder_config)
+tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+```
+
+모든 컴포넌트들을 [`StableDiffusionPipeline`]에 전달하고 [`~diffusers.utils.PushToHubMixin.push_to_hub`]를 호출하여 파이프라인을 Hub로 푸시합니다:
+
+```py
+components = {
+    "unet": unet,
+    "scheduler": scheduler,
+    "vae": vae,
+    "text_encoder": text_encoder,
+    "tokenizer": tokenizer,
+    "safety_checker": None,
+    "feature_extractor": None,
+}
+
+pipeline = StableDiffusionPipeline(**components)
+pipeline.push_to_hub("my-pipeline")
+```
+
+[`~diffusers.utils.PushToHubMixin.push_to_hub`] 함수는 각 컴포넌트를 리포지토리의 하위 폴더에 저장합니다. 이제 Hub의 리포지토리에서 파이프라인을 다시 불러올 수 있습니다:
+
+```py
+pipeline = StableDiffusionPipeline.from_pretrained("your-namespace/my-pipeline")
+```
+
+## 비공개
+
+모델, 스케줄러 또는 파이프라인 파일들을 비공개로 두려면 [`~diffusers.utils.PushToHubMixin.push_to_hub`] 함수에서 `private=True`를 설정하세요:
+
+```py
+controlnet.push_to_hub("my-controlnet-model-private", private=True)
+```
+
+비공개 리포지토리는 본인만 볼 수 있으며 다른 사용자는 리포지토리를 복제할 수 없고 리포지토리가 검색 결과에 표시되지 않습니다. 사용자가 비공개 리포지토리의 URL을 가지고 있더라도 `404 - Sorry, we can't find the page you are looking for`라는 메시지가 표시됩니다. 비공개 리포지토리에서 모델을 로드하려면 [로그인](https://huggingface.co/docs/huggingface_hub/quick-start#login) 상태여야 합니다.

docs/source/ko/using-diffusers/reproducibility.md

@@ -1,201 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# 재현 가능한 파이프라인 생성하기
-
-[[open-in-colab]]
-
-재현성은 테스트, 결과 재현, 그리고 [이미지 퀄리티 높이기](resuing_seeds)에서 중요합니다.
-그러나 diffusion 모델의 무작위성은 매번 모델이 돌아갈 때마다 파이프라인이 다른 이미지를 생성할 수 있도록 하는 이유로 필요합니다.
-플랫폼 간에 정확하게 동일한 결과를 얻을 수는 없지만, 특정 허용 범위 내에서 릴리스 및 플랫폼 간에 결과를 재현할 수는 있습니다.
-그럼에도 diffusion 파이프라인과 체크포인트에 따라 허용 오차가 달라집니다.
-
-diffusion 모델에서 무작위성의 원천을 제어하거나 결정론적 알고리즘을 사용하는 방법을 이해하는 것이 중요한 이유입니다.
-
-<Tip>
-
-💡 Pytorch의 [재현성에 대한 선언](https://pytorch.org/docs/stable/notes/randomness.html)를 꼭 읽어보길 추천합니다:
-
-> 완전하게 재현가능한 결과는 Pytorch 배포, 개별적인 커밋, 혹은 다른 플랫폼들에서 보장되지 않습니다.
-> 또한, 결과는 CPU와 GPU 실행간에 심지어 같은 seed를 사용할 때도 재현 가능하지 않을 수 있습니다.
-
-</Tip>
-
-## 무작위성 제어하기
-
-추론에서, 파이프라인은 노이즈를 줄이기 위해 가우시안 노이즈를 생성하거나 스케줄링 단계에 노이즈를 더하는 등의 랜덤 샘플링 실행에 크게 의존합니다,
-
-[DDIMPipeline](https://huggingface.co/docs/diffusers/v0.18.0/en/api/pipelines/ddim#diffusers.DDIMPipeline)에서 두 추론 단계 이후의 텐서 값을 살펴보세요:
-
-```python
-from diffusers import DDIMPipeline
-import numpy as np
-
-model_id = "google/ddpm-cifar10-32"
-
-# 모델과 스케줄러를 불러오기
-ddim = DDIMPipeline.from_pretrained(model_id)
-
-# 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기
-image = ddim(num_inference_steps=2, output_type="np").images
-print(np.abs(image).sum())
-```
-
-위의 코드를 실행하면 하나의 값이 나오지만, 다시 실행하면 다른 값이 나옵니다. 무슨 일이 일어나고 있는 걸까요?
-
-파이프라인이 실행될 때마다, [torch.randn](https://pytorch.org/docs/stable/generated/torch.randn.html)은
-단계적으로 노이즈 제거되는 가우시안 노이즈가 생성하기 위한 다른 랜덤 seed를 사용합니다.
-
-그러나 동일한 이미지를 안정적으로 생성해야 하는 경우에는 CPU에서 파이프라인을 실행하는지 GPU에서 실행하는지에 따라 달라집니다.
-
-### CPU
-
-CPU에서 재현 가능한 결과를 생성하려면, PyTorch [Generator](https://pytorch.org/docs/stable/generated/torch.randn.html)로 seed를 고정합니다:
-
-```python
-import torch
-from diffusers import DDIMPipeline
-import numpy as np
-
-model_id = "google/ddpm-cifar10-32"
-
-# 모델과 스케줄러 불러오기
-ddim = DDIMPipeline.from_pretrained(model_id)
-
-# 재현성을 위해 generator 만들기
-generator = torch.Generator(device="cpu").manual_seed(0)
-
-# 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기
-image = ddim(num_inference_steps=2, output_type="np", generator=generator).images
-print(np.abs(image).sum())
-```
-
-이제 위의 코드를 실행하면 seed를 가진 `Generator` 객체가 파이프라인의 모든 랜덤 함수에 전달되므로 항상 `1491.1711` 값이 출력됩니다.
-
-특정 하드웨어 및 PyTorch 버전에서 이 코드 예제를 실행하면 동일하지는 않더라도 유사한 결과를 얻을 수 있습니다.
-
-<Tip>
-
-💡 처음에는 시드를 나타내는 정수값 대신에 `Generator` 개체를 파이프라인에 전달하는 것이 약간 비직관적일 수 있지만,
-`Generator`는 순차적으로 여러 파이프라인에 전달될 수 있는 \랜덤상태\이기 때문에 PyTorch에서 확률론적 모델을 다룰 때 권장되는 설계입니다.
-
-</Tip>
-
-### GPU
-
-예를 들면, GPU 상에서 같은 코드 예시를 실행하면:
-
-```python
-import torch
-from diffusers import DDIMPipeline
-import numpy as np
-
-model_id = "google/ddpm-cifar10-32"
-
-# 모델과 스케줄러 불러오기
-ddim = DDIMPipeline.from_pretrained(model_id)
-ddim.to("cuda")
-
-# 재현성을 위한 generator 만들기
-generator = torch.Generator(device="cuda").manual_seed(0)
-
-# 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기
-image = ddim(num_inference_steps=2, output_type="np", generator=generator).images
-print(np.abs(image).sum())
-```
-
-GPU가 CPU와 다른 난수 생성기를 사용하기 때문에 동일한 시드를 사용하더라도 결과가 같지 않습니다.
-
-이 문제를 피하기 위해 🧨 Diffusers는 CPU에 임의의 노이즈를 생성한 다음 필요에 따라 텐서를 GPU로 이동시키는
-[randn_tensor()](https://huggingface.co/docs/diffusers/v0.18.0/en/api/utilities#diffusers.utils.randn_tensor)기능을 가지고 있습니다.
-`randn_tensor` 기능은 파이프라인 내부 어디에서나 사용되므로 파이프라인이 GPU에서 실행되더라도 **항상** CPU `Generator`를 통과할 수 있습니다.
-
-이제 결과에 훨씬 더 다가왔습니다!
-
-```python
-import torch
-from diffusers import DDIMPipeline
-import numpy as np
-
-model_id = "google/ddpm-cifar10-32"
-
-# 모델과 스케줄러 불러오기
-ddim = DDIMPipeline.from_pretrained(model_id)
-ddim.to("cuda")
-
-#재현성을 위한 generator 만들기 (GPU에 올리지 않도록 조심한다!)
-generator = torch.manual_seed(0)
-
-# 두 개의 단계에 대해서 파이프라인을 실행하고 numpy tensor로 값을 반환하기
-image = ddim(num_inference_steps=2, output_type="np", generator=generator).images
-print(np.abs(image).sum())
-```
-
-<Tip>
-
-💡 재현성이 중요한 경우에는 항상 CPU generator를 전달하는 것이 좋습니다.
-성능 손실은 무시할 수 없는 경우가 많으며 파이프라인이 GPU에서 실행되었을 때보다 훨씬 더 비슷한 값을 생성할 수 있습니다.
-
-</Tip>
-
-마지막으로 [UnCLIPPipeline](https://huggingface.co/docs/diffusers/v0.18.0/en/api/pipelines/unclip#diffusers.UnCLIPPipeline)과 같은
-더 복잡한 파이프라인의 경우, 이들은 종종 정밀 오차 전파에 극도로 취약합니다.
-다른 GPU 하드웨어 또는 PyTorch 버전에서 유사한 결과를 기대하지 마세요.
-이 경우 완전한 재현성을 위해 완전히 동일한 하드웨어 및 PyTorch 버전을 실행해야 합니다.
-
-## 결정론적 알고리즘
-
-결정론적 알고리즘을 사용하여 재현 가능한 파이프라인을 생성하도록 PyTorch를 구성할 수도 있습니다.
-그러나 결정론적 알고리즘은 비결정론적 알고리즘보다 느리고 성능이 저하될 수 있습니다.
-하지만 재현성이 중요하다면, 이것이 최선의 방법입니다!
-
-둘 이상의 CUDA 스트림에서 작업이 시작될 때 비결정론적 동작이 발생합니다.
-이 문제를 방지하려면 환경 변수 [CUBLAS_WORKSPACE_CONFIG](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility)를 `:16:8`로 설정해서
-런타임 중에 오직 하나의 버퍼 크리만 사용하도록 설정합니다.
-
-PyTorch는 일반적으로 가장 빠른 알고리즘을 선택하기 위해 여러 알고리즘을 벤치마킹합니다.
-하지만 재현성을 원하는 경우, 벤치마크가 매 순간 다른 알고리즘을 선택할 수 있기 때문에 이 기능을 사용하지 않도록 설정해야 합니다.
-마지막으로, [torch.use_deterministic_algorithms](https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html)에
-`True`를 통과시켜 결정론적 알고리즘이 활성화 되도록 합니다.
-
-```py
-import os
-
-os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8"
-
-torch.backends.cudnn.benchmark = False
-torch.use_deterministic_algorithms(True)
-```
-
-이제 동일한 파이프라인을 두번 실행하면 동일한 결과를 얻을 수 있습니다.
-
-```py
-import torch
-from diffusers import DDIMScheduler, StableDiffusionPipeline
-import numpy as np
-
-model_id = "runwayml/stable-diffusion-v1-5"
-pipe = StableDiffusionPipeline.from_pretrained(model_id).to("cuda")
-pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
-g = torch.Generator(device="cuda")
-
-prompt = "A bear is playing a guitar on Times Square"
-
-g.manual_seed(0)
-result1 = pipe(prompt=prompt, num_inference_steps=50, generator=g, output_type="latent").images
-
-g.manual_seed(0)
-result2 = pipe(prompt=prompt, num_inference_steps=50, generator=g, output_type="latent").images
-
-print("L_inf dist = ", abs(result1 - result2).max())
-"L_inf dist =  tensor(0., device='cuda:0')"
-```

docs/source/ko/using-diffusers/reusing_seeds.md

@@ -1,63 +0,0 @@
-<!--Copyright 2024 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.
--->
-
-# Deterministic(결정적) 생성을 통한 이미지 품질 개선
-
-생성된 이미지의 품질을 개선하는 일반적인 방법은 *결정적 batch(배치) 생성*을 사용하는 것입니다. 이 방법은 이미지 batch(배치)를 생성하고 두 번째 추론 라운드에서 더 자세한 프롬프트와 함께 개선할 이미지 하나를 선택하는 것입니다. 핵심은 일괄 이미지 생성을 위해 파이프라인에 [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html#generator) 목록을 전달하고, 각 `Generator`를 시드에 연결하여 이미지에 재사용할 수 있도록 하는 것입니다.
-
-예를 들어 [`runwayml/stable-diffusion-v1-5`](runwayml/stable-diffusion-v1-5)를 사용하여 다음 프롬프트의 여러 버전을 생성해 봅시다.
-
-```py
-prompt = "Labrador in the style of Vermeer"
-```
-
-(가능하다면) 파이프라인을 [`DiffusionPipeline.from_pretrained`]로 인스턴스화하여 GPU에 배치합니다.
-
-```python
->>> from diffusers import DiffusionPipeline
-
->>> pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
->>> pipe = pipe.to("cuda")
-```
-
-이제 네 개의 서로 다른 `Generator`를 정의하고 각 `Generator`에 시드(`0` ~ `3`)를 할당하여 나중에 특정 이미지에 대해 `Generator`를 재사용할 수 있도록 합니다.
-
-```python
->>> import torch
-
->>> generator = [torch.Generator(device="cuda").manual_seed(i) for i in range(4)]
-```
-
-이미지를 생성하고 살펴봅니다.
-
-```python
->>> images = pipe(prompt, generator=generator, num_images_per_prompt=4).images
->>> images
-```
-
-![img](https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/reusabe_seeds.jpg)
-
-이 예제에서는 첫 번째 이미지를 개선했지만 실제로는 원하는 모든 이미지를 사용할 수 있습니다(심지어 두 개의 눈이 있는 이미지도!). 첫 번째 이미지에서는 시드가 '0'인 '생성기'를 사용했기 때문에 두 번째 추론 라운드에서는 이 '생성기'를 재사용할 것입니다. 이미지의 품질을 개선하려면 프롬프트에 몇 가지 텍스트를 추가합니다:
-
-```python
-prompt = [prompt + t for t in [", highly realistic", ", artsy", ", trending", ", colorful"]]
-generator = [torch.Generator(device="cuda").manual_seed(0) for i in range(4)]
-```
-
-시드가 `0`인 제너레이터 4개를 생성하고, 이전 라운드의 첫 번째 이미지처럼 보이는 다른 이미지 batch(배치)를 생성합니다!
-
-```python
->>> images = pipe(prompt, generator=generator).images
->>> images
-```
-
-![img](https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/reusabe_seeds_2.jpg)

docs/source/ko/using-diffusers/sdxl_turbo.md

@@ -0,0 +1,114 @@
+<!--Copyright 2023 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 XL Turbo
+
+[[open-in-colab]]
+
+SDXL Turbo는 adversarial time-distilled(적대적 시간 전이) [Stable Diffusion XL](https://huggingface.co/papers/2307.01952)(SDXL) 모델로, 단 한 번의 스텝만으로 추론을 실행할 수 있습니다.
+
+이 가이드에서는 text-to-image와 image-to-image를 위한 SDXL-Turbo를 사용하는 방법을 설명합니다.
+
+시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:
+
+```py
+# Colab에서 필요한 라이브러리를 설치하기 위해 주석을 제외하세요
+#!pip install -q diffusers transformers accelerate
+```
+
+## 모델 체크포인트 불러오기
+
+모델 가중치는 Hub의 별도 하위 폴더 또는 로컬에 저장할 수 있으며, 이 경우 [`~StableDiffusionXLPipeline.from_pretrained`] 메서드를 사용해야 합니다:
+
+```py
+from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
+pipeline = pipeline.to("cuda")
+```
+
+또한 [`~StableDiffusionXLPipeline.from_single_file`] 메서드를 사용하여 허브 또는 로컬에서 단일 파일 형식(`.ckpt` 또는 `.safetensors`)으로 저장된 모델 체크포인트를 불러올 수도 있습니다:
+
+```py
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+pipeline = StableDiffusionXLPipeline.from_single_file(
+    "https://huggingface.co/stabilityai/sdxl-turbo/blob/main/sd_xl_turbo_1.0_fp16.safetensors", torch_dtype=torch.float16)
+pipeline = pipeline.to("cuda")
+```
+
+## Text-to-image
+
+Text-to-image의 경우 텍스트 프롬프트를 전달합니다. 기본적으로 SDXL Turbo는 512x512 이미지를 생성하며, 이 해상도에서 최상의 결과를 제공합니다. `height` 및 `width` 매개 변수를 768x768 또는 1024x1024로 설정할 수 있지만 이 경우 품질 저하를 예상할 수 있습니다.
+
+모델이 `guidance_scale` 없이 학습되었으므로 이를 0.0으로 설정해 비활성화해야 합니다. 단일 추론 스텝만으로도 고품질 이미지를 생성할 수 있습니다. 
+스텝 수를 2, 3 또는 4로 늘리면 이미지 품질이 향상됩니다.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline_text2image = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
+pipeline_text2image = pipeline_text2image.to("cuda")
+
+prompt = "A cinematic shot of a baby racoon wearing an intricate italian priest robe."
+
+image = pipeline_text2image(prompt=prompt, guidance_scale=0.0, num_inference_steps=1).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/sdxl-turbo-text2img.png" alt="generated image of a racoon in a robe"/>
+</div>
+
+## Image-to-image
+
+Image-to-image 생성의 경우 `num_inference_steps * strength`가 1보다 크거나 같은지 확인하세요. 
+Image-to-image 파이프라인은 아래 예제에서 `0.5 * 2.0 = 1` 스텝과 같이 `int(num_inference_steps * strength)` 스텝으로 실행됩니다.
+
+```py
+from diffusers import AutoPipelineForImage2Image
+from diffusers.utils import load_image, make_image_grid
+
+# 체크포인트를 불러올 때 추가 메모리 소모를 피하려면 from_pipe를 사용하세요.
+pipeline = AutoPipelineForImage2Image.from_pipe(pipeline_text2image).to("cuda")
+
+init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
+init_image = init_image.resize((512, 512))
+
+prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k"
+
+image = pipeline(prompt, image=init_image, strength=0.5, guidance_scale=0.0, num_inference_steps=2).images[0]
+make_image_grid([init_image, image], rows=1, cols=2)
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/sdxl-turbo-img2img.png" alt="Image-to-image generation sample using SDXL Turbo"/>
+</div>
+
+## SDXL Turbo 속도 훨씬 더 빠르게 하기
+
+- PyTorch 버전 2 이상을 사용하는 경우 UNet을 컴파일합니다. 첫 번째 추론 실행은 매우 느리지만 이후 실행은 훨씬 빨라집니다.
+
+```py
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+```
+
+- 기본 VAE를 사용하는 경우, 각 생성 전후에 비용이 많이 드는 `dtype` 변환을 피하기 위해 `float32`로 유지하세요. 이 작업은 첫 생성 이전에 한 번만 수행하면 됩니다:
+
+```py
+pipe.upcast_vae()
+```
+
+또는, 커뮤니티 회원인 [`@madebyollin`](https://huggingface.co/madebyollin)이 만든 [16비트 VAE](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)를 사용할 수도 있으며, 이는 `float32`로 업캐스트할 필요가 없습니다.

docs/source/ko/using-diffusers/shap-e.md

@@ -0,0 +1,192 @@
+<!--Copyright 2023 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.
+-->
+
+# Shap-E
+
+[[open-in-colab]]
+
+Shap-E는 비디오 게임 개발, 인테리어 디자인, 건축에 사용할 수 있는 3D 에셋을 생성하기 위한 conditional 모델입니다. 대규모 3D 에셋 데이터셋을 학습되었고, 각 오브젝트의 더 많은 뷰를 렌더링하고 4K point cloud 대신 16K를 생성하도록 후처리합니다. Shap-E 모델은 두 단계로 학습됩니다:
+
+1. 인코더가 3D 에셋의 포인트 클라우드와 렌더링된 뷰를 받아들이고 에셋을 나타내는 implicit functions의 파라미터를 출력합니다.
+2. 인코더가 생성한 latents를 바탕으로 diffusion 모델을 훈련하여 neural radiance fields(NeRF) 또는 textured 3D 메시를 생성하여 다운스트림 애플리케이션에서 3D 에셋을 더 쉽게 렌더링하고 사용할 수 있도록 합니다.
+
+이 가이드에서는 Shap-E를 사용하여 나만의 3D 에셋을 생성하는 방법을 보입니다!
+
+시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:
+
+```py
+# Colab에서 필요한 라이브러리를 설치하기 위해 주석을 제외하세요
+#!pip install -q diffusers transformers accelerate trimesh
+```
+
+## Text-to-3D
+
+3D 객체의 gif를 생성하려면 텍스트 프롬프트를 [`ShapEPipeline`]에 전달합니다. 파이프라인은 3D 객체를 생성하는 데 사용되는 이미지 프레임 리스트를 생성합니다.
+
+```py
+import torch
+from diffusers import ShapEPipeline
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+pipe = ShapEPipeline.from_pretrained("openai/shap-e", torch_dtype=torch.float16, variant="fp16")
+pipe = pipe.to(device)
+
+guidance_scale = 15.0
+prompt = ["A firecracker", "A birthday cupcake"]
+
+images = pipe(
+    prompt,
+    guidance_scale=guidance_scale,
+    num_inference_steps=64,
+    frame_size=256,
+).images
+```
+
+이제 [`~utils.export_to_gif`] 함수를 사용하여 이미지 프레임 리스트를 3D 객체의 gif로 변환합니다.
+
+```py
+from diffusers.utils import export_to_gif
+
+export_to_gif(images[0], "firecracker_3d.gif")
+export_to_gif(images[1], "cake_3d.gif")
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/shap_e/firecracker_out.gif"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">prompt = "A firecracker"</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/shap_e/cake_out.gif"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">prompt = "A birthday cupcake"</figcaption>
+  </div>
+</div>
+
+## Image-to-3D
+
+다른 이미지로부터 3D 개체를 생성하려면 [`ShapEImg2ImgPipeline`]을 사용합니다. 기존 이미지를 사용하거나 완전히 새로운 이미지를 생성할 수 있습니다. [Kandinsky 2.1](../api/pipelines/kandinsky) 모델을 사용하여 새 이미지를 생성해 보겠습니다.
+
+```py
+from diffusers import DiffusionPipeline
+import torch
+
+prior_pipeline = DiffusionPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipeline = DiffusionPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+
+prompt = "A cheeseburger, white background"
+
+image_embeds, negative_image_embeds = prior_pipeline(prompt, guidance_scale=1.0).to_tuple()
+image = pipeline(
+    prompt,
+    image_embeds=image_embeds,
+    negative_image_embeds=negative_image_embeds,
+).images[0]
+
+image.save("burger.png")
+```
+
+치즈버거를 [`ShapEImg2ImgPipeline`]에 전달하여 3D representation을 생성합니다.
+
+```py
+from PIL import Image
+from diffusers import ShapEImg2ImgPipeline
+from diffusers.utils import export_to_gif
+
+pipe = ShapEImg2ImgPipeline.from_pretrained("openai/shap-e-img2img", torch_dtype=torch.float16, variant="fp16").to("cuda")
+
+guidance_scale = 3.0
+image = Image.open("burger.png").resize((256, 256))
+
+images = pipe(
+    image,
+    guidance_scale=guidance_scale,
+    num_inference_steps=64,
+    frame_size=256,
+).images
+
+gif_path = export_to_gif(images[0], "burger_3d.gif")
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/shap_e/burger_in.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">cheeseburger</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/shap_e/burger_out.gif"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">3D cheeseburger</figcaption>
+  </div>
+</div>
+
+## 메시 생성하기
+
+Shap-E는 다운스트림 애플리케이션에 렌더링할 textured 메시 출력을 생성할 수도 있는 유연한 모델입니다. 이 예제에서는 🤗 Datasets 라이브러리에서 [Dataset viewer](https://huggingface.co/docs/hub/datasets-viewer#dataset-preview)를 사용해 메시 시각화를 지원하는 `glb` 파일로 변환합니다.
+
+`output_type` 매개변수를 `"mesh"`로 지정함으로써 [`ShapEPipeline`]과 [`ShapEImg2ImgPipeline`] 모두에 대한 메시 출력을 생성할 수 있습니다:
+
+```py
+import torch
+from diffusers import ShapEPipeline
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+pipe = ShapEPipeline.from_pretrained("openai/shap-e", torch_dtype=torch.float16, variant="fp16")
+pipe = pipe.to(device)
+
+guidance_scale = 15.0
+prompt = "A birthday cupcake"
+
+images = pipe(prompt, guidance_scale=guidance_scale, num_inference_steps=64, frame_size=256, output_type="mesh").images
+```
+
+메시 출력을 `ply` 파일로 저장하려면 [`~utils.export_to_ply`] 함수를 사용합니다:
+
+<Tip>
+
+선택적으로 [`~utils.export_to_obj`] 함수를 사용하여 메시 출력을 `obj` 파일로 저장할 수 있습니다. 다양한 형식으로 메시 출력을 저장할 수 있어 다운스트림에서 더욱 유연하게 사용할 수 있습니다!
+
+</Tip>
+
+```py
+from diffusers.utils import export_to_ply
+
+ply_path = export_to_ply(images[0], "3d_cake.ply")
+print(f"Saved to folder: {ply_path}")
+```
+
+그 다음 trimesh 라이브러리를 사용하여 `ply` 파일을 `glb` 파일로 변환할 수 있습니다:
+
+```py
+import trimesh
+
+mesh = trimesh.load("3d_cake.ply")
+mesh_export = mesh.export("3d_cake.glb", file_type="glb")
+```
+
+기본적으로 메시 출력은 아래쪽 시점에 초점이 맞춰져 있지만 회전 변환을 적용하여 기본 시점을 변경할 수 있습니다:
+
+```py
+import trimesh
+import numpy as np
+
+mesh = trimesh.load("3d_cake.ply")
+rot = trimesh.transformations.rotation_matrix(-np.pi / 2, [1, 0, 0])
+mesh = mesh.apply_transform(rot)
+mesh_export = mesh.export("3d_cake.glb", file_type="glb")
+```
+
+메시 파일을 데이터셋 레포지토리에 업로드해 Dataset viewer로 시각화하세요!
+
+<div class="flex justify-center">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/3D-cake.gif"/>
+</div>

docs/source/ko/using-diffusers/svd.md

@@ -0,0 +1,121 @@
+<!--Copyright 2023 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 Video Diffusion
+
+[[open-in-colab]]
+
+[Stable Video Diffusion (SVD)](https://huggingface.co/papers/2311.15127)은 입력 이미지에 맞춰 2~4초 분량의 고해상도(576x1024) 비디오를 생성할 수 있는 강력한 image-to-video 생성 모델입니다.
+
+이 가이드에서는 SVD를 사용하여 이미지에서 짧은 동영상을 생성하는 방법을 설명합니다.
+
+시작하기 전에 다음 라이브러리가 설치되어 있는지 확인하세요:
+
+```py
+!pip install -q -U diffusers transformers accelerate 
+```
+
+이 모델에는 [SVD](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid)와 [SVD-XT](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt) 두 가지 종류가 있습니다. SVD 체크포인트는 14개의 프레임을 생성하도록 학습되었고, SVD-XT 체크포인트는 25개의 프레임을 생성하도록 파인튜닝되었습니다.
+
+이 가이드에서는 SVD-XT 체크포인트를 사용합니다.
+
+```python
+import torch
+
+from diffusers import StableVideoDiffusionPipeline
+from diffusers.utils import load_image, export_to_video
+
+pipe = StableVideoDiffusionPipeline.from_pretrained(
+    "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16"
+)
+pipe.enable_model_cpu_offload()
+
+# Conditioning 이미지 불러오기
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png")
+image = image.resize((1024, 576))
+
+generator = torch.manual_seed(42)
+frames = pipe(image, decode_chunk_size=8, generator=generator).frames[0]
+
+export_to_video(frames, "generated.mp4", fps=7)
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">"source image of a rocket"</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/output_rocket.gif"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">"generated video from source image"</figcaption>
+  </div>
+</div>
+
+## torch.compile
+
+UNet을 [컴파일](../optimization/torch2.0#torchcompile)하면 메모리 사용량이 살짝 증가하지만, 20~25%의 속도 향상을 얻을 수 있습니다.
+
+```diff
+- pipe.enable_model_cpu_offload()
++ pipe.to("cuda")
++ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+```
+
+## 메모리 사용량 줄이기
+
+비디오 생성은 기본적으로 배치 크기가 큰 text-to-image 생성과 유사하게 'num_frames'를 한 번에 생성하기 때문에 메모리 사용량이 매우 높습니다. 메모리 사용량을 줄이기 위해 추론 속도와 메모리 사용량을 절충하는 여러 가지 옵션이 있습니다:
+
+- 모델 오프로링 활성화: 파이프라인의 각 구성 요소가 더 이상 필요하지 않을 때 CPU로 오프로드됩니다.
+- Feed-forward chunking 활성화: feed-forward 레이어가 배치 크기가 큰 단일 feed-forward를 실행하는 대신 루프로 반복해서 실행됩니다.
+- `decode_chunk_size` 감소: VAE가 프레임들을 한꺼번에 디코딩하는 대신 chunk 단위로 디코딩합니다. `decode_chunk_size=1`을 설정하면 한 번에 한 프레임씩 디코딩하고 최소한의 메모리만 사용하지만(GPU 메모리에 따라 이 값을 조정하는 것이 좋습니다), 동영상에 약간의 깜박임이 발생할 수 있습니다.
+
+```diff
+- pipe.enable_model_cpu_offload()
+- frames = pipe(image, decode_chunk_size=8, generator=generator).frames[0]
++ pipe.enable_model_cpu_offload()
++ pipe.unet.enable_forward_chunking()
++ frames = pipe(image, decode_chunk_size=2, generator=generator, num_frames=25).frames[0]
+```
+
+이러한 모든 방법들을 사용하면 메모리 사용량이 8GAM VRAM보다 적을 것입니다.
+
+## Micro-conditioning
+
+Stable Diffusion Video는 또한 이미지 conditoning 외에도 micro-conditioning을 허용하므로 생성된 비디오를 더 잘 제어할 수 있습니다:
+
+- `fps`: 생성된 비디오의 초당 프레임 수입니다.
+- `motion_bucket_id`: 생성된 동영상에 사용할 모션 버킷 아이디입니다. 생성된 동영상의 모션을 제어하는 데 사용할 수 있습니다. 모션 버킷 아이디를 늘리면 생성되는 동영상의 모션이 증가합니다.
+- `noise_aug_strength`: Conditioning 이미지에 추가되는 노이즈의 양입니다. 값이 클수록 비디오가 conditioning 이미지와 덜 유사해집니다. 이 값을 높이면 생성된 비디오의 움직임도 증가합니다.
+
+예를 들어, 모션이 더 많은 동영상을 생성하려면 `motion_bucket_id` 및 `noise_aug_strength` micro-conditioning 파라미터를 사용합니다:
+
+```python
+import torch
+
+from diffusers import StableVideoDiffusionPipeline
+from diffusers.utils import load_image, export_to_video
+
+pipe = StableVideoDiffusionPipeline.from_pretrained(
+  "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16"
+)
+pipe.enable_model_cpu_offload()
+
+# Conditioning 이미지 불러오기
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png")
+image = image.resize((1024, 576))
+
+generator = torch.manual_seed(42)
+frames = pipe(image, decode_chunk_size=8, generator=generator, motion_bucket_id=180, noise_aug_strength=0.1).frames[0]
+export_to_video(frames, "generated.mp4", fps=7)
+```
+
+![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/output_rocket_with_conditions.gif)

docs/source/ko/using-diffusers/using_safetensors.md

@@ -1,67 +0,0 @@
-# 세이프텐서 로드
-
-[safetensors](https://github.com/huggingface/safetensors)는 텐서를 저장하고 로드하기 위한 안전하고 빠른 파일 형식입니다. 일반적으로 PyTorch 모델 가중치는 Python의 [`pickle`](https://docs.python.org/3/library/pickle.html) 유틸리티를 사용하여 `.bin` 파일에 저장되거나 `피클`됩니다. 그러나 `피클`은 안전하지 않으며 피클된 파일에는 실행될 수 있는 악성 코드가 포함될 수 있습니다. 세이프텐서는 `피클`의 안전한 대안으로 모델 가중치를 공유하는 데 이상적입니다.
-
-이 가이드에서는 `.safetensor` 파일을 로드하는 방법과 다른 형식으로 저장된 안정적 확산 모델 가중치를 `.safetensor`로 변환하는 방법을 보여드리겠습니다. 시작하기 전에 세이프텐서가 설치되어 있는지 확인하세요:
-
-```bash
-!pip install safetensors
-```
-
-['runwayml/stable-diffusion-v1-5`] (https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main) 리포지토리를 보면 `text_encoder`, `unet` 및 `vae` 하위 폴더에 가중치가 `.safetensors` 형식으로 저장되어 있는 것을 볼 수 있습니다. 기본적으로 🤗 디퓨저는 모델 저장소에서 사용할 수 있는 경우 해당 하위 폴더에서 이러한 '.safetensors` 파일을 자동으로 로드합니다.
-
-보다 명시적인 제어를 위해 선택적으로 `사용_세이프텐서=True`를 설정할 수 있습니다(`세이프텐서`가 설치되지 않은 경우 설치하라는 오류 메시지가 표시됨):
-
-```py
-from diffusers import DiffusionPipeline
-
-pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", use_safetensors=True)
-```
-
-그러나 모델 가중치가 위의 예시처럼 반드시 별도의 하위 폴더에 저장되는 것은 아닙니다. 모든 가중치가 하나의 '.safetensors` 파일에 저장되는 경우도 있습니다. 이 경우 가중치가 Stable Diffusion 가중치인 경우 [`~diffusers.loaders.FromCkptMixin.from_ckpt`] 메서드를 사용하여 파일을 직접 로드할 수 있습니다:
-
-```py
-from diffusers import StableDiffusionPipeline
-
-pipeline = StableDiffusionPipeline.from_ckpt(
-    "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
-)
-```
-
-## 세이프텐서로 변환
-
-허브의 모든 가중치를 '.safetensors` 형식으로 사용할 수 있는 것은 아니며, '.bin`으로 저장된 가중치가 있을 수 있습니다. 이 경우 [Convert Space](https://huggingface.co/spaces/diffusers/convert)을 사용하여 가중치를 '.safetensors'로 변환하세요. Convert Space는 피클된 가중치를 다운로드하여 변환한 후 풀 리퀘스트를 열어 허브에 새로 변환된 `.safetensors` 파일을 업로드합니다. 이렇게 하면 피클된 파일에 악성 코드가 포함되어 있는 경우, 안전하지 않은 파일과 의심스러운 피클 가져오기를 탐지하는 [보안 스캐너](https://huggingface.co/docs/hub/security-pickle#hubs-security-scanner)가 있는 허브로 업로드됩니다. - 개별 컴퓨터가 아닌.
-
-개정` 매개변수에 풀 리퀘스트에 대한 참조를 지정하여 새로운 '.safetensors` 가중치가 적용된 모델을 사용할 수 있습니다(허브의 [Check PR](https://huggingface.co/spaces/diffusers/check_pr) 공간에서 테스트할 수도 있음)(예: `refs/pr/22`):
-
-```py
-from diffusers import DiffusionPipeline
-
-pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", revision="refs/pr/22")
-```
-
-## 세이프센서를 사용하는 이유는 무엇인가요?
-
-세이프티 센서를 사용하는 데에는 여러 가지 이유가 있습니다:
-
-- 세이프텐서를 사용하는 가장 큰 이유는 안전입니다.오픈 소스 및 모델 배포가 증가함에 따라 다운로드한 모델 가중치에 악성 코드가 포함되어 있지 않다는 것을 신뢰할 수 있는 것이 중요해졌습니다.세이프센서의 현재 헤더 크기는 매우 큰 JSON 파일을 구문 분석하지 못하게 합니다.
-- 모델 전환 간의 로딩 속도는 텐서의 제로 카피를 수행하는 세이프텐서를 사용해야 하는 또 다른 이유입니다. 가중치를 CPU(기본값)로 로드하는 경우 '피클'에 비해 특히 빠르며, 가중치를 GPU로 직접 로드하는 경우에도 빠르지는 않더라도 비슷하게 빠릅니다. 모델이 이미 로드된 경우에만 성능 차이를 느낄 수 있으며, 가중치를 다운로드하거나 모델을 처음 로드하는 경우에는 성능 차이를 느끼지 못할 것입니다.
-
-	전체 파이프라인을 로드하는 데 걸리는 시간입니다:
-
-	```py
- from diffusers import StableDiffusionPipeline
-
- pipeline = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")
- "Loaded in safetensors 0:00:02.033658"
- "Loaded in PyTorch 0:00:02.663379"
-	```
-
-	하지만 실제로 500MB의 모델 가중치를 로드하는 데 걸리는 시간은 얼마 되지 않습니다:
-
-	```bash
-	safetensors: 3.4873ms
-	PyTorch: 172.7537ms
-	```
-
-지연 로딩은 세이프텐서에서도 지원되며, 이는 분산 설정에서 일부 텐서만 로드하는 데 유용합니다. 이 형식을 사용하면 [BLOOM](https://huggingface.co/bigscience/bloom) 모델을 일반 PyTorch 가중치를 사용하여 10분이 걸리던 것을 8개의 GPU에서 45초 만에 로드할 수 있습니다.

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -1290,6 +1290,7 @@ def main(args):
                         text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
                             get_peft_model_state_dict(model)
                         )
+                else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
                 # make sure to pop weight so that corresponding model is not saved again
@@ -1301,7 +1302,7 @@ def main(args):
                 text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
             )
         if args.train_text_encoder_ti:
-            embedding_handler.save_embeddings(f"{output_dir}/{args.output_dir}_emb.safetensors")
+            embedding_handler.save_embeddings(f"{args.output_dir}/{Path(args.output_dir).name}_emb.safetensors")
 
     def load_model_hook(models, input_dir):
         unet_ = None
@@ -1524,17 +1525,22 @@ def main(args):
                 torch.cuda.empty_cache()
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     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
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1551,8 +1557,14 @@ def main(args):
 
     # 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:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+        logger.warning(
+            f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+            f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+            f"This inconsistency may result in the learning rate scheduler not functioning properly."
+        )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
@@ -1845,10 +1857,10 @@ def main(args):
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
 
-            if torch.backends.mps.is_available():
-                autocast_ctx = nullcontext()
-            else:
-                autocast_ctx = torch.autocast(accelerator.device.type)
+                if torch.backends.mps.is_available():
+                    autocast_ctx = nullcontext()
+                else:
+                    autocast_ctx = torch.autocast(accelerator.device.type)
 
                 with autocast_ctx:
                     images = [
@@ -1869,7 +1881,6 @@ def main(args):
                                 ]
                             }
                         )
-
                 del pipeline
                 torch.cuda.empty_cache()
 
@@ -1971,7 +1982,7 @@ def main(args):
         lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
         peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
         kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
-        save_file(kohya_state_dict, f"{args.output_dir}/{args.output_dir}.safetensors")
+        save_file(kohya_state_dict, f"{args.output_dir}/{Path(args.output_dir).name}.safetensors")
 
         save_model_card(
             model_id if not args.push_to_hub else repo_id,

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -31,8 +31,6 @@ from typing import List, Optional
 import numpy as np
 import torch
 import torch.nn.functional as F
-
-# imports of the TokenEmbeddingsHandler class
 import torch.utils.checkpoint
 import transformers
 from accelerate import Accelerator
@@ -77,6 +75,9 @@ from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
+if is_wandb_available():
+    import wandb
+
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
 check_min_version("0.30.0.dev0")
 
@@ -101,12 +102,12 @@ def save_model_card(
     repo_id: str,
     use_dora: bool,
     images=None,
-    base_model=str,
+    base_model: str = None,
     train_text_encoder=False,
     train_text_encoder_ti=False,
     token_abstraction_dict=None,
-    instance_prompt=str,
-    validation_prompt=str,
+    instance_prompt: str = None,
+    validation_prompt: str = None,
     repo_folder=None,
     vae_path=None,
 ):
@@ -135,6 +136,14 @@ def save_model_card(
     diffusers_imports_pivotal = ""
     diffusers_example_pivotal = ""
     webui_example_pivotal = ""
+    license = ""
+    if "playground" in base_model:
+        license = """\n
+    ## License
+
+    Please adhere to the licensing terms as described [here](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic/blob/main/LICENSE.md).
+    """
+
     if train_text_encoder_ti:
         trigger_str = (
             "To trigger image generation of trained concept(or concepts) replace each concept identifier "
@@ -223,11 +232,75 @@ Pivotal tuning was enabled: {train_text_encoder_ti}.
 
 Special VAE used for training: {vae_path}.
 
+{license}
 """
     with open(os.path.join(repo_folder, "README.md"), "w") as f:
         f.write(yaml + model_card)
 
 
+def log_validation(
+    pipeline,
+    args,
+    accelerator,
+    pipeline_args,
+    epoch,
+    is_final_validation=False,
+):
+    logger.info(
+        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+        f" {args.validation_prompt}."
+    )
+
+    # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
+    scheduler_args = {}
+
+    if not args.do_edm_style_training:
+        if "variance_type" in pipeline.scheduler.config:
+            variance_type = pipeline.scheduler.config.variance_type
+
+            if variance_type in ["learned", "learned_range"]:
+                variance_type = "fixed_small"
+
+            scheduler_args["variance_type"] = variance_type
+
+        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
+
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    # run inference
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
+    # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
+    if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
+        autocast_ctx = nullcontext()
+    else:
+        autocast_ctx = torch.autocast(accelerator.device.type)
+
+    with autocast_ctx:
+        images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+
+    for tracker in accelerator.trackers:
+        phase_name = "test" if is_final_validation else "validation"
+        if tracker.name == "tensorboard":
+            np_images = np.stack([np.asarray(img) for img in images])
+            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
+        if tracker.name == "wandb":
+            tracker.log(
+                {
+                    phase_name: [
+                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
+                    ]
+                }
+            )
+
+    del pipeline
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    return images
+
+
 def import_model_class_from_model_name_or_path(
     pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
 ):
@@ -390,6 +463,7 @@ def parse_args(input_args=None):
     )
     parser.add_argument(
         "--do_edm_style_training",
+        default=False,
         action="store_true",
         help="Flag to conduct training using the EDM formulation as introduced in https://arxiv.org/abs/2206.00364.",
     )
@@ -499,6 +573,13 @@ def parse_args(input_args=None):
         default=1e-4,
         help="Initial learning rate (after the potential warmup period) to use.",
     )
+    parser.add_argument(
+        "--clip_skip",
+        type=int,
+        default=None,
+        help="Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that "
+        "the output of the pre-final layer will be used for computing the prompt embeddings.",
+    )
 
     parser.add_argument(
         "--text_encoder_lr",
@@ -571,7 +652,7 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--optimizer",
         type=str,
-        default="adamW",
+        default="AdamW",
         help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
     )
 
@@ -906,11 +987,6 @@ class DreamBoothDataset(Dataset):
         instance_data_root,
         instance_prompt,
         class_prompt,
-        dataset_name,
-        dataset_config_name,
-        cache_dir,
-        image_column,
-        caption_column,
         train_text_encoder_ti,
         class_data_root=None,
         class_num=None,
@@ -929,7 +1005,7 @@ class DreamBoothDataset(Dataset):
         self.train_text_encoder_ti = train_text_encoder_ti
         # if --dataset_name is provided or a metadata jsonl file is provided in the local --instance_data directory,
         # we load the training data using load_dataset
-        if dataset_name is not None:
+        if args.dataset_name is not None:
             try:
                 from datasets import load_dataset
             except ImportError:
@@ -942,25 +1018,26 @@ class DreamBoothDataset(Dataset):
             # See more about loading custom images at
             # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
             dataset = load_dataset(
-                dataset_name,
-                dataset_config_name,
-                cache_dir=cache_dir,
+                args.dataset_name,
+                args.dataset_config_name,
+                cache_dir=args.cache_dir,
             )
             # Preprocessing the datasets.
             column_names = dataset["train"].column_names
 
             # 6. Get the column names for input/target.
-            if image_column is None:
+            if args.image_column is None:
                 image_column = column_names[0]
                 logger.info(f"image column defaulting to {image_column}")
             else:
+                image_column = args.image_column
                 if image_column not in column_names:
                     raise ValueError(
-                        f"`--image_column` value '{image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                        f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
                     )
             instance_images = dataset["train"][image_column]
 
-            if caption_column is None:
+            if args.caption_column is None:
                 logger.info(
                     "No caption column provided, defaulting to instance_prompt for all images. If your dataset "
                     "contains captions/prompts for the images, make sure to specify the "
@@ -968,11 +1045,11 @@ class DreamBoothDataset(Dataset):
                 )
                 self.custom_instance_prompts = None
             else:
-                if caption_column not in column_names:
+                if args.caption_column not in column_names:
                     raise ValueError(
-                        f"`--caption_column` value '{caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                        f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
                     )
-                custom_instance_prompts = dataset["train"][caption_column]
+                custom_instance_prompts = dataset["train"][args.caption_column]
                 # create final list of captions according to --repeats
                 self.custom_instance_prompts = []
                 for caption in custom_instance_prompts:
@@ -1166,7 +1243,7 @@ def tokenize_prompt(tokenizer, prompt, add_special_tokens=False):
 
 
 # Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
-def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
+def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None, clip_skip=None):
     prompt_embeds_list = []
 
     for i, text_encoder in enumerate(text_encoders):
@@ -1178,13 +1255,16 @@ def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
             text_input_ids = text_input_ids_list[i]
 
         prompt_embeds = text_encoder(
-            text_input_ids.to(text_encoder.device),
-            output_hidden_states=True,
+            text_input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=False
         )
 
         # We are only ALWAYS interested in the pooled output of the final text encoder
         pooled_prompt_embeds = prompt_embeds[0]
-        prompt_embeds = prompt_embeds.hidden_states[-2]
+        if clip_skip is None:
+            prompt_embeds = prompt_embeds[-1][-2]
+        else:
+            # "2" because SDXL always indexes from the penultimate layer.
+            prompt_embeds = prompt_embeds[-1][-(clip_skip + 2)]
         bs_embed, seq_len, _ = prompt_embeds.shape
         prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
         prompt_embeds_list.append(prompt_embeds)
@@ -1200,9 +1280,16 @@ def main(args):
             "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
             " Please use `huggingface-cli login` to authenticate with the Hub."
         )
+
     if args.do_edm_style_training and args.snr_gamma is not None:
         raise ValueError("Min-SNR formulation is not supported when conducting EDM-style training.")
 
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     logging_dir = Path(args.output_dir, args.logging_dir)
 
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
@@ -1215,10 +1302,13 @@ def main(args):
         kwargs_handlers=[kwargs],
     )
 
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
     if args.report_to == "wandb":
         if not is_wandb_available():
             raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
-        import wandb
 
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
@@ -1246,7 +1336,8 @@ def main(args):
         cur_class_images = len(list(class_images_dir.iterdir()))
 
         if cur_class_images < args.num_class_images:
-            torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
+            has_supported_fp16_accelerator = torch.cuda.is_available() or torch.backends.mps.is_available()
+            torch_dtype = torch.float16 if has_supported_fp16_accelerator else torch.float32
             if args.prior_generation_precision == "fp32":
                 torch_dtype = torch.float32
             elif args.prior_generation_precision == "fp16":
@@ -1404,6 +1495,12 @@ def main(args):
     elif accelerator.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
+    if torch.backends.mps.is_available() and weight_dtype == torch.bfloat16:
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     # Move unet, vae and text_encoder to device and cast to weight_dtype
     unet.to(accelerator.device, dtype=weight_dtype)
 
@@ -1530,7 +1627,7 @@ def main(args):
                 text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
             )
         if args.train_text_encoder_ti:
-            embedding_handler.save_embeddings(f"{output_dir}/{args.output_dir}_emb.safetensors")
+            embedding_handler.save_embeddings(f"{args.output_dir}/{Path(args.output_dir).name}_emb.safetensors")
 
     def load_model_hook(models, input_dir):
         unet_ = None
@@ -1564,6 +1661,7 @@ def main(args):
                 )
 
         if args.train_text_encoder:
+            # Do we need to call `scale_lora_layers()` here?
             _set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
 
             _set_state_dict_into_text_encoder(
@@ -1578,14 +1676,14 @@ def main(args):
             if args.train_text_encoder:
                 models.extend([text_encoder_one_, text_encoder_two_])
                 # only upcast trainable parameters (LoRA) into fp32
-            cast_training_params(models)
+                cast_training_params(models)
 
     accelerator.register_save_state_pre_hook(save_model_hook)
     accelerator.register_load_state_pre_hook(load_model_hook)
 
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
-    if args.allow_tf32:
+    if args.allow_tf32 and torch.cuda.is_available():
         torch.backends.cuda.matmul.allow_tf32 = True
 
     if args.scale_lr:
@@ -1711,12 +1809,7 @@ def main(args):
         instance_data_root=args.instance_data_dir,
         instance_prompt=args.instance_prompt,
         class_prompt=args.class_prompt,
-        dataset_name=args.dataset_name,
-        dataset_config_name=args.dataset_config_name,
-        cache_dir=args.cache_dir,
-        image_column=args.image_column,
         train_text_encoder_ti=args.train_text_encoder_ti,
-        caption_column=args.caption_column,
         class_data_root=args.class_data_dir if args.with_prior_preservation else None,
         token_abstraction_dict=token_abstraction_dict if args.train_text_encoder_ti else None,
         class_num=args.num_class_images,
@@ -1740,8 +1833,6 @@ def main(args):
 
     def compute_time_ids(crops_coords_top_left, original_size=None):
         # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
-        if original_size is None:
-            original_size = (args.resolution, args.resolution)
         target_size = (args.resolution, args.resolution)
         add_time_ids = list(original_size + crops_coords_top_left + target_size)
         add_time_ids = torch.tensor([add_time_ids])
@@ -1752,9 +1843,9 @@ def main(args):
         tokenizers = [tokenizer_one, tokenizer_two]
         text_encoders = [text_encoder_one, text_encoder_two]
 
-        def compute_text_embeddings(prompt, text_encoders, tokenizers):
+        def compute_text_embeddings(prompt, text_encoders, tokenizers, clip_skip):
             with torch.no_grad():
-                prompt_embeds, pooled_prompt_embeds = encode_prompt(text_encoders, tokenizers, prompt)
+                prompt_embeds, pooled_prompt_embeds = encode_prompt(text_encoders, tokenizers, prompt, clip_skip)
                 prompt_embeds = prompt_embeds.to(accelerator.device)
                 pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
             return prompt_embeds, pooled_prompt_embeds
@@ -1764,7 +1855,7 @@ def main(args):
     # the redundant encoding.
     if freeze_text_encoder and not train_dataset.custom_instance_prompts:
         instance_prompt_hidden_states, instance_pooled_prompt_embeds = compute_text_embeddings(
-            args.instance_prompt, text_encoders, tokenizers
+            args.instance_prompt, text_encoders, tokenizers, args.clip_skip
         )
 
     # Handle class prompt for prior-preservation.
@@ -1778,7 +1869,8 @@ def main(args):
     if freeze_text_encoder and not train_dataset.custom_instance_prompts:
         del tokenizers, text_encoders
         gc.collect()
-        torch.cuda.empty_cache()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
 
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
@@ -1820,17 +1912,22 @@ def main(args):
                 torch.cuda.empty_cache()
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     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
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1847,8 +1944,14 @@ def main(args):
 
     # 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:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
@@ -1946,8 +2049,8 @@ def main(args):
                 text_encoder_two.train()
                 # set top parameter requires_grad = True for gradient checkpointing works
                 if args.train_text_encoder:
-                    text_encoder_one.text_model.embeddings.requires_grad_(True)
-                    text_encoder_two.text_model.embeddings.requires_grad_(True)
+                    accelerator.unwrap_model(text_encoder_one).text_model.embeddings.requires_grad_(True)
+                    accelerator.unwrap_model(text_encoder_two).text_model.embeddings.requires_grad_(True)
 
         for step, batch in enumerate(train_dataloader):
             if pivoted:
@@ -1962,7 +2065,7 @@ def main(args):
                 if train_dataset.custom_instance_prompts:
                     if freeze_text_encoder:
                         prompt_embeds, unet_add_text_embeds = compute_text_embeddings(
-                            prompts, text_encoders, tokenizers
+                            prompts, text_encoders, tokenizers, args.clip_skip
                         )
 
                     else:
@@ -2040,7 +2143,6 @@ def main(args):
                 if freeze_text_encoder:
                     unet_added_conditions = {
                         "time_ids": add_time_ids,
-                        # "time_ids": add_time_ids.repeat(elems_to_repeat_time_ids, 1),
                         "text_embeds": unet_add_text_embeds.repeat(elems_to_repeat_text_embeds, 1),
                     }
                     prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
@@ -2058,6 +2160,7 @@ def main(args):
                         tokenizers=None,
                         prompt=None,
                         text_input_ids_list=[tokens_one, tokens_two],
+                        clip_skip=args.clip_skip,
                     )
                     unet_added_conditions.update(
                         {"text_embeds": pooled_prompt_embeds.repeat(elems_to_repeat_text_embeds, 1)}
@@ -2220,10 +2323,6 @@ def main(args):
 
         if accelerator.is_main_process:
             if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
-                logger.info(
-                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
-                    f" {args.validation_prompt}."
-                )
                 # create pipeline
                 if freeze_text_encoder:
                     text_encoder_one = text_encoder_cls_one.from_pretrained(
@@ -2250,70 +2349,29 @@ def main(args):
                     variant=args.variant,
                     torch_dtype=weight_dtype,
                 )
-
-                # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
-                scheduler_args = {}
-
-                if not args.do_edm_style_training:
-                    if "variance_type" in pipeline.scheduler.config:
-                        variance_type = pipeline.scheduler.config.variance_type
-
-                        if variance_type in ["learned", "learned_range"]:
-                            variance_type = "fixed_small"
-
-                        scheduler_args["variance_type"] = variance_type
-
-                    pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
-                        pipeline.scheduler.config, **scheduler_args
-                    )
-
-                pipeline = pipeline.to(accelerator.device)
-                pipeline.set_progress_bar_config(disable=True)
-
-                # run inference
-                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
-                if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
-                    autocast_ctx = nullcontext()
-                else:
-                    autocast_ctx = torch.autocast(accelerator.device.type)
 
-                with autocast_ctx:
-                    images = [
-                        pipeline(**pipeline_args, generator=generator).images[0]
-                        for _ in range(args.num_validation_images)
-                    ]
-
-                for tracker in accelerator.trackers:
-                    if tracker.name == "tensorboard":
-                        np_images = np.stack([np.asarray(img) for img in images])
-                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
-                    if tracker.name == "wandb":
-                        tracker.log(
-                            {
-                                "validation": [
-                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                    for i, image in enumerate(images)
-                                ]
-                            }
-                        )
-
-                del pipeline
-                torch.cuda.empty_cache()
+                images = log_validation(
+                    pipeline,
+                    args,
+                    accelerator,
+                    pipeline_args,
+                    epoch,
+                )
 
     # Save the lora layers
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
-        unet = accelerator.unwrap_model(unet)
+        unet = unwrap_model(unet)
         unet = unet.to(torch.float32)
         unet_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
 
         if args.train_text_encoder:
-            text_encoder_one = accelerator.unwrap_model(text_encoder_one)
+            text_encoder_one = unwrap_model(text_encoder_one)
             text_encoder_lora_layers = convert_state_dict_to_diffusers(
                 get_peft_model_state_dict(text_encoder_one.to(torch.float32))
             )
-            text_encoder_two = accelerator.unwrap_model(text_encoder_two)
+            text_encoder_two = unwrap_model(text_encoder_two)
             text_encoder_2_lora_layers = convert_state_dict_to_diffusers(
                 get_peft_model_state_dict(text_encoder_two.to(torch.float32))
             )
@@ -2332,90 +2390,44 @@ def main(args):
             embeddings_path = f"{args.output_dir}/{args.output_dir}_emb.safetensors"
             embedding_handler.save_embeddings(embeddings_path)
 
+        # Final inference
+        # Load previous pipeline
+        vae = AutoencoderKL.from_pretrained(
+            vae_path,
+            subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
+            revision=args.revision,
+            variant=args.variant,
+            torch_dtype=weight_dtype,
+        )
+        pipeline = StableDiffusionXLPipeline.from_pretrained(
+            args.pretrained_model_name_or_path,
+            vae=vae,
+            revision=args.revision,
+            variant=args.variant,
+            torch_dtype=weight_dtype,
+        )
+
+        # load attention processors
+        pipeline.load_lora_weights(args.output_dir)
+
+        # run inference
         images = []
         if args.validation_prompt and args.num_validation_images > 0:
-            # Final inference
-            # Load previous pipeline
-            vae = AutoencoderKL.from_pretrained(
-                vae_path,
-                subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
-                revision=args.revision,
-                variant=args.variant,
-                torch_dtype=weight_dtype,
+            pipeline_args = {"prompt": args.validation_prompt, "num_inference_steps": 25}
+            images = log_validation(
+                pipeline,
+                args,
+                accelerator,
+                pipeline_args,
+                epoch,
+                is_final_validation=True,
             )
-            pipeline = StableDiffusionXLPipeline.from_pretrained(
-                args.pretrained_model_name_or_path,
-                vae=vae,
-                revision=args.revision,
-                variant=args.variant,
-                torch_dtype=weight_dtype,
-            )
-
-            # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
-            scheduler_args = {}
-
-            if not args.do_edm_style_training:
-                if "variance_type" in pipeline.scheduler.config:
-                    variance_type = pipeline.scheduler.config.variance_type
-
-                    if variance_type in ["learned", "learned_range"]:
-                        variance_type = "fixed_small"
-
-                    scheduler_args["variance_type"] = variance_type
-
-                pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
-                    pipeline.scheduler.config, **scheduler_args
-                )
-
-            # load attention processors
-            pipeline.load_lora_weights(args.output_dir)
-
-            # load new tokens
-            if args.train_text_encoder_ti:
-                state_dict = load_file(embeddings_path)
-                all_new_tokens = []
-                for key, value in token_abstraction_dict.items():
-                    all_new_tokens.extend(value)
-                pipeline.load_textual_inversion(
-                    state_dict["clip_l"],
-                    token=all_new_tokens,
-                    text_encoder=pipeline.text_encoder,
-                    tokenizer=pipeline.tokenizer,
-                )
-                pipeline.load_textual_inversion(
-                    state_dict["clip_g"],
-                    token=all_new_tokens,
-                    text_encoder=pipeline.text_encoder_2,
-                    tokenizer=pipeline.tokenizer_2,
-                )
-
-            # run inference
-            pipeline = pipeline.to(accelerator.device)
-            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-            images = [
-                pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
-                for _ in range(args.num_validation_images)
-            ]
-
-            for tracker in accelerator.trackers:
-                if tracker.name == "tensorboard":
-                    np_images = np.stack([np.asarray(img) for img in images])
-                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
-                if tracker.name == "wandb":
-                    tracker.log(
-                        {
-                            "test": [
-                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                for i, image in enumerate(images)
-                            ]
-                        }
-                    )
 
         # Convert to WebUI format
         lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
         peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
         kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
-        save_file(kohya_state_dict, f"{args.output_dir}/{args.output_dir}.safetensors")
+        save_file(kohya_state_dict, f"{args.output_dir}/{Path(args.output_dir).name}.safetensors")
 
         save_model_card(
             model_id if not args.push_to_hub else repo_id,
@@ -2430,6 +2442,7 @@ def main(args):
             repo_folder=args.output_dir,
             vae_path=args.pretrained_vae_model_name_or_path,
         )
+
         if args.push_to_hub:
             upload_folder(
                 repo_id=repo_id,

examples/community/README.md

@@ -70,6 +70,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Stable Diffusion XL IPEX Pipeline | Accelerate Stable Diffusion XL inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion XL on IPEX](#stable-diffusion-xl-on-ipex) | - | [Dan Li](https://github.com/ustcuna/) |
 | Stable Diffusion BoxDiff Pipeline | Training-free controlled generation with bounding boxes using [BoxDiff](https://github.com/showlab/BoxDiff) | [Stable Diffusion BoxDiff Pipeline](#stable-diffusion-boxdiff) | - | [Jingyang Zhang](https://github.com/zjysteven/) |
 |   FRESCO V2V Pipeline                                                                                                    | Implementation of [[CVPR 2024] FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation](https://arxiv.org/abs/2403.12962)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [FRESCO V2V Pipeline](#fresco)      | - |              [Yifan Zhou](https://github.com/SingleZombie) |
+| AnimateDiff IPEX Pipeline | Accelerate AnimateDiff inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [AnimateDiff on IPEX](#animatediff-on-ipex) | - | [Dan Li](https://github.com/ustcuna/) |
 
 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.
 
@@ -1640,18 +1641,18 @@ from io import BytesIO
 from PIL import Image
 import torch
 from diffusers import DDIMScheduler
-from diffusers.pipelines.stable_diffusion import StableDiffusionImg2ImgPipeline
+from diffusers import DiffusionPipeline
 
 # Use the DDIMScheduler scheduler here instead
 scheduler = DDIMScheduler.from_pretrained("stabilityai/stable-diffusion-2-1",
                                             subfolder="scheduler")
 
 
-pipe = StableDiffusionImg2ImgPipeline.from_pretrained("stabilityai/stable-diffusion-2-1",
-                                                custom_pipeline="stable_diffusion_tensorrt_img2img",
-                                                variant='fp16',
-                                                torch_dtype=torch.float16,
-                                                scheduler=scheduler,)
+pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1",
+                                            custom_pipeline="stable_diffusion_tensorrt_img2img",
+                                            variant='fp16',
+                                            torch_dtype=torch.float16,
+                                            scheduler=scheduler,)
 
 # re-use cached folder to save ONNX models and TensorRT Engines
 pipe.set_cached_folder("stabilityai/stable-diffusion-2-1", variant='fp16',)
@@ -4099,6 +4100,117 @@ output_frames[0].save(output_video_path, save_all=True,
                  append_images=output_frames[1:], duration=100, loop=0)
 ```
 
+### AnimateDiff on IPEX
+
+This diffusion pipeline aims to accelerate the inference of AnimateDiff on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
+
+To use this pipeline, you need to:
+1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
+
+**Note:** For each PyTorch release, there is a corresponding release of IPEX. Here is the mapping relationship. It is recommended to install Pytorch/IPEX2.3 to get the best performance.
+
+|PyTorch Version|IPEX Version|
+|--|--|
+|[v2.3.\*](https://github.com/pytorch/pytorch/tree/v2.3.0 "v2.3.0")|[v2.3.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v2.3.0+cpu)|
+|[v1.13.\*](https://github.com/pytorch/pytorch/tree/v1.13.0 "v1.13.0")|[v1.13.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v1.13.100+cpu)|
+
+You can simply use pip to install IPEX with the latest version.
+```python
+python -m pip install intel_extension_for_pytorch
+```
+**Note:** To install a specific version, run with the following command:
+```
+python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
+```
+2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX accelaration. Supported inference datatypes are Float32 and BFloat16.
+
+```python
+pipe = AnimateDiffPipelineIpex.from_pretrained(base, motion_adapter=adapter, torch_dtype=dtype).to(device)
+# For Float32
+pipe.prepare_for_ipex(torch.float32, prompt="A girl smiling")
+# For BFloat16
+pipe.prepare_for_ipex(torch.bfloat16, prompt="A girl smiling")
+```
+
+Then you can use the ipex pipeline in a similar way to the default animatediff pipeline.
+```python
+# For Float32
+output = pipe(prompt="A girl smiling", guidance_scale=1.0, num_inference_steps=step)
+# For BFloat16
+with torch.cpu.amp.autocast(enabled = True, dtype = torch.bfloat16):
+    output = pipe(prompt="A girl smiling", guidance_scale=1.0, num_inference_steps=step)
+```
+
+The following code compares the performance of the original animatediff pipeline with the ipex-optimized pipeline.
+By using this optimized pipeline, we can get about 1.5-2.2 times performance boost with BFloat16 on the fifth generation of Intel Xeon CPUs, code-named Emerald Rapids.
+
+```python
+import torch
+from diffusers import MotionAdapter, AnimateDiffPipeline, EulerDiscreteScheduler
+from safetensors.torch import load_file
+from pipeline_animatediff_ipex import AnimateDiffPipelineIpex
+import time
+
+device = "cpu"
+dtype = torch.float32
+
+prompt = "A girl smiling"
+step = 8  # Options: [1,2,4,8]
+repo = "ByteDance/AnimateDiff-Lightning"
+ckpt = f"animatediff_lightning_{step}step_diffusers.safetensors"
+base = "emilianJR/epiCRealism"  # Choose to your favorite base model.
+
+adapter = MotionAdapter().to(device, dtype)
+adapter.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device))
+
+# Helper function for time evaluation
+def elapsed_time(pipeline, nb_pass=3, num_inference_steps=1):
+    # warmup
+    for _ in range(2):
+        output = pipeline(prompt = prompt, guidance_scale=1.0, num_inference_steps = num_inference_steps)
+    #time evaluation
+    start = time.time()
+    for _ in range(nb_pass):
+        pipeline(prompt = prompt, guidance_scale=1.0, num_inference_steps = num_inference_steps)
+    end = time.time()
+    return (end - start) / nb_pass
+
+##############     bf16 inference performance    ###############
+
+# 1. IPEX Pipeline initialization
+pipe = AnimateDiffPipelineIpex.from_pretrained(base, motion_adapter=adapter, torch_dtype=dtype).to(device)
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
+pipe.prepare_for_ipex(torch.bfloat16, prompt = prompt)
+
+# 2. Original Pipeline initialization
+pipe2 = AnimateDiffPipeline.from_pretrained(base, motion_adapter=adapter, torch_dtype=dtype).to(device)
+pipe2.scheduler = EulerDiscreteScheduler.from_config(pipe2.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
+
+# 3. Compare performance between Original Pipeline and IPEX Pipeline
+with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
+    latency = elapsed_time(pipe, num_inference_steps=step)
+    print("Latency of AnimateDiffPipelineIpex--bf16", latency, "s for total", step, "steps")
+    latency = elapsed_time(pipe2, num_inference_steps=step)
+    print("Latency of AnimateDiffPipeline--bf16", latency, "s for total", step, "steps")
+
+##############     fp32 inference performance    ###############
+
+# 1. IPEX Pipeline initialization
+pipe3 = AnimateDiffPipelineIpex.from_pretrained(base, motion_adapter=adapter, torch_dtype=dtype).to(device)
+pipe3.scheduler = EulerDiscreteScheduler.from_config(pipe3.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
+pipe3.prepare_for_ipex(torch.float32, prompt = prompt)
+
+# 2. Original Pipeline initialization
+pipe4 = AnimateDiffPipeline.from_pretrained(base, motion_adapter=adapter, torch_dtype=dtype).to(device)
+pipe4.scheduler = EulerDiscreteScheduler.from_config(pipe4.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
+
+# 3. Compare performance between Original Pipeline and IPEX Pipeline
+latency = elapsed_time(pipe3, num_inference_steps=step)
+print("Latency of AnimateDiffPipelineIpex--fp32", latency, "s for total", step, "steps")
+latency = elapsed_time(pipe4, num_inference_steps=step)
+print("Latency of AnimateDiffPipeline--fp32",latency, "s for total", step, "steps")
+```
+
 # Perturbed-Attention Guidance
 
 [Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)

examples/community/checkpoint_merger.py

@@ -71,7 +71,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
             **kwargs:
                 Supports all the default DiffusionPipeline.get_config_dict kwargs viz..
 
-                cache_dir, resume_download, force_download, proxies, local_files_only, token, revision, torch_dtype, device_map.
+                cache_dir, force_download, proxies, local_files_only, token, revision, torch_dtype, device_map.
 
                 alpha - The interpolation parameter. Ranges from 0 to 1.  It affects the ratio in which the checkpoints are merged. A 0.8 alpha
                     would mean that the first model checkpoints would affect the final result far less than an alpha of 0.2
@@ -86,7 +86,6 @@ class CheckpointMergerPipeline(DiffusionPipeline):
         """
         # Default kwargs from DiffusionPipeline
         cache_dir = kwargs.pop("cache_dir", None)
-        resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
@@ -124,7 +123,6 @@ class CheckpointMergerPipeline(DiffusionPipeline):
             config_dict = DiffusionPipeline.load_config(
                 pretrained_model_name_or_path,
                 cache_dir=cache_dir,
-                resume_download=resume_download,
                 force_download=force_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
@@ -160,7 +158,6 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                 else snapshot_download(
                     pretrained_model_name_or_path,
                     cache_dir=cache_dir,
-                    resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
                     token=token,

examples/community/ip_adapter_face_id.py

@@ -267,7 +267,6 @@ class IPAdapterFaceIDStableDiffusionPipeline(
     def load_ip_adapter_face_id(self, pretrained_model_name_or_path_or_dict, weight_name, **kwargs):
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", None)
         token = kwargs.pop("token", None)
@@ -283,7 +282,6 @@ class IPAdapterFaceIDStableDiffusionPipeline(
             weights_name=weight_name,
             cache_dir=cache_dir,
             force_download=force_download,
-            resume_download=resume_download,
             proxies=proxies,
             local_files_only=local_files_only,
             token=token,

examples/community/pipeline_stable_diffusion_3_differential_img2img.py

@@ -0,0 +1,981 @@
+# Copyright 2024 Stability AI and 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 inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import (
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.models.transformers import SD3Transformer2DModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+
+        >>> from diffusers import AutoPipelineForImage2Image
+        >>> from diffusers.utils import load_image
+
+        >>> device = "cuda"
+        >>> model_id_or_path = "stabilityai/stable-diffusion-3-medium-diffusers"
+        >>> pipe = AutoPipelineForImage2Image.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
+        >>> pipe = pipe.to(device)
+
+        >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+        >>> init_image = load_image(url).resize((512, 512))
+
+        >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k"
+
+        >>> images = pipe(prompt=prompt, image=init_image, strength=0.95, guidance_scale=7.5).images[0]
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusion3DifferentialImg2ImgPipeline(DiffusionPipeline):
+    r"""
+    Args:
+        transformer ([`SD3Transformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant,
+            with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size`
+            as its dimension.
+        text_encoder_2 ([`CLIPTextModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        text_encoder_3 ([`T5EncoderModel`]):
+            Frozen text-encoder. Stable Diffusion 3 uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_3 (`T5TokenizerFast`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"]
+
+    def __init__(
+        self,
+        transformer: SD3Transformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer_2: CLIPTokenizer,
+        text_encoder_3: T5EncoderModel,
+        tokenizer_3: T5TokenizerFast,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            text_encoder_3=text_encoder_3,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            tokenizer_3=tokenizer_3,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, vae_latent_channels=self.vae.config.latent_channels
+        )
+        self.mask_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_convert_grayscale=True
+        )
+
+        self.tokenizer_max_length = self.tokenizer.model_max_length
+        self.default_sample_size = self.transformer.config.sample_size
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_images_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        if self.text_encoder_3 is None:
+            return torch.zeros(
+                (
+                    batch_size * num_images_per_prompt,
+                    self.tokenizer_max_length,
+                    self.transformer.config.joint_attention_dim,
+                ),
+                device=device,
+                dtype=dtype,
+            )
+
+        text_inputs = self.tokenizer_3(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0]
+
+        dtype = self.text_encoder_3.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        clip_skip: Optional[int] = None,
+        clip_model_index: int = 0,
+    ):
+        device = device or self._execution_device
+
+        clip_tokenizers = [self.tokenizer, self.tokenizer_2]
+        clip_text_encoders = [self.text_encoder, self.text_encoder_2]
+
+        tokenizer = clip_tokenizers[clip_model_index]
+        text_encoder = clip_text_encoders[clip_model_index]
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_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_max_length} tokens: {removed_text}"
+            )
+        prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+        pooled_prompt_embeds = prompt_embeds[0]
+
+        if clip_skip is None:
+            prompt_embeds = prompt_embeds.hidden_states[-2]
+        else:
+            prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1)
+
+        return prompt_embeds, pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]],
+        prompt_3: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        clip_skip: Optional[int] = None,
+        max_sequence_length: int = 256,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in all text-encoders
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                used in all text-encoders
+            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]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            prompt_3 = prompt_3 or prompt
+            prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3
+
+            prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=0,
+            )
+            prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                prompt=prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=clip_skip,
+                clip_model_index=1,
+            )
+            clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1)
+
+            t5_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            clip_prompt_embeds = torch.nn.functional.pad(
+                clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1])
+            )
+
+            prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2)
+            pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+            negative_prompt_3 = negative_prompt_3 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+            negative_prompt_3 = (
+                batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3
+            )
+
+            if prompt is not None and 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 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`."
+                )
+
+            negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds(
+                negative_prompt,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=0,
+            )
+            negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds(
+                negative_prompt_2,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                clip_skip=None,
+                clip_model_index=1,
+            )
+            negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1)
+
+            t5_negative_prompt_embed = self._get_t5_prompt_embeds(
+                prompt=negative_prompt_3,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+            negative_clip_prompt_embeds = torch.nn.functional.pad(
+                negative_clip_prompt_embeds,
+                (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]),
+            )
+
+            negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2)
+            negative_pooled_prompt_embeds = torch.cat(
+                [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1
+            )
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        prompt_3,
+        strength,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        negative_prompt_3=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_3 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (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)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+        elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)):
+            raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_3 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 512:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
+
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(num_inference_steps * strength, num_inference_steps)
+
+        t_start = int(max(num_inference_steps - init_timestep, 0))
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, height, width, image, timestep, dtype, device, generator=None
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+
+        image = image.to(device=device, dtype=dtype)
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+        elif isinstance(generator, list):
+            init_latents = [
+                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) for i in range(batch_size)
+            ]
+            init_latents = torch.cat(init_latents, dim=0)
+        else:
+            init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+        init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+
+        if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+            # expand init_latents for batch_size
+            additional_image_per_prompt = batch_size // init_latents.shape[0]
+            init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+        elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            init_latents = torch.cat([init_latents], dim=0)
+
+        shape = init_latents.shape
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+        init_latents = self.scheduler.scale_noise(init_latents, timestep, noise)
+        latents = init_latents.to(device=device, dtype=dtype)
+
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # 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.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        image: PipelineImageInput = None,
+        strength: float = 0.6,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_prompt_3: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        map: PipelineImageInput = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead
+            prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is
+                will be used instead
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            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.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                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. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used instead
+            negative_prompt_3 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and
+                `text_encoder_3`. If not defined, `negative_prompt` is used instead
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](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`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            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_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        # 0. Default height and width
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            prompt_3,
+            strength,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._clip_skip = clip_skip
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_3=prompt_3,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            negative_prompt_3=negative_prompt_3,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            clip_skip=self.clip_skip,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+        )
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
+
+        # 3. Preprocess image
+        init_image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+
+        map = self.mask_processor.preprocess(
+            map, height=height // self.vae_scale_factor, width=width // self.vae_scale_factor
+        ).to(device)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # begin diff diff change
+        total_time_steps = num_inference_steps
+        # end diff diff change
+
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        if latents is None:
+            latents = self.prepare_latents(
+                batch_size * num_images_per_prompt,
+                num_channels_latents,
+                height,
+                width,
+                init_image,
+                latent_timestep,
+                prompt_embeds.dtype,
+                device,
+                generator,
+            )
+
+        # 6. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # preparations for diff diff
+        original_with_noise = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            init_image,
+            timesteps,
+            prompt_embeds.dtype,
+            device,
+            generator,
+        )
+        thresholds = torch.arange(total_time_steps, dtype=map.dtype) / total_time_steps
+        thresholds = thresholds.unsqueeze(1).unsqueeze(1).to(device)
+        masks = map.squeeze() > thresholds
+        # end diff diff preparations
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # diff diff
+                if i == 0:
+                    latents = original_with_noise[:1]
+                else:
+                    mask = masks[i].unsqueeze(0).to(latents.dtype)
+                    mask = mask.unsqueeze(1)  # fit shape
+                    latents = original_with_noise[i] * mask + latents * (1 - mask)
+                # end diff diff
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    pooled_projections=pooled_prompt_embeds,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusion3PipelineOutput(images=image)

examples/community/regional_prompting_stable_diffusion.py

@@ -467,8 +467,6 @@ def make_emblist(self, prompts):
 
 
 def split_dims(xs, height, width):
-    xs = xs
-
     def repeat_div(x, y):
         while y > 0:
             x = math.ceil(x / 2)

examples/community/stable_diffusion_tensorrt_img2img.py

@@ -18,8 +18,7 @@
 import gc
 import os
 from collections import OrderedDict
-from copy import copy
-from typing import List, Optional, Union
+from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import onnx
@@ -27,9 +26,11 @@ import onnx_graphsurgeon as gs
 import PIL.Image
 import tensorrt as trt
 import torch
+from cuda import cudart
 from huggingface_hub import snapshot_download
 from huggingface_hub.utils import validate_hf_hub_args
 from onnx import shape_inference
+from packaging import version
 from polygraphy import cuda
 from polygraphy.backend.common import bytes_from_path
 from polygraphy.backend.onnx.loader import fold_constants
@@ -41,12 +42,13 @@ from polygraphy.backend.trt import (
     network_from_onnx_path,
     save_engine,
 )
-from polygraphy.backend.trt import util as trt_util
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
+from diffusers import DiffusionPipeline
+from diffusers.configuration_utils import FrozenDict, deprecate
+from diffusers.image_processor import VaeImageProcessor
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.pipelines.stable_diffusion import (
-    StableDiffusionImg2ImgPipeline,
     StableDiffusionPipelineOutput,
     StableDiffusionSafetyChecker,
 )
@@ -58,7 +60,7 @@ from diffusers.utils import logging
 """
 Installation instructions
 python3 -m pip install --upgrade transformers diffusers>=0.16.0
-python3 -m pip install --upgrade tensorrt>=8.6.1
+python3 -m pip install --upgrade tensorrt-cu12==10.2.0
 python3 -m pip install --upgrade polygraphy>=0.47.0 onnx-graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com
 python3 -m pip install onnxruntime
 """
@@ -88,10 +90,6 @@ else:
 torch_to_numpy_dtype_dict = {value: key for (key, value) in numpy_to_torch_dtype_dict.items()}
 
 
-def device_view(t):
-    return cuda.DeviceView(ptr=t.data_ptr(), shape=t.shape, dtype=torch_to_numpy_dtype_dict[t.dtype])
-
-
 def preprocess_image(image):
     """
     image: torch.Tensor
@@ -125,10 +123,8 @@ class Engine:
         onnx_path,
         fp16,
         input_profile=None,
-        enable_preview=False,
         enable_all_tactics=False,
         timing_cache=None,
-        workspace_size=0,
     ):
         logger.warning(f"Building TensorRT engine for {onnx_path}: {self.engine_path}")
         p = Profile()
@@ -137,20 +133,13 @@ class Engine:
                 assert len(dims) == 3
                 p.add(name, min=dims[0], opt=dims[1], max=dims[2])
 
-        config_kwargs = {}
-
-        config_kwargs["preview_features"] = [trt.PreviewFeature.DISABLE_EXTERNAL_TACTIC_SOURCES_FOR_CORE_0805]
-        if enable_preview:
-            # Faster dynamic shapes made optional since it increases engine build time.
-            config_kwargs["preview_features"].append(trt.PreviewFeature.FASTER_DYNAMIC_SHAPES_0805)
-        if workspace_size > 0:
-            config_kwargs["memory_pool_limits"] = {trt.MemoryPoolType.WORKSPACE: workspace_size}
+        extra_build_args = {}
         if not enable_all_tactics:
-            config_kwargs["tactic_sources"] = []
+            extra_build_args["tactic_sources"] = []
 
         engine = engine_from_network(
             network_from_onnx_path(onnx_path, flags=[trt.OnnxParserFlag.NATIVE_INSTANCENORM]),
-            config=CreateConfig(fp16=fp16, profiles=[p], load_timing_cache=timing_cache, **config_kwargs),
+            config=CreateConfig(fp16=fp16, profiles=[p], load_timing_cache=timing_cache, **extra_build_args),
             save_timing_cache=timing_cache,
         )
         save_engine(engine, path=self.engine_path)
@@ -163,28 +152,24 @@ class Engine:
         self.context = self.engine.create_execution_context()
 
     def allocate_buffers(self, shape_dict=None, device="cuda"):
-        for idx in range(trt_util.get_bindings_per_profile(self.engine)):
-            binding = self.engine[idx]
-            if shape_dict and binding in shape_dict:
-                shape = shape_dict[binding]
+        for binding in range(self.engine.num_io_tensors):
+            name = self.engine.get_tensor_name(binding)
+            if shape_dict and name in shape_dict:
+                shape = shape_dict[name]
             else:
-                shape = self.engine.get_binding_shape(binding)
-            dtype = trt.nptype(self.engine.get_binding_dtype(binding))
-            if self.engine.binding_is_input(binding):
-                self.context.set_binding_shape(idx, shape)
+                shape = self.engine.get_tensor_shape(name)
+            dtype = trt.nptype(self.engine.get_tensor_dtype(name))
+            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
+                self.context.set_input_shape(name, shape)
             tensor = torch.empty(tuple(shape), dtype=numpy_to_torch_dtype_dict[dtype]).to(device=device)
-            self.tensors[binding] = tensor
-            self.buffers[binding] = cuda.DeviceView(ptr=tensor.data_ptr(), shape=shape, dtype=dtype)
+            self.tensors[name] = tensor
 
     def infer(self, feed_dict, stream):
-        start_binding, end_binding = trt_util.get_active_profile_bindings(self.context)
-        # shallow copy of ordered dict
-        device_buffers = copy(self.buffers)
         for name, buf in feed_dict.items():
-            assert isinstance(buf, cuda.DeviceView)
-            device_buffers[name] = buf
-        bindings = [0] * start_binding + [buf.ptr for buf in device_buffers.values()]
-        noerror = self.context.execute_async_v2(bindings=bindings, stream_handle=stream.ptr)
+            self.tensors[name].copy_(buf)
+        for name, tensor in self.tensors.items():
+            self.context.set_tensor_address(name, tensor.data_ptr())
+        noerror = self.context.execute_async_v3(stream)
         if not noerror:
             raise ValueError("ERROR: inference failed.")
 
@@ -325,10 +310,8 @@ def build_engines(
     force_engine_rebuild=False,
     static_batch=False,
     static_shape=True,
-    enable_preview=False,
     enable_all_tactics=False,
     timing_cache=None,
-    max_workspace_size=0,
 ):
     built_engines = {}
     if not os.path.isdir(onnx_dir):
@@ -393,9 +376,7 @@ def build_engines(
                     static_batch=static_batch,
                     static_shape=static_shape,
                 ),
-                enable_preview=enable_preview,
                 timing_cache=timing_cache,
-                workspace_size=max_workspace_size,
             )
         built_engines[model_name] = engine
 
@@ -674,7 +655,7 @@ def make_VAEEncoder(model, device, max_batch_size, embedding_dim, inpaint=False)
     return VAEEncoder(model, device=device, max_batch_size=max_batch_size, embedding_dim=embedding_dim)
 
 
-class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
+class TensorRTStableDiffusionImg2ImgPipeline(DiffusionPipeline):
     r"""
     Pipeline for image-to-image generation using TensorRT accelerated Stable Diffusion.
 
@@ -702,6 +683,8 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+
     def __init__(
         self,
         vae: AutoencoderKL,
@@ -722,24 +705,86 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         onnx_dir: str = "onnx",
         # TensorRT engine build parameters
         engine_dir: str = "engine",
-        build_preview_features: bool = True,
         force_engine_rebuild: bool = False,
         timing_cache: str = "timing_cache",
     ):
-        super().__init__(
-            vae,
-            text_encoder,
-            tokenizer,
-            unet,
-            scheduler,
+        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, "clip_sample") and scheduler.config.clip_sample is True:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+                " config accordingly as not setting `clip_sample` 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("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["clip_sample"] = False
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                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 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+            version.parse(unet.config._diffusers_version).base_version
+        ) < version.parse("0.9.0.dev0")
+        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+            deprecation_message = (
+                "The configuration file of the unet has set the default `sample_size` to smaller than"
+                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+                " 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 `unet/config.json` file"
+            )
+            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(unet.config)
+            new_config["sample_size"] = 64
+            unet._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,
             image_encoder=image_encoder,
-            requires_safety_checker=requires_safety_checker,
         )
 
-        self.vae.forward = self.vae.decode
-
         self.stages = stages
         self.image_height, self.image_width = image_height, image_width
         self.inpaint = False
@@ -750,7 +795,6 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         self.timing_cache = timing_cache
         self.build_static_batch = False
         self.build_dynamic_shape = False
-        self.build_preview_features = build_preview_features
 
         self.max_batch_size = max_batch_size
         # TODO: Restrict batch size to 4 for larger image dimensions as a WAR for TensorRT limitation.
@@ -761,6 +805,11 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         self.models = {}  # loaded in __loadModels()
         self.engine = {}  # loaded in build_engines()
 
+        self.vae.forward = self.vae.decode
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
+
     def __loadModels(self):
         # Load pipeline models
         self.embedding_dim = self.text_encoder.config.hidden_size
@@ -779,11 +828,37 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         if "vae_encoder" in self.stages:
             self.models["vae_encoder"] = make_VAEEncoder(self.vae, **models_args)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+    def run_safety_checker(
+        self, image: Union[torch.Tensor, PIL.Image.Image], device: torch.device, dtype: torch.dtype
+    ) -> Tuple[Union[torch.Tensor, PIL.Image.Image], Optional[bool]]:
+        r"""
+        Runs the safety checker on the given image.
+        Args:
+            image (Union[torch.Tensor, PIL.Image.Image]): The input image to be checked.
+            device (torch.device): The device to run the safety checker on.
+            dtype (torch.dtype): The data type of the input image.
+        Returns:
+            (image, has_nsfw_concept) Tuple[Union[torch.Tensor, PIL.Image.Image], Optional[bool]]: A tuple containing the processed image and
+            a boolean indicating whether the image has a NSFW (Not Safe for Work) concept.
+        """
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
     @classmethod
     @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         cache_dir = kwargs.pop("cache_dir", None)
-        resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
         token = kwargs.pop("token", None)
@@ -795,7 +870,6 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             else snapshot_download(
                 pretrained_model_name_or_path,
                 cache_dir=cache_dir,
-                resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
                 token=token,
@@ -828,7 +902,6 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             force_engine_rebuild=self.force_engine_rebuild,
             static_batch=self.build_static_batch,
             static_shape=not self.build_dynamic_shape,
-            enable_preview=self.build_preview_features,
             timing_cache=self.timing_cache,
         )
 
@@ -852,9 +925,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         return tuple(init_images)
 
     def __encode_image(self, init_image):
-        init_latents = runEngine(self.engine["vae_encoder"], {"images": device_view(init_image)}, self.stream)[
-            "latent"
-        ]
+        init_latents = runEngine(self.engine["vae_encoder"], {"images": init_image}, self.stream)["latent"]
         init_latents = 0.18215 * init_latents
         return init_latents
 
@@ -883,9 +954,8 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             .to(self.torch_device)
         )
 
-        text_input_ids_inp = device_view(text_input_ids)
         # NOTE: output tensor for CLIP must be cloned because it will be overwritten when called again for negative prompt
-        text_embeddings = runEngine(self.engine["clip"], {"input_ids": text_input_ids_inp}, self.stream)[
+        text_embeddings = runEngine(self.engine["clip"], {"input_ids": text_input_ids}, self.stream)[
             "text_embeddings"
         ].clone()
 
@@ -901,8 +971,7 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             .input_ids.type(torch.int32)
             .to(self.torch_device)
         )
-        uncond_input_ids_inp = device_view(uncond_input_ids)
-        uncond_embeddings = runEngine(self.engine["clip"], {"input_ids": uncond_input_ids_inp}, self.stream)[
+        uncond_embeddings = runEngine(self.engine["clip"], {"input_ids": uncond_input_ids}, self.stream)[
             "text_embeddings"
         ]
 
@@ -926,18 +995,15 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             # Predict the noise residual
             timestep_float = timestep.float() if timestep.dtype != torch.float32 else timestep
 
-            sample_inp = device_view(latent_model_input)
-            timestep_inp = device_view(timestep_float)
-            embeddings_inp = device_view(text_embeddings)
             noise_pred = runEngine(
                 self.engine["unet"],
-                {"sample": sample_inp, "timestep": timestep_inp, "encoder_hidden_states": embeddings_inp},
+                {"sample": latent_model_input, "timestep": timestep_float, "encoder_hidden_states": text_embeddings},
                 self.stream,
             )["latent"]
 
             # Perform guidance
             noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-            noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+            noise_pred = noise_pred_uncond + self._guidance_scale * (noise_pred_text - noise_pred_uncond)
 
             latents = self.scheduler.step(noise_pred, timestep, latents).prev_sample
 
@@ -945,12 +1011,12 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
         return latents
 
     def __decode_latent(self, latents):
-        images = runEngine(self.engine["vae"], {"latent": device_view(latents)}, self.stream)["images"]
+        images = runEngine(self.engine["vae"], {"latent": latents}, self.stream)["images"]
         images = (images / 2 + 0.5).clamp(0, 1)
         return images.cpu().permute(0, 2, 3, 1).float().numpy()
 
     def __loadResources(self, image_height, image_width, batch_size):
-        self.stream = cuda.Stream()
+        self.stream = cudart.cudaStreamCreate()[1]
 
         # Allocate buffers for TensorRT engine bindings
         for model_name, obj in self.models.items():
@@ -1063,5 +1129,6 @@ class TensorRTStableDiffusionImg2ImgPipeline(StableDiffusionImg2ImgPipeline):
             # VAE decode latent
             images = self.__decode_latent(latents)
 
+        images, has_nsfw_concept = self.run_safety_checker(images, self.torch_device, text_embeddings.dtype)
         images = self.numpy_to_pil(images)
-        return StableDiffusionPipelineOutput(images=images, nsfw_content_detected=None)
+        return StableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept)

examples/community/stable_diffusion_tensorrt_inpaint.py

@@ -783,7 +783,6 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
     @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         cache_dir = kwargs.pop("cache_dir", None)
-        resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
         token = kwargs.pop("token", None)
@@ -795,7 +794,6 @@ class TensorRTStableDiffusionInpaintPipeline(StableDiffusionInpaintPipeline):
             else snapshot_download(
                 pretrained_model_name_or_path,
                 cache_dir=cache_dir,
-                resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
                 token=token,

examples/community/stable_diffusion_tensorrt_txt2img.py

@@ -695,7 +695,6 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
     @validate_hf_hub_args
     def set_cached_folder(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
         cache_dir = kwargs.pop("cache_dir", None)
-        resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
         token = kwargs.pop("token", None)
@@ -707,7 +706,6 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
             else snapshot_download(
                 pretrained_model_name_or_path,
                 cache_dir=cache_dir,
-                resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
                 token=token,

examples/controlnet/train_controlnet_sdxl.py

@@ -1088,17 +1088,22 @@ def main(args):
     )
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     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
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1110,8 +1115,14 @@ def main(args):
 
     # 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:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 

examples/dreambooth/README_sd3.md

@@ -183,4 +183,6 @@ accelerate launch train_dreambooth_lora_sd3.py \
 
 ## Other notes
 
-We default to the "logit_normal" weighting scheme for the loss following the SD3 paper. Thanks to @bghira for helping us discover that for other weighting schemes supported from the training script, training may incur numerical instabilities.
+1. We default to the "logit_normal" weighting scheme for the loss following the SD3 paper. Thanks to @bghira for helping us discover that for other weighting schemes supported from the training script, training may incur numerical instabilities.
+2. Thanks to `bghira`, `JinxuXiang`, and `bendanzzc` for helping us discover a bug in how VAE encoding was being done previously. This has been fixed in [#8917](https://github.com/huggingface/diffusers/pull/8917).
+3. Additionally, we now have the option to control if we want to apply preconditioning to the model outputs via a `--precondition_outputs` CLI arg. It affects how the model `target` is calculated as well. 

examples/dreambooth/test_dreambooth_lora_sd3.py

@@ -0,0 +1,165 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 logging
+import os
+import sys
+import tempfile
+
+import safetensors
+
+
+sys.path.append("..")
+from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+
+
+class DreamBoothLoRASD3(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    instance_prompt = "photo"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-sd3-pipe"
+    script_path = "examples/dreambooth/train_dreambooth_lora_sd3.py"
+
+    def test_dreambooth_lora_sd3(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_text_encoder_sd3(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --train_text_encoder
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            starts_with_expected_prefix = all(
+                (key.startswith("transformer") or key.startswith("text_encoder")) for key in lora_state_dict.keys()
+            )
+            self.assertTrue(starts_with_expected_prefix)
+
+    def test_dreambooth_lora_sd3_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_lora_sd3_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-2", "checkpoint-4"})
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})

examples/dreambooth/test_dreambooth_sd3.py

@@ -0,0 +1,203 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 logging
+import os
+import shutil
+import sys
+import tempfile
+
+from diffusers import DiffusionPipeline, SD3Transformer2DModel
+
+
+sys.path.append("..")
+from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+
+
+class DreamBoothSD3(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    instance_prompt = "photo"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-sd3-pipe"
+    script_path = "examples/dreambooth/train_dreambooth_sd3.py"
+
+    def test_dreambooth(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "transformer", "diffusion_pytorch_model.safetensors")))
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))
+
+    def test_dreambooth_checkpointing(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            # Run training script with checkpointing
+            # max_train_steps == 4, checkpointing_steps == 2
+            # Should create checkpoints at steps 2, 4
+
+            initial_run_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 4
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --checkpointing_steps=2
+                --seed=0
+                """.split()
+
+            run_command(self._launch_args + initial_run_args)
+
+            # check can run the original fully trained output pipeline
+            pipe = DiffusionPipeline.from_pretrained(tmpdir)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # check checkpoint directories exist
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-2")))
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-4")))
+
+            # check can run an intermediate checkpoint
+            transformer = SD3Transformer2DModel.from_pretrained(tmpdir, subfolder="checkpoint-2/transformer")
+            pipe = DiffusionPipeline.from_pretrained(self.pretrained_model_name_or_path, transformer=transformer)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # Remove checkpoint 2 so that we can check only later checkpoints exist after resuming
+            shutil.rmtree(os.path.join(tmpdir, "checkpoint-2"))
+
+            # Run training script for 7 total steps resuming from checkpoint 4
+
+            resume_run_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 6
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                --checkpointing_steps=2
+                --resume_from_checkpoint=checkpoint-4
+                --seed=0
+                """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            # check can run new fully trained pipeline
+            pipe = DiffusionPipeline.from_pretrained(tmpdir)
+            pipe(self.instance_prompt, num_inference_steps=1)
+
+            # check old checkpoints do not exist
+            self.assertFalse(os.path.isdir(os.path.join(tmpdir, "checkpoint-2")))
+
+            # check new checkpoints exist
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-4")))
+            self.assertTrue(os.path.isdir(os.path.join(tmpdir, "checkpoint-6")))
+
+    def test_dreambooth_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-2", "checkpoint-4"},
+            )
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -101,19 +101,37 @@ def save_model_card(
 
 ## Model description
 
-These are {repo_id} DreamBooth weights for {base_model}.
+These are {repo_id} DreamBooth LoRA weights for {base_model}.
 
-The weights were trained  using [DreamBooth](https://dreambooth.github.io/).
+The weights were trained using [DreamBooth](https://dreambooth.github.io/) with the [SD3 diffusers trainer](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sd3.md).
 
-LoRA for the text encoder was enabled: {train_text_encoder}.
+Was LoRA for the text encoder enabled? {train_text_encoder}.
 
 ## Trigger words
 
-You should use {instance_prompt} to trigger the image generation.
+You should use `{instance_prompt}` to trigger the image generation.
 
 ## Download model
 
-[Download]({repo_id}/tree/main) them in the Files & versions tab.
+[Download the *.safetensors LoRA]({repo_id}/tree/main) in the Files & versions tab.
+
+## Use it with the [🧨 diffusers library](https://github.com/huggingface/diffusers)
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+pipeline = AutoPipelineForText2Image.from_pretrained('stabilityai/stable-diffusion-3-medium-diffusers', torch_dtype=torch.float16).to('cuda')
+pipeline.load_lora_weights('{repo_id}', weight_name='pytorch_lora_weights.safetensors')
+image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
+```
+
+### Use it with UIs such as AUTOMATIC1111, Comfy UI, SD.Next, Invoke
+
+- **LoRA**: download **[`diffusers_lora_weights.safetensors` here 💾](/{repo_id}/blob/main/diffusers_lora_weights.safetensors)**.
+    - Rename it and place it on your `models/Lora` folder.
+    - On AUTOMATIC1111, load the LoRA by adding `<lora:your_new_name:1>` to your prompt. On ComfyUI just [load it as a regular LoRA](https://comfyanonymous.github.io/ComfyUI_examples/lora/).
+
+For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
 
 ## License
 
@@ -505,6 +523,13 @@ def parse_args(input_args=None):
         default=1.29,
         help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
     )
+    parser.add_argument(
+        "--precondition_outputs",
+        type=int,
+        default=1,
+        help="Flag indicating if we are preconditioning the model outputs or not as done in EDM. This affects how "
+        "model `target` is calculated.",
+    )
     parser.add_argument(
         "--optimizer",
         type=str,
@@ -962,7 +987,7 @@ def encode_prompt(
             prompt=prompt,
             device=device if device is not None else text_encoder.device,
             num_images_per_prompt=num_images_per_prompt,
-            text_input_ids=text_input_ids_list[i],
+            text_input_ids=text_input_ids_list[i] if text_input_ids_list else None,
         )
         clip_prompt_embeds_list.append(prompt_embeds)
         clip_pooled_prompt_embeds_list.append(pooled_prompt_embeds)
@@ -976,7 +1001,7 @@ def encode_prompt(
         max_sequence_length,
         prompt=prompt,
         num_images_per_prompt=num_images_per_prompt,
-        text_input_ids=text_input_ids_list[:-1],
+        text_input_ids=text_input_ids_list[-1] if text_input_ids_list else None,
         device=device if device is not None else text_encoders[-1].device,
     )
 
@@ -1491,6 +1516,9 @@ def main(args):
         ) = accelerator.prepare(
             transformer, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
         )
+        assert text_encoder_one is not None
+        assert text_encoder_two is not None
+        assert text_encoder_three is not None
     else:
         transformer, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
             transformer, optimizer, train_dataloader, lr_scheduler
@@ -1598,7 +1626,7 @@ def main(args):
                         tokens_three = tokenize_prompt(tokenizer_three, prompts)
                         prompt_embeds, pooled_prompt_embeds = encode_prompt(
                             text_encoders=[text_encoder_one, text_encoder_two, text_encoder_three],
-                            tokenizers=[None, None, tokenizer_three],
+                            tokenizers=[None, None, None],
                             prompt=prompts,
                             max_sequence_length=args.max_sequence_length,
                             text_input_ids_list=[tokens_one, tokens_two, tokens_three],
@@ -1608,14 +1636,14 @@ def main(args):
                         prompt_embeds, pooled_prompt_embeds = encode_prompt(
                             text_encoders=[text_encoder_one, text_encoder_two, text_encoder_three],
                             tokenizers=[None, None, tokenizer_three],
-                            prompt=prompts,
+                            prompt=args.instance_prompt,
                             max_sequence_length=args.max_sequence_length,
                             text_input_ids_list=[tokens_one, tokens_two, tokens_three],
                         )
 
                 # Convert images to latent space
                 model_input = vae.encode(pixel_values).latent_dist.sample()
-                model_input = model_input * vae.config.scaling_factor
+                model_input = (model_input - vae.config.shift_factor) * vae.config.scaling_factor
                 model_input = model_input.to(dtype=weight_dtype)
 
                 # Sample noise that we'll add to the latents
@@ -1635,8 +1663,9 @@ def main(args):
                 timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)
 
                 # Add noise according to flow matching.
+                # zt = (1 - texp) * x + texp * z1
                 sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
-                noisy_model_input = sigmas * noise + (1.0 - sigmas) * model_input
+                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
 
                 # Predict the noise residual
                 model_pred = transformer(
@@ -1649,14 +1678,18 @@ def main(args):
 
                 # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
                 # Preconditioning of the model outputs.
-                model_pred = model_pred * (-sigmas) + noisy_model_input
+                if args.precondition_outputs:
+                    model_pred = model_pred * (-sigmas) + noisy_model_input
 
                 # these weighting schemes use a uniform timestep sampling
                 # and instead post-weight the loss
                 weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas)
 
                 # flow matching loss
-                target = model_input
+                if args.precondition_outputs:
+                    target = model_input
+                else:
+                    target = noise - model_input
 
                 if args.with_prior_preservation:
                     # Chunk the noise and model_pred into two parts and compute the loss on each part separately.
@@ -1685,10 +1718,12 @@ def main(args):
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    params_to_clip = itertools.chain(
-                        transformer_lora_parameters,
-                        text_lora_parameters_one,
-                        text_lora_parameters_two if args.train_text_encoder else transformer_lora_parameters,
+                    params_to_clip = (
+                        itertools.chain(
+                            transformer_lora_parameters, text_lora_parameters_one, text_lora_parameters_two
+                        )
+                        if args.train_text_encoder
+                        else transformer_lora_parameters
                     )
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
 
@@ -1741,13 +1776,6 @@ def main(args):
                     text_encoder_one, text_encoder_two, text_encoder_three = load_text_encoders(
                         text_encoder_cls_one, text_encoder_cls_two, text_encoder_cls_three
                     )
-                else:
-                    text_encoder_three = text_encoder_cls_three.from_pretrained(
-                        args.pretrained_model_name_or_path,
-                        subfolder="text_encoder_3",
-                        revision=args.revision,
-                        variant=args.variant,
-                    )
                 pipeline = StableDiffusion3Pipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
                     vae=vae,
@@ -1767,7 +1795,9 @@ def main(args):
                     pipeline_args=pipeline_args,
                     epoch=epoch,
                 )
-                del text_encoder_one, text_encoder_two, text_encoder_three
+                if not args.train_text_encoder:
+                    del text_encoder_one, text_encoder_two, text_encoder_three
+
                 torch.cuda.empty_cache()
                 gc.collect()
 

examples/dreambooth/train_dreambooth_sd3.py

@@ -95,17 +95,22 @@ def save_model_card(
 
 These are {repo_id} DreamBooth weights for {base_model}.
 
-The weights were trained  using [DreamBooth](https://dreambooth.github.io/).
+The weights were trained using [DreamBooth](https://dreambooth.github.io/) with the [SD3 diffusers trainer](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sd3.md).
 
-Text encoder was fine-tuned: {train_text_encoder}.
+Was the text encoder fine-tuned? {train_text_encoder}.
 
 ## Trigger words
 
-You should use {instance_prompt} to trigger the image generation.
+You should use `{instance_prompt}` to trigger the image generation.
 
-## Download model
+## Use it with the [🧨 diffusers library](https://github.com/huggingface/diffusers)
 
-[Download]({repo_id}/tree/main) them in the Files & versions tab.
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+pipeline = AutoPipelineForText2Image.from_pretrained('{repo_id}', torch_dtype=torch.float16).to('cuda')
+image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
+```
 
 ## License
 
@@ -489,6 +494,13 @@ def parse_args(input_args=None):
         default=1.29,
         help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
     )
+    parser.add_argument(
+        "--precondition_outputs",
+        type=int,
+        default=1,
+        help="Flag indicating if we are preconditioning the model outputs or not as done in EDM. This affects how "
+        "model `target` is calculated.",
+    )
     parser.add_argument(
         "--optimizer",
         type=str,
@@ -1544,7 +1556,7 @@ def main(args):
 
                 # Convert images to latent space
                 model_input = vae.encode(pixel_values).latent_dist.sample()
-                model_input = model_input * vae.config.scaling_factor
+                model_input = (model_input - vae.config.shift_factor) * vae.config.scaling_factor
                 model_input = model_input.to(dtype=weight_dtype)
 
                 # Sample noise that we'll add to the latents
@@ -1564,8 +1576,9 @@ def main(args):
                 timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)
 
                 # Add noise according to flow matching.
+                # zt = (1 - texp) * x + texp * z1
                 sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
-                noisy_model_input = sigmas * noise + (1.0 - sigmas) * model_input
+                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
 
                 # Predict the noise residual
                 if not args.train_text_encoder:
@@ -1593,13 +1606,18 @@ def main(args):
 
                 # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
                 # Preconditioning of the model outputs.
-                model_pred = model_pred * (-sigmas) + noisy_model_input
+                if args.precondition_outputs:
+                    model_pred = model_pred * (-sigmas) + noisy_model_input
+
                 # these weighting schemes use a uniform timestep sampling
                 # and instead post-weight the loss
                 weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas)
 
                 # flow matching loss
-                target = model_input
+                if args.precondition_outputs:
+                    target = model_input
+                else:
+                    target = noise - model_input
 
                 if args.with_prior_preservation:
                     # Chunk the noise and model_pred into two parts and compute the loss on each part separately.

examples/research_projects/consistency_training/train_cm_ct_unconditional.py

@@ -1195,7 +1195,7 @@ def main(args):
 
     # Resolve the c parameter for the Pseudo-Huber loss
     if args.huber_c is None:
-        args.huber_c = 0.00054 * args.resolution * math.sqrt(unet.config.in_channels)
+        args.huber_c = 0.00054 * args.resolution * math.sqrt(unwrap_model(unet).config.in_channels)
 
     # Get current number of discretization steps N according to our discretization curriculum
     current_discretization_steps = get_discretization_steps(

examples/research_projects/sd3_lora_colab/README.md

@@ -0,0 +1,38 @@
+# Running Stable Diffusion 3 DreamBooth LoRA training under 16GB
+
+This is an **EDUCATIONAL** project that provides utilities for DreamBooth LoRA training for [Stable Diffusion 3 (SD3)](ttps://huggingface.co/papers/2403.03206) under 16GB GPU VRAM. This means you can successfully try out this project using a [free-tier Colab Notebook](https://colab.research.google.com/github/huggingface/diffusers/blob/main/examples/research_projects/sd3_lora_colab/sd3_dreambooth_lora_16gb.ipynb) instance. 🤗
+
+> [!NOTE]
+> SD3 is gated, so you need to make sure you agree to [share your contact info](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) to access the model before using it with Diffusers. Once you have access, you need to log in so your system knows you’re authorized. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+For setup, inference code, and details on how to run the code, please follow the Colab Notebook provided above. 
+
+## How
+
+We make use of several techniques to make this possible:
+
+* Compute the embeddings from the instance prompt and serialize them for later reuse. This is implemented in the [`compute_embeddings.py`](./compute_embeddings.py) script. We use an 8bit (as introduced in [`LLM.int8()`](https://arxiv.org/abs/2208.07339)) T5 to reduce memory requirements to ~10.5GB. 
+* In the `train_dreambooth_sd3_lora_miniature.py` script, we make use of:
+  * 8bit Adam for optimization through the `bitsandbytes` library.
+  * Gradient checkpointing and gradient accumulation.
+  * FP16 precision.
+  * Flash attention through `F.scaled_dot_product_attention()`. 
+
+Computing the text embeddings is arguably the most memory-intensive part in the pipeline as SD3 employs three text encoders. If we run them in FP32, it will take about 20GB of VRAM. With FP16, we are down to 12GB. 
+
+
+## Gotchas
+
+This project is educational. It exists to showcase the possibility of fine-tuning a big diffusion system on consumer GPUs. But additional components might have to be added to obtain state-of-the-art performance. Below are some commonly known gotchas that users should be aware of:
+
+* Training of text encoders is purposefully disabled. 
+* Techniques such as prior-preservation is unsupported. 
+* Custom instance captions for instance images are unsupported, but this should be relatively easy to integrate.
+
+Hopefully, this project gives you a template to extend it further to suit your needs.

examples/research_projects/sd3_lora_colab/compute_embeddings.py

@@ -0,0 +1,123 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. 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 argparse
+import glob
+import hashlib
+
+import pandas as pd
+import torch
+from transformers import T5EncoderModel
+
+from diffusers import StableDiffusion3Pipeline
+
+
+PROMPT = "a photo of sks dog"
+MAX_SEQ_LENGTH = 77
+LOCAL_DATA_DIR = "dog"
+OUTPUT_PATH = "sample_embeddings.parquet"
+
+
+def bytes_to_giga_bytes(bytes):
+    return bytes / 1024 / 1024 / 1024
+
+
+def generate_image_hash(image_path):
+    with open(image_path, "rb") as f:
+        img_data = f.read()
+    return hashlib.sha256(img_data).hexdigest()
+
+
+def load_sd3_pipeline():
+    id = "stabilityai/stable-diffusion-3-medium-diffusers"
+    text_encoder = T5EncoderModel.from_pretrained(id, subfolder="text_encoder_3", load_in_8bit=True, device_map="auto")
+    pipeline = StableDiffusion3Pipeline.from_pretrained(
+        id, text_encoder_3=text_encoder, transformer=None, vae=None, device_map="balanced"
+    )
+    return pipeline
+
+
+@torch.no_grad()
+def compute_embeddings(pipeline, prompt, max_sequence_length):
+    (
+        prompt_embeds,
+        negative_prompt_embeds,
+        pooled_prompt_embeds,
+        negative_pooled_prompt_embeds,
+    ) = pipeline.encode_prompt(prompt=prompt, prompt_2=None, prompt_3=None, max_sequence_length=max_sequence_length)
+
+    print(
+        f"{prompt_embeds.shape=}, {negative_prompt_embeds.shape=}, {pooled_prompt_embeds.shape=}, {negative_pooled_prompt_embeds.shape}"
+    )
+
+    max_memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
+    print(f"Max memory allocated: {max_memory:.3f} GB")
+    return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+
+def run(args):
+    pipeline = load_sd3_pipeline()
+    prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = compute_embeddings(
+        pipeline, args.prompt, args.max_sequence_length
+    )
+
+    # Assumes that the images within `args.local_image_dir` have a JPEG extension. Change
+    # as needed.
+    image_paths = glob.glob(f"{args.local_data_dir}/*.jpeg")
+    data = []
+    for image_path in image_paths:
+        img_hash = generate_image_hash(image_path)
+        data.append(
+            (img_hash, prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds)
+        )
+
+    # Create a DataFrame
+    embedding_cols = [
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "pooled_prompt_embeds",
+        "negative_pooled_prompt_embeds",
+    ]
+    df = pd.DataFrame(
+        data,
+        columns=["image_hash"] + embedding_cols,
+    )
+
+    # Convert embedding lists to arrays (for proper storage in parquet)
+    for col in embedding_cols:
+        df[col] = df[col].apply(lambda x: x.cpu().numpy().flatten().tolist())
+
+    # Save the dataframe to a parquet file
+    df.to_parquet(args.output_path)
+    print(f"Data successfully serialized to {args.output_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--prompt", type=str, default=PROMPT, help="The instance prompt.")
+    parser.add_argument(
+        "--max_sequence_length",
+        type=int,
+        default=MAX_SEQ_LENGTH,
+        help="Maximum sequence length to use for computing the embeddings. The more the higher computational costs.",
+    )
+    parser.add_argument(
+        "--local_data_dir", type=str, default=LOCAL_DATA_DIR, help="Path to the directory containing instance images."
+    )
+    parser.add_argument("--output_path", type=str, default=OUTPUT_PATH, help="Path to serialize the parquet file.")
+    args = parser.parse_args()
+
+    run(args)

examples/research_projects/vae/README.md

@@ -0,0 +1,11 @@
+# VAE
+
+`vae_roundtrip.py` Demonstrates the use of a VAE by roundtripping an image through the encoder and decoder. Original and reconstructed images are displayed side by side.
+
+```
+cd examples/research_projects/vae
+python vae_roundtrip.py \
+    --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
+    --subfolder="vae" \
+    --input_image="/path/to/your/input.png"
+```

examples/research_projects/vae/vae_roundtrip.py

@@ -0,0 +1,282 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. 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
+
+import argparse
+import typing
+from typing import Optional, Union
+
+import torch
+from PIL import Image
+from torchvision import transforms  # type: ignore
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models.autoencoders.autoencoder_kl import (
+    AutoencoderKL,
+    AutoencoderKLOutput,
+)
+from diffusers.models.autoencoders.autoencoder_tiny import (
+    AutoencoderTiny,
+    AutoencoderTinyOutput,
+)
+from diffusers.models.autoencoders.vae import DecoderOutput
+
+
+SupportedAutoencoder = Union[AutoencoderKL, AutoencoderTiny]
+
+
+def load_vae_model(
+    *,
+    device: torch.device,
+    model_name_or_path: str,
+    revision: Optional[str],
+    variant: Optional[str],
+    # NOTE: use subfolder="vae" if the pointed model is for stable diffusion as a whole instead of just the VAE
+    subfolder: Optional[str],
+    use_tiny_nn: bool,
+) -> SupportedAutoencoder:
+    if use_tiny_nn:
+        # NOTE: These scaling factors don't have to be the same as each other.
+        down_scale = 2
+        up_scale = 2
+        vae = AutoencoderTiny.from_pretrained(  # type: ignore
+            model_name_or_path,
+            subfolder=subfolder,
+            revision=revision,
+            variant=variant,
+            downscaling_scaling_factor=down_scale,
+            upsampling_scaling_factor=up_scale,
+        )
+        assert isinstance(vae, AutoencoderTiny)
+    else:
+        vae = AutoencoderKL.from_pretrained(  # type: ignore
+            model_name_or_path,
+            subfolder=subfolder,
+            revision=revision,
+            variant=variant,
+        )
+        assert isinstance(vae, AutoencoderKL)
+    vae = vae.to(device)
+    vae.eval()  # Set the model to inference mode
+    return vae
+
+
+def pil_to_nhwc(
+    *,
+    device: torch.device,
+    image: Image.Image,
+) -> torch.Tensor:
+    assert image.mode == "RGB"
+    transform = transforms.ToTensor()
+    nhwc = transform(image).unsqueeze(0).to(device)  # type: ignore
+    assert isinstance(nhwc, torch.Tensor)
+    return nhwc
+
+
+def nhwc_to_pil(
+    *,
+    nhwc: torch.Tensor,
+) -> Image.Image:
+    assert nhwc.shape[0] == 1
+    hwc = nhwc.squeeze(0).cpu()
+    return transforms.ToPILImage()(hwc)  # type: ignore
+
+
+def concatenate_images(
+    *,
+    left: Image.Image,
+    right: Image.Image,
+    vertical: bool = False,
+) -> Image.Image:
+    width1, height1 = left.size
+    width2, height2 = right.size
+    if vertical:
+        total_height = height1 + height2
+        max_width = max(width1, width2)
+        new_image = Image.new("RGB", (max_width, total_height))
+        new_image.paste(left, (0, 0))
+        new_image.paste(right, (0, height1))
+    else:
+        total_width = width1 + width2
+        max_height = max(height1, height2)
+        new_image = Image.new("RGB", (total_width, max_height))
+        new_image.paste(left, (0, 0))
+        new_image.paste(right, (width1, 0))
+    return new_image
+
+
+def to_latent(
+    *,
+    rgb_nchw: torch.Tensor,
+    vae: SupportedAutoencoder,
+) -> torch.Tensor:
+    rgb_nchw = VaeImageProcessor.normalize(rgb_nchw)  # type: ignore
+    encoding_nchw = vae.encode(typing.cast(torch.FloatTensor, rgb_nchw))
+    if isinstance(encoding_nchw, AutoencoderKLOutput):
+        latent = encoding_nchw.latent_dist.sample()  # type: ignore
+        assert isinstance(latent, torch.Tensor)
+    elif isinstance(encoding_nchw, AutoencoderTinyOutput):
+        latent = encoding_nchw.latents
+        do_internal_vae_scaling = False  # Is this needed?
+        if do_internal_vae_scaling:
+            latent = vae.scale_latents(latent).mul(255).round().byte()  # type: ignore
+            latent = vae.unscale_latents(latent / 255.0)  # type: ignore
+            assert isinstance(latent, torch.Tensor)
+    else:
+        assert False, f"Unknown encoding type: {type(encoding_nchw)}"
+    return latent
+
+
+def from_latent(
+    *,
+    latent_nchw: torch.Tensor,
+    vae: SupportedAutoencoder,
+) -> torch.Tensor:
+    decoding_nchw = vae.decode(latent_nchw)  # type: ignore
+    assert isinstance(decoding_nchw, DecoderOutput)
+    rgb_nchw = VaeImageProcessor.denormalize(decoding_nchw.sample)  # type: ignore
+    assert isinstance(rgb_nchw, torch.Tensor)
+    return rgb_nchw
+
+
+def main_kwargs(
+    *,
+    device: torch.device,
+    input_image_path: str,
+    pretrained_model_name_or_path: str,
+    revision: Optional[str],
+    variant: Optional[str],
+    subfolder: Optional[str],
+    use_tiny_nn: bool,
+) -> None:
+    vae = load_vae_model(
+        device=device,
+        model_name_or_path=pretrained_model_name_or_path,
+        revision=revision,
+        variant=variant,
+        subfolder=subfolder,
+        use_tiny_nn=use_tiny_nn,
+    )
+    original_pil = Image.open(input_image_path).convert("RGB")
+    original_image = pil_to_nhwc(
+        device=device,
+        image=original_pil,
+    )
+    print(f"Original image shape: {original_image.shape}")
+    reconstructed_image: Optional[torch.Tensor] = None
+
+    with torch.no_grad():
+        latent_image = to_latent(rgb_nchw=original_image, vae=vae)
+        print(f"Latent shape: {latent_image.shape}")
+        reconstructed_image = from_latent(latent_nchw=latent_image, vae=vae)
+        reconstructed_pil = nhwc_to_pil(nhwc=reconstructed_image)
+    combined_image = concatenate_images(
+        left=original_pil,
+        right=reconstructed_pil,
+        vertical=False,
+    )
+    combined_image.show("Original | Reconstruction")
+    print(f"Reconstructed image shape: {reconstructed_image.shape}")
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Inference with VAE")
+    parser.add_argument(
+        "--input_image",
+        type=str,
+        required=True,
+        help="Path to the input image for inference.",
+    )
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        required=True,
+        help="Path to pretrained VAE model.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        help="Model version.",
+    )
+    parser.add_argument(
+        "--variant",
+        type=str,
+        default=None,
+        help="Model file variant, e.g., 'fp16'.",
+    )
+    parser.add_argument(
+        "--subfolder",
+        type=str,
+        default=None,
+        help="Subfolder in the model file.",
+    )
+    parser.add_argument(
+        "--use_cuda",
+        action="store_true",
+        help="Use CUDA if available.",
+    )
+    parser.add_argument(
+        "--use_tiny_nn",
+        action="store_true",
+        help="Use tiny neural network.",
+    )
+    return parser.parse_args()
+
+
+# EXAMPLE USAGE:
+#
+# python vae_roundtrip.py --use_cuda --pretrained_model_name_or_path "runwayml/stable-diffusion-v1-5" --subfolder "vae" --input_image "foo.png"
+#
+# python vae_roundtrip.py --use_cuda --pretrained_model_name_or_path "madebyollin/taesd" --use_tiny_nn --input_image "foo.png"
+#
+def main_cli() -> None:
+    args = parse_args()
+
+    input_image_path = args.input_image
+    assert isinstance(input_image_path, str)
+
+    pretrained_model_name_or_path = args.pretrained_model_name_or_path
+    assert isinstance(pretrained_model_name_or_path, str)
+
+    revision = args.revision
+    assert isinstance(revision, (str, type(None)))
+
+    variant = args.variant
+    assert isinstance(variant, (str, type(None)))
+
+    subfolder = args.subfolder
+    assert isinstance(subfolder, (str, type(None)))
+
+    use_cuda = args.use_cuda
+    assert isinstance(use_cuda, bool)
+
+    use_tiny_nn = args.use_tiny_nn
+    assert isinstance(use_tiny_nn, bool)
+
+    device = torch.device("cuda" if use_cuda else "cpu")
+
+    main_kwargs(
+        device=device,
+        input_image_path=input_image_path,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        revision=revision,
+        variant=variant,
+        subfolder=subfolder,
+        use_tiny_nn=use_tiny_nn,
+    )
+
+
+if __name__ == "__main__":
+    main_cli()

scripts/convert_aura_flow_to_diffusers.py

@@ -0,0 +1,131 @@
+import argparse
+
+import torch
+from huggingface_hub import hf_hub_download
+
+from diffusers.models.transformers.auraflow_transformer_2d import AuraFlowTransformer2DModel
+
+
+def load_original_state_dict(args):
+    model_pt = hf_hub_download(repo_id=args.original_state_dict_repo_id, filename="aura_diffusion_pytorch_model.bin")
+    state_dict = torch.load(model_pt, map_location="cpu")
+    return state_dict
+
+
+def calculate_layers(state_dict_keys, key_prefix):
+    dit_layers = set()
+    for k in state_dict_keys:
+        if key_prefix in k:
+            dit_layers.add(int(k.split(".")[2]))
+    print(f"{key_prefix}: {len(dit_layers)}")
+    return len(dit_layers)
+
+
+# similar to SD3 but only for the last norm layer
+def swap_scale_shift(weight, dim):
+    shift, scale = weight.chunk(2, dim=0)
+    new_weight = torch.cat([scale, shift], dim=0)
+    return new_weight
+
+
+def convert_transformer(state_dict):
+    converted_state_dict = {}
+    state_dict_keys = list(state_dict.keys())
+
+    converted_state_dict["register_tokens"] = state_dict.pop("model.register_tokens")
+    converted_state_dict["pos_embed.pos_embed"] = state_dict.pop("model.positional_encoding")
+    converted_state_dict["pos_embed.proj.weight"] = state_dict.pop("model.init_x_linear.weight")
+    converted_state_dict["pos_embed.proj.bias"] = state_dict.pop("model.init_x_linear.bias")
+
+    converted_state_dict["time_step_proj.linear_1.weight"] = state_dict.pop("model.t_embedder.mlp.0.weight")
+    converted_state_dict["time_step_proj.linear_1.bias"] = state_dict.pop("model.t_embedder.mlp.0.bias")
+    converted_state_dict["time_step_proj.linear_2.weight"] = state_dict.pop("model.t_embedder.mlp.2.weight")
+    converted_state_dict["time_step_proj.linear_2.bias"] = state_dict.pop("model.t_embedder.mlp.2.bias")
+
+    converted_state_dict["context_embedder.weight"] = state_dict.pop("model.cond_seq_linear.weight")
+
+    mmdit_layers = calculate_layers(state_dict_keys, key_prefix="double_layers")
+    single_dit_layers = calculate_layers(state_dict_keys, key_prefix="single_layers")
+
+    # MMDiT blocks 🎸.
+    for i in range(mmdit_layers):
+        # feed-forward
+        path_mapping = {"mlpX": "ff", "mlpC": "ff_context"}
+        weight_mapping = {"c_fc1": "linear_1", "c_fc2": "linear_2", "c_proj": "out_projection"}
+        for orig_k, diffuser_k in path_mapping.items():
+            for k, v in weight_mapping.items():
+                converted_state_dict[f"joint_transformer_blocks.{i}.{diffuser_k}.{v}.weight"] = state_dict.pop(
+                    f"model.double_layers.{i}.{orig_k}.{k}.weight"
+                )
+
+        # norms
+        path_mapping = {"modX": "norm1", "modC": "norm1_context"}
+        for orig_k, diffuser_k in path_mapping.items():
+            converted_state_dict[f"joint_transformer_blocks.{i}.{diffuser_k}.linear.weight"] = state_dict.pop(
+                f"model.double_layers.{i}.{orig_k}.1.weight"
+            )
+
+        # attns
+        x_attn_mapping = {"w2q": "to_q", "w2k": "to_k", "w2v": "to_v", "w2o": "to_out.0"}
+        context_attn_mapping = {"w1q": "add_q_proj", "w1k": "add_k_proj", "w1v": "add_v_proj", "w1o": "to_add_out"}
+        for attn_mapping in [x_attn_mapping, context_attn_mapping]:
+            for k, v in attn_mapping.items():
+                converted_state_dict[f"joint_transformer_blocks.{i}.attn.{v}.weight"] = state_dict.pop(
+                    f"model.double_layers.{i}.attn.{k}.weight"
+                )
+
+    # Single-DiT blocks.
+    for i in range(single_dit_layers):
+        # feed-forward
+        mapping = {"c_fc1": "linear_1", "c_fc2": "linear_2", "c_proj": "out_projection"}
+        for k, v in mapping.items():
+            converted_state_dict[f"single_transformer_blocks.{i}.ff.{v}.weight"] = state_dict.pop(
+                f"model.single_layers.{i}.mlp.{k}.weight"
+            )
+
+        # norms
+        converted_state_dict[f"single_transformer_blocks.{i}.norm1.linear.weight"] = state_dict.pop(
+            f"model.single_layers.{i}.modCX.1.weight"
+        )
+
+        # attns
+        x_attn_mapping = {"w1q": "to_q", "w1k": "to_k", "w1v": "to_v", "w1o": "to_out.0"}
+        for k, v in x_attn_mapping.items():
+            converted_state_dict[f"single_transformer_blocks.{i}.attn.{v}.weight"] = state_dict.pop(
+                f"model.single_layers.{i}.attn.{k}.weight"
+            )
+
+    # Final blocks.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("model.final_linear.weight")
+    converted_state_dict["norm_out.linear.weight"] = swap_scale_shift(state_dict.pop("model.modF.1.weight"), dim=None)
+
+    return converted_state_dict
+
+
+@torch.no_grad()
+def populate_state_dict(args):
+    original_state_dict = load_original_state_dict(args)
+    state_dict_keys = list(original_state_dict.keys())
+    mmdit_layers = calculate_layers(state_dict_keys, key_prefix="double_layers")
+    single_dit_layers = calculate_layers(state_dict_keys, key_prefix="single_layers")
+
+    converted_state_dict = convert_transformer(original_state_dict)
+    model_diffusers = AuraFlowTransformer2DModel(
+        num_mmdit_layers=mmdit_layers, num_single_dit_layers=single_dit_layers
+    )
+    model_diffusers.load_state_dict(converted_state_dict, strict=True)
+
+    return model_diffusers
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--original_state_dict_repo_id", default="AuraDiffusion/auradiffusion-v0.1a0", type=str)
+    parser.add_argument("--dump_path", default="aura-flow", type=str)
+    parser.add_argument("--hub_id", default=None, type=str)
+    args = parser.parse_args()
+
+    model_diffusers = populate_state_dict(args)
+    model_diffusers.save_pretrained(args.dump_path)
+    if args.hub_id is not None:
+        model_diffusers.push_to_hub(args.hub_id)

scripts/convert_hunyuandit_controlnet_to_diffusers.py

@@ -0,0 +1,241 @@
+import argparse
+
+import torch
+
+from diffusers import HunyuanDiT2DControlNetModel
+
+
+def main(args):
+    state_dict = torch.load(args.pt_checkpoint_path, map_location="cpu")
+
+    if args.load_key != "none":
+        try:
+            state_dict = state_dict[args.load_key]
+        except KeyError:
+            raise KeyError(
+                f"{args.load_key} not found in the checkpoint."
+                "Please load from the following keys:{state_dict.keys()}"
+            )
+    device = "cuda"
+
+    model_config = HunyuanDiT2DControlNetModel.load_config(
+        "Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers", subfolder="transformer"
+    )
+    model_config[
+        "use_style_cond_and_image_meta_size"
+    ] = args.use_style_cond_and_image_meta_size  ### version <= v1.1: True; version >= v1.2: False
+    print(model_config)
+
+    for key in state_dict:
+        print("local:", key)
+
+    model = HunyuanDiT2DControlNetModel.from_config(model_config).to(device)
+
+    for key in model.state_dict():
+        print("diffusers:", key)
+
+    num_layers = 19
+    for i in range(num_layers):
+        # attn1
+        # Wkqv -> to_q, to_k, to_v
+        q, k, v = torch.chunk(state_dict[f"blocks.{i}.attn1.Wqkv.weight"], 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict[f"blocks.{i}.attn1.Wqkv.bias"], 3, dim=0)
+        state_dict[f"blocks.{i}.attn1.to_q.weight"] = q
+        state_dict[f"blocks.{i}.attn1.to_q.bias"] = q_bias
+        state_dict[f"blocks.{i}.attn1.to_k.weight"] = k
+        state_dict[f"blocks.{i}.attn1.to_k.bias"] = k_bias
+        state_dict[f"blocks.{i}.attn1.to_v.weight"] = v
+        state_dict[f"blocks.{i}.attn1.to_v.bias"] = v_bias
+        state_dict.pop(f"blocks.{i}.attn1.Wqkv.weight")
+        state_dict.pop(f"blocks.{i}.attn1.Wqkv.bias")
+
+        # q_norm, k_norm -> norm_q, norm_k
+        state_dict[f"blocks.{i}.attn1.norm_q.weight"] = state_dict[f"blocks.{i}.attn1.q_norm.weight"]
+        state_dict[f"blocks.{i}.attn1.norm_q.bias"] = state_dict[f"blocks.{i}.attn1.q_norm.bias"]
+        state_dict[f"blocks.{i}.attn1.norm_k.weight"] = state_dict[f"blocks.{i}.attn1.k_norm.weight"]
+        state_dict[f"blocks.{i}.attn1.norm_k.bias"] = state_dict[f"blocks.{i}.attn1.k_norm.bias"]
+
+        state_dict.pop(f"blocks.{i}.attn1.q_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn1.q_norm.bias")
+        state_dict.pop(f"blocks.{i}.attn1.k_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn1.k_norm.bias")
+
+        # out_proj -> to_out
+        state_dict[f"blocks.{i}.attn1.to_out.0.weight"] = state_dict[f"blocks.{i}.attn1.out_proj.weight"]
+        state_dict[f"blocks.{i}.attn1.to_out.0.bias"] = state_dict[f"blocks.{i}.attn1.out_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn1.out_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn1.out_proj.bias")
+
+        # attn2
+        # kq_proj -> to_k, to_v
+        k, v = torch.chunk(state_dict[f"blocks.{i}.attn2.kv_proj.weight"], 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict[f"blocks.{i}.attn2.kv_proj.bias"], 2, dim=0)
+        state_dict[f"blocks.{i}.attn2.to_k.weight"] = k
+        state_dict[f"blocks.{i}.attn2.to_k.bias"] = k_bias
+        state_dict[f"blocks.{i}.attn2.to_v.weight"] = v
+        state_dict[f"blocks.{i}.attn2.to_v.bias"] = v_bias
+        state_dict.pop(f"blocks.{i}.attn2.kv_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.kv_proj.bias")
+
+        # q_proj -> to_q
+        state_dict[f"blocks.{i}.attn2.to_q.weight"] = state_dict[f"blocks.{i}.attn2.q_proj.weight"]
+        state_dict[f"blocks.{i}.attn2.to_q.bias"] = state_dict[f"blocks.{i}.attn2.q_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn2.q_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.q_proj.bias")
+
+        # q_norm, k_norm -> norm_q, norm_k
+        state_dict[f"blocks.{i}.attn2.norm_q.weight"] = state_dict[f"blocks.{i}.attn2.q_norm.weight"]
+        state_dict[f"blocks.{i}.attn2.norm_q.bias"] = state_dict[f"blocks.{i}.attn2.q_norm.bias"]
+        state_dict[f"blocks.{i}.attn2.norm_k.weight"] = state_dict[f"blocks.{i}.attn2.k_norm.weight"]
+        state_dict[f"blocks.{i}.attn2.norm_k.bias"] = state_dict[f"blocks.{i}.attn2.k_norm.bias"]
+
+        state_dict.pop(f"blocks.{i}.attn2.q_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn2.q_norm.bias")
+        state_dict.pop(f"blocks.{i}.attn2.k_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn2.k_norm.bias")
+
+        # out_proj -> to_out
+        state_dict[f"blocks.{i}.attn2.to_out.0.weight"] = state_dict[f"blocks.{i}.attn2.out_proj.weight"]
+        state_dict[f"blocks.{i}.attn2.to_out.0.bias"] = state_dict[f"blocks.{i}.attn2.out_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn2.out_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.out_proj.bias")
+
+        # switch norm 2 and norm 3
+        norm2_weight = state_dict[f"blocks.{i}.norm2.weight"]
+        norm2_bias = state_dict[f"blocks.{i}.norm2.bias"]
+        state_dict[f"blocks.{i}.norm2.weight"] = state_dict[f"blocks.{i}.norm3.weight"]
+        state_dict[f"blocks.{i}.norm2.bias"] = state_dict[f"blocks.{i}.norm3.bias"]
+        state_dict[f"blocks.{i}.norm3.weight"] = norm2_weight
+        state_dict[f"blocks.{i}.norm3.bias"] = norm2_bias
+
+        # norm1 -> norm1.norm
+        # default_modulation.1 -> norm1.linear
+        state_dict[f"blocks.{i}.norm1.norm.weight"] = state_dict[f"blocks.{i}.norm1.weight"]
+        state_dict[f"blocks.{i}.norm1.norm.bias"] = state_dict[f"blocks.{i}.norm1.bias"]
+        state_dict[f"blocks.{i}.norm1.linear.weight"] = state_dict[f"blocks.{i}.default_modulation.1.weight"]
+        state_dict[f"blocks.{i}.norm1.linear.bias"] = state_dict[f"blocks.{i}.default_modulation.1.bias"]
+        state_dict.pop(f"blocks.{i}.norm1.weight")
+        state_dict.pop(f"blocks.{i}.norm1.bias")
+        state_dict.pop(f"blocks.{i}.default_modulation.1.weight")
+        state_dict.pop(f"blocks.{i}.default_modulation.1.bias")
+
+        # mlp.fc1 -> ff.net.0, mlp.fc2 -> ff.net.2
+        state_dict[f"blocks.{i}.ff.net.0.proj.weight"] = state_dict[f"blocks.{i}.mlp.fc1.weight"]
+        state_dict[f"blocks.{i}.ff.net.0.proj.bias"] = state_dict[f"blocks.{i}.mlp.fc1.bias"]
+        state_dict[f"blocks.{i}.ff.net.2.weight"] = state_dict[f"blocks.{i}.mlp.fc2.weight"]
+        state_dict[f"blocks.{i}.ff.net.2.bias"] = state_dict[f"blocks.{i}.mlp.fc2.bias"]
+        state_dict.pop(f"blocks.{i}.mlp.fc1.weight")
+        state_dict.pop(f"blocks.{i}.mlp.fc1.bias")
+        state_dict.pop(f"blocks.{i}.mlp.fc2.weight")
+        state_dict.pop(f"blocks.{i}.mlp.fc2.bias")
+
+        # after_proj_list -> controlnet_blocks
+        state_dict[f"controlnet_blocks.{i}.weight"] = state_dict[f"after_proj_list.{i}.weight"]
+        state_dict[f"controlnet_blocks.{i}.bias"] = state_dict[f"after_proj_list.{i}.bias"]
+        state_dict.pop(f"after_proj_list.{i}.weight")
+        state_dict.pop(f"after_proj_list.{i}.bias")
+
+    # before_proj -> input_block
+    state_dict["input_block.weight"] = state_dict["before_proj.weight"]
+    state_dict["input_block.bias"] = state_dict["before_proj.bias"]
+    state_dict.pop("before_proj.weight")
+    state_dict.pop("before_proj.bias")
+
+    # pooler -> time_extra_emb
+    state_dict["time_extra_emb.pooler.positional_embedding"] = state_dict["pooler.positional_embedding"]
+    state_dict["time_extra_emb.pooler.k_proj.weight"] = state_dict["pooler.k_proj.weight"]
+    state_dict["time_extra_emb.pooler.k_proj.bias"] = state_dict["pooler.k_proj.bias"]
+    state_dict["time_extra_emb.pooler.q_proj.weight"] = state_dict["pooler.q_proj.weight"]
+    state_dict["time_extra_emb.pooler.q_proj.bias"] = state_dict["pooler.q_proj.bias"]
+    state_dict["time_extra_emb.pooler.v_proj.weight"] = state_dict["pooler.v_proj.weight"]
+    state_dict["time_extra_emb.pooler.v_proj.bias"] = state_dict["pooler.v_proj.bias"]
+    state_dict["time_extra_emb.pooler.c_proj.weight"] = state_dict["pooler.c_proj.weight"]
+    state_dict["time_extra_emb.pooler.c_proj.bias"] = state_dict["pooler.c_proj.bias"]
+    state_dict.pop("pooler.k_proj.weight")
+    state_dict.pop("pooler.k_proj.bias")
+    state_dict.pop("pooler.q_proj.weight")
+    state_dict.pop("pooler.q_proj.bias")
+    state_dict.pop("pooler.v_proj.weight")
+    state_dict.pop("pooler.v_proj.bias")
+    state_dict.pop("pooler.c_proj.weight")
+    state_dict.pop("pooler.c_proj.bias")
+    state_dict.pop("pooler.positional_embedding")
+
+    # t_embedder -> time_embedding (`TimestepEmbedding`)
+    state_dict["time_extra_emb.timestep_embedder.linear_1.bias"] = state_dict["t_embedder.mlp.0.bias"]
+    state_dict["time_extra_emb.timestep_embedder.linear_1.weight"] = state_dict["t_embedder.mlp.0.weight"]
+    state_dict["time_extra_emb.timestep_embedder.linear_2.bias"] = state_dict["t_embedder.mlp.2.bias"]
+    state_dict["time_extra_emb.timestep_embedder.linear_2.weight"] = state_dict["t_embedder.mlp.2.weight"]
+
+    state_dict.pop("t_embedder.mlp.0.bias")
+    state_dict.pop("t_embedder.mlp.0.weight")
+    state_dict.pop("t_embedder.mlp.2.bias")
+    state_dict.pop("t_embedder.mlp.2.weight")
+
+    # x_embedder -> pos_embd (`PatchEmbed`)
+    state_dict["pos_embed.proj.weight"] = state_dict["x_embedder.proj.weight"]
+    state_dict["pos_embed.proj.bias"] = state_dict["x_embedder.proj.bias"]
+    state_dict.pop("x_embedder.proj.weight")
+    state_dict.pop("x_embedder.proj.bias")
+
+    # mlp_t5 -> text_embedder
+    state_dict["text_embedder.linear_1.bias"] = state_dict["mlp_t5.0.bias"]
+    state_dict["text_embedder.linear_1.weight"] = state_dict["mlp_t5.0.weight"]
+    state_dict["text_embedder.linear_2.bias"] = state_dict["mlp_t5.2.bias"]
+    state_dict["text_embedder.linear_2.weight"] = state_dict["mlp_t5.2.weight"]
+    state_dict.pop("mlp_t5.0.bias")
+    state_dict.pop("mlp_t5.0.weight")
+    state_dict.pop("mlp_t5.2.bias")
+    state_dict.pop("mlp_t5.2.weight")
+
+    # extra_embedder -> extra_embedder
+    state_dict["time_extra_emb.extra_embedder.linear_1.bias"] = state_dict["extra_embedder.0.bias"]
+    state_dict["time_extra_emb.extra_embedder.linear_1.weight"] = state_dict["extra_embedder.0.weight"]
+    state_dict["time_extra_emb.extra_embedder.linear_2.bias"] = state_dict["extra_embedder.2.bias"]
+    state_dict["time_extra_emb.extra_embedder.linear_2.weight"] = state_dict["extra_embedder.2.weight"]
+    state_dict.pop("extra_embedder.0.bias")
+    state_dict.pop("extra_embedder.0.weight")
+    state_dict.pop("extra_embedder.2.bias")
+    state_dict.pop("extra_embedder.2.weight")
+
+    # style_embedder
+    if model_config["use_style_cond_and_image_meta_size"]:
+        print(state_dict["style_embedder.weight"])
+        print(state_dict["style_embedder.weight"].shape)
+        state_dict["time_extra_emb.style_embedder.weight"] = state_dict["style_embedder.weight"][0:1]
+        state_dict.pop("style_embedder.weight")
+
+    model.load_state_dict(state_dict)
+
+    if args.save:
+        model.save_pretrained(args.output_checkpoint_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--save", default=True, type=bool, required=False, help="Whether to save the converted pipeline or not."
+    )
+    parser.add_argument(
+        "--pt_checkpoint_path", default=None, type=str, required=True, help="Path to the .pt pretrained model."
+    )
+    parser.add_argument(
+        "--output_checkpoint_path",
+        default=None,
+        type=str,
+        required=False,
+        help="Path to the output converted diffusers pipeline.",
+    )
+    parser.add_argument(
+        "--load_key", default="none", type=str, required=False, help="The key to load from the pretrained .pt file"
+    )
+    parser.add_argument(
+        "--use_style_cond_and_image_meta_size",
+        type=bool,
+        default=False,
+        help="version <= v1.1: True; version >= v1.2: False",
+    )
+
+    args = parser.parse_args()
+    main(args)

scripts/convert_hunyuandit_to_diffusers.py

@@ -0,0 +1,267 @@
+import argparse
+
+import torch
+
+from diffusers import HunyuanDiT2DModel
+
+
+def main(args):
+    state_dict = torch.load(args.pt_checkpoint_path, map_location="cpu")
+
+    if args.load_key != "none":
+        try:
+            state_dict = state_dict[args.load_key]
+        except KeyError:
+            raise KeyError(
+                f"{args.load_key} not found in the checkpoint."
+                f"Please load from the following keys:{state_dict.keys()}"
+            )
+
+    device = "cuda"
+    model_config = HunyuanDiT2DModel.load_config("Tencent-Hunyuan/HunyuanDiT-Diffusers", subfolder="transformer")
+    model_config[
+        "use_style_cond_and_image_meta_size"
+    ] = args.use_style_cond_and_image_meta_size  ### version <= v1.1: True; version >= v1.2: False
+
+    # input_size -> sample_size, text_dim -> cross_attention_dim
+    for key in state_dict:
+        print("local:", key)
+
+    model = HunyuanDiT2DModel.from_config(model_config).to(device)
+
+    for key in model.state_dict():
+        print("diffusers:", key)
+
+    num_layers = 40
+    for i in range(num_layers):
+        # attn1
+        # Wkqv -> to_q, to_k, to_v
+        q, k, v = torch.chunk(state_dict[f"blocks.{i}.attn1.Wqkv.weight"], 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict[f"blocks.{i}.attn1.Wqkv.bias"], 3, dim=0)
+        state_dict[f"blocks.{i}.attn1.to_q.weight"] = q
+        state_dict[f"blocks.{i}.attn1.to_q.bias"] = q_bias
+        state_dict[f"blocks.{i}.attn1.to_k.weight"] = k
+        state_dict[f"blocks.{i}.attn1.to_k.bias"] = k_bias
+        state_dict[f"blocks.{i}.attn1.to_v.weight"] = v
+        state_dict[f"blocks.{i}.attn1.to_v.bias"] = v_bias
+        state_dict.pop(f"blocks.{i}.attn1.Wqkv.weight")
+        state_dict.pop(f"blocks.{i}.attn1.Wqkv.bias")
+
+        # q_norm, k_norm -> norm_q, norm_k
+        state_dict[f"blocks.{i}.attn1.norm_q.weight"] = state_dict[f"blocks.{i}.attn1.q_norm.weight"]
+        state_dict[f"blocks.{i}.attn1.norm_q.bias"] = state_dict[f"blocks.{i}.attn1.q_norm.bias"]
+        state_dict[f"blocks.{i}.attn1.norm_k.weight"] = state_dict[f"blocks.{i}.attn1.k_norm.weight"]
+        state_dict[f"blocks.{i}.attn1.norm_k.bias"] = state_dict[f"blocks.{i}.attn1.k_norm.bias"]
+
+        state_dict.pop(f"blocks.{i}.attn1.q_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn1.q_norm.bias")
+        state_dict.pop(f"blocks.{i}.attn1.k_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn1.k_norm.bias")
+
+        # out_proj -> to_out
+        state_dict[f"blocks.{i}.attn1.to_out.0.weight"] = state_dict[f"blocks.{i}.attn1.out_proj.weight"]
+        state_dict[f"blocks.{i}.attn1.to_out.0.bias"] = state_dict[f"blocks.{i}.attn1.out_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn1.out_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn1.out_proj.bias")
+
+        # attn2
+        # kq_proj -> to_k, to_v
+        k, v = torch.chunk(state_dict[f"blocks.{i}.attn2.kv_proj.weight"], 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict[f"blocks.{i}.attn2.kv_proj.bias"], 2, dim=0)
+        state_dict[f"blocks.{i}.attn2.to_k.weight"] = k
+        state_dict[f"blocks.{i}.attn2.to_k.bias"] = k_bias
+        state_dict[f"blocks.{i}.attn2.to_v.weight"] = v
+        state_dict[f"blocks.{i}.attn2.to_v.bias"] = v_bias
+        state_dict.pop(f"blocks.{i}.attn2.kv_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.kv_proj.bias")
+
+        # q_proj -> to_q
+        state_dict[f"blocks.{i}.attn2.to_q.weight"] = state_dict[f"blocks.{i}.attn2.q_proj.weight"]
+        state_dict[f"blocks.{i}.attn2.to_q.bias"] = state_dict[f"blocks.{i}.attn2.q_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn2.q_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.q_proj.bias")
+
+        # q_norm, k_norm -> norm_q, norm_k
+        state_dict[f"blocks.{i}.attn2.norm_q.weight"] = state_dict[f"blocks.{i}.attn2.q_norm.weight"]
+        state_dict[f"blocks.{i}.attn2.norm_q.bias"] = state_dict[f"blocks.{i}.attn2.q_norm.bias"]
+        state_dict[f"blocks.{i}.attn2.norm_k.weight"] = state_dict[f"blocks.{i}.attn2.k_norm.weight"]
+        state_dict[f"blocks.{i}.attn2.norm_k.bias"] = state_dict[f"blocks.{i}.attn2.k_norm.bias"]
+
+        state_dict.pop(f"blocks.{i}.attn2.q_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn2.q_norm.bias")
+        state_dict.pop(f"blocks.{i}.attn2.k_norm.weight")
+        state_dict.pop(f"blocks.{i}.attn2.k_norm.bias")
+
+        # out_proj -> to_out
+        state_dict[f"blocks.{i}.attn2.to_out.0.weight"] = state_dict[f"blocks.{i}.attn2.out_proj.weight"]
+        state_dict[f"blocks.{i}.attn2.to_out.0.bias"] = state_dict[f"blocks.{i}.attn2.out_proj.bias"]
+        state_dict.pop(f"blocks.{i}.attn2.out_proj.weight")
+        state_dict.pop(f"blocks.{i}.attn2.out_proj.bias")
+
+        # switch norm 2 and norm 3
+        norm2_weight = state_dict[f"blocks.{i}.norm2.weight"]
+        norm2_bias = state_dict[f"blocks.{i}.norm2.bias"]
+        state_dict[f"blocks.{i}.norm2.weight"] = state_dict[f"blocks.{i}.norm3.weight"]
+        state_dict[f"blocks.{i}.norm2.bias"] = state_dict[f"blocks.{i}.norm3.bias"]
+        state_dict[f"blocks.{i}.norm3.weight"] = norm2_weight
+        state_dict[f"blocks.{i}.norm3.bias"] = norm2_bias
+
+        # norm1 -> norm1.norm
+        # default_modulation.1 -> norm1.linear
+        state_dict[f"blocks.{i}.norm1.norm.weight"] = state_dict[f"blocks.{i}.norm1.weight"]
+        state_dict[f"blocks.{i}.norm1.norm.bias"] = state_dict[f"blocks.{i}.norm1.bias"]
+        state_dict[f"blocks.{i}.norm1.linear.weight"] = state_dict[f"blocks.{i}.default_modulation.1.weight"]
+        state_dict[f"blocks.{i}.norm1.linear.bias"] = state_dict[f"blocks.{i}.default_modulation.1.bias"]
+        state_dict.pop(f"blocks.{i}.norm1.weight")
+        state_dict.pop(f"blocks.{i}.norm1.bias")
+        state_dict.pop(f"blocks.{i}.default_modulation.1.weight")
+        state_dict.pop(f"blocks.{i}.default_modulation.1.bias")
+
+        # mlp.fc1 -> ff.net.0, mlp.fc2 -> ff.net.2
+        state_dict[f"blocks.{i}.ff.net.0.proj.weight"] = state_dict[f"blocks.{i}.mlp.fc1.weight"]
+        state_dict[f"blocks.{i}.ff.net.0.proj.bias"] = state_dict[f"blocks.{i}.mlp.fc1.bias"]
+        state_dict[f"blocks.{i}.ff.net.2.weight"] = state_dict[f"blocks.{i}.mlp.fc2.weight"]
+        state_dict[f"blocks.{i}.ff.net.2.bias"] = state_dict[f"blocks.{i}.mlp.fc2.bias"]
+        state_dict.pop(f"blocks.{i}.mlp.fc1.weight")
+        state_dict.pop(f"blocks.{i}.mlp.fc1.bias")
+        state_dict.pop(f"blocks.{i}.mlp.fc2.weight")
+        state_dict.pop(f"blocks.{i}.mlp.fc2.bias")
+
+    # pooler -> time_extra_emb
+    state_dict["time_extra_emb.pooler.positional_embedding"] = state_dict["pooler.positional_embedding"]
+    state_dict["time_extra_emb.pooler.k_proj.weight"] = state_dict["pooler.k_proj.weight"]
+    state_dict["time_extra_emb.pooler.k_proj.bias"] = state_dict["pooler.k_proj.bias"]
+    state_dict["time_extra_emb.pooler.q_proj.weight"] = state_dict["pooler.q_proj.weight"]
+    state_dict["time_extra_emb.pooler.q_proj.bias"] = state_dict["pooler.q_proj.bias"]
+    state_dict["time_extra_emb.pooler.v_proj.weight"] = state_dict["pooler.v_proj.weight"]
+    state_dict["time_extra_emb.pooler.v_proj.bias"] = state_dict["pooler.v_proj.bias"]
+    state_dict["time_extra_emb.pooler.c_proj.weight"] = state_dict["pooler.c_proj.weight"]
+    state_dict["time_extra_emb.pooler.c_proj.bias"] = state_dict["pooler.c_proj.bias"]
+    state_dict.pop("pooler.k_proj.weight")
+    state_dict.pop("pooler.k_proj.bias")
+    state_dict.pop("pooler.q_proj.weight")
+    state_dict.pop("pooler.q_proj.bias")
+    state_dict.pop("pooler.v_proj.weight")
+    state_dict.pop("pooler.v_proj.bias")
+    state_dict.pop("pooler.c_proj.weight")
+    state_dict.pop("pooler.c_proj.bias")
+    state_dict.pop("pooler.positional_embedding")
+
+    # t_embedder -> time_embedding (`TimestepEmbedding`)
+    state_dict["time_extra_emb.timestep_embedder.linear_1.bias"] = state_dict["t_embedder.mlp.0.bias"]
+    state_dict["time_extra_emb.timestep_embedder.linear_1.weight"] = state_dict["t_embedder.mlp.0.weight"]
+    state_dict["time_extra_emb.timestep_embedder.linear_2.bias"] = state_dict["t_embedder.mlp.2.bias"]
+    state_dict["time_extra_emb.timestep_embedder.linear_2.weight"] = state_dict["t_embedder.mlp.2.weight"]
+
+    state_dict.pop("t_embedder.mlp.0.bias")
+    state_dict.pop("t_embedder.mlp.0.weight")
+    state_dict.pop("t_embedder.mlp.2.bias")
+    state_dict.pop("t_embedder.mlp.2.weight")
+
+    # x_embedder -> pos_embd (`PatchEmbed`)
+    state_dict["pos_embed.proj.weight"] = state_dict["x_embedder.proj.weight"]
+    state_dict["pos_embed.proj.bias"] = state_dict["x_embedder.proj.bias"]
+    state_dict.pop("x_embedder.proj.weight")
+    state_dict.pop("x_embedder.proj.bias")
+
+    # mlp_t5 -> text_embedder
+    state_dict["text_embedder.linear_1.bias"] = state_dict["mlp_t5.0.bias"]
+    state_dict["text_embedder.linear_1.weight"] = state_dict["mlp_t5.0.weight"]
+    state_dict["text_embedder.linear_2.bias"] = state_dict["mlp_t5.2.bias"]
+    state_dict["text_embedder.linear_2.weight"] = state_dict["mlp_t5.2.weight"]
+    state_dict.pop("mlp_t5.0.bias")
+    state_dict.pop("mlp_t5.0.weight")
+    state_dict.pop("mlp_t5.2.bias")
+    state_dict.pop("mlp_t5.2.weight")
+
+    # extra_embedder -> extra_embedder
+    state_dict["time_extra_emb.extra_embedder.linear_1.bias"] = state_dict["extra_embedder.0.bias"]
+    state_dict["time_extra_emb.extra_embedder.linear_1.weight"] = state_dict["extra_embedder.0.weight"]
+    state_dict["time_extra_emb.extra_embedder.linear_2.bias"] = state_dict["extra_embedder.2.bias"]
+    state_dict["time_extra_emb.extra_embedder.linear_2.weight"] = state_dict["extra_embedder.2.weight"]
+    state_dict.pop("extra_embedder.0.bias")
+    state_dict.pop("extra_embedder.0.weight")
+    state_dict.pop("extra_embedder.2.bias")
+    state_dict.pop("extra_embedder.2.weight")
+
+    # model.final_adaLN_modulation.1 -> norm_out.linear
+    def swap_scale_shift(weight):
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        return new_weight
+
+    state_dict["norm_out.linear.weight"] = swap_scale_shift(state_dict["final_layer.adaLN_modulation.1.weight"])
+    state_dict["norm_out.linear.bias"] = swap_scale_shift(state_dict["final_layer.adaLN_modulation.1.bias"])
+    state_dict.pop("final_layer.adaLN_modulation.1.weight")
+    state_dict.pop("final_layer.adaLN_modulation.1.bias")
+
+    # final_linear -> proj_out
+    state_dict["proj_out.weight"] = state_dict["final_layer.linear.weight"]
+    state_dict["proj_out.bias"] = state_dict["final_layer.linear.bias"]
+    state_dict.pop("final_layer.linear.weight")
+    state_dict.pop("final_layer.linear.bias")
+
+    # style_embedder
+    if model_config["use_style_cond_and_image_meta_size"]:
+        print(state_dict["style_embedder.weight"])
+        print(state_dict["style_embedder.weight"].shape)
+        state_dict["time_extra_emb.style_embedder.weight"] = state_dict["style_embedder.weight"][0:1]
+        state_dict.pop("style_embedder.weight")
+
+    model.load_state_dict(state_dict)
+
+    from diffusers import HunyuanDiTPipeline
+
+    if args.use_style_cond_and_image_meta_size:
+        pipe = HunyuanDiTPipeline.from_pretrained(
+            "Tencent-Hunyuan/HunyuanDiT-Diffusers", transformer=model, torch_dtype=torch.float32
+        )
+    else:
+        pipe = HunyuanDiTPipeline.from_pretrained(
+            "Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers", transformer=model, torch_dtype=torch.float32
+        )
+    pipe.to("cuda")
+    pipe.to(dtype=torch.float32)
+
+    if args.save:
+        pipe.save_pretrained(args.output_checkpoint_path)
+
+    # ### NOTE: HunyuanDiT supports both Chinese and English inputs
+    prompt = "一个宇航员在骑马"
+    # prompt = "An astronaut riding a horse"
+    generator = torch.Generator(device="cuda").manual_seed(0)
+    image = pipe(
+        height=1024, width=1024, prompt=prompt, generator=generator, num_inference_steps=25, guidance_scale=5.0
+    ).images[0]
+
+    image.save("img.png")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--save", default=True, type=bool, required=False, help="Whether to save the converted pipeline or not."
+    )
+    parser.add_argument(
+        "--pt_checkpoint_path", default=None, type=str, required=True, help="Path to the .pt pretrained model."
+    )
+    parser.add_argument(
+        "--output_checkpoint_path",
+        default=None,
+        type=str,
+        required=False,
+        help="Path to the output converted diffusers pipeline.",
+    )
+    parser.add_argument(
+        "--load_key", default="none", type=str, required=False, help="The key to load from the pretrained .pt file"
+    )
+    parser.add_argument(
+        "--use_style_cond_and_image_meta_size",
+        type=bool,
+        default=False,
+        help="version <= v1.1: True; version >= v1.2: False",
+    )
+
+    args = parser.parse_args()
+    main(args)

scripts/convert_lumina_to_diffusers.py

@@ -0,0 +1,142 @@
+import argparse
+import os
+
+import torch
+from safetensors.torch import load_file
+from transformers import AutoModel, AutoTokenizer
+
+from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, LuminaNextDiT2DModel, LuminaText2ImgPipeline
+
+
+def main(args):
+    # checkpoint from https://huggingface.co/Alpha-VLLM/Lumina-Next-SFT or https://huggingface.co/Alpha-VLLM/Lumina-Next-T2I
+    all_sd = load_file(args.origin_ckpt_path, device="cpu")
+    converted_state_dict = {}
+    # pad token
+    converted_state_dict["pad_token"] = all_sd["pad_token"]
+
+    # patch embed
+    converted_state_dict["patch_embedder.weight"] = all_sd["x_embedder.weight"]
+    converted_state_dict["patch_embedder.bias"] = all_sd["x_embedder.bias"]
+
+    # time and caption embed
+    converted_state_dict["time_caption_embed.timestep_embedder.linear_1.weight"] = all_sd["t_embedder.mlp.0.weight"]
+    converted_state_dict["time_caption_embed.timestep_embedder.linear_1.bias"] = all_sd["t_embedder.mlp.0.bias"]
+    converted_state_dict["time_caption_embed.timestep_embedder.linear_2.weight"] = all_sd["t_embedder.mlp.2.weight"]
+    converted_state_dict["time_caption_embed.timestep_embedder.linear_2.bias"] = all_sd["t_embedder.mlp.2.bias"]
+    converted_state_dict["time_caption_embed.caption_embedder.0.weight"] = all_sd["cap_embedder.0.weight"]
+    converted_state_dict["time_caption_embed.caption_embedder.0.bias"] = all_sd["cap_embedder.0.bias"]
+    converted_state_dict["time_caption_embed.caption_embedder.1.weight"] = all_sd["cap_embedder.1.weight"]
+    converted_state_dict["time_caption_embed.caption_embedder.1.bias"] = all_sd["cap_embedder.1.bias"]
+
+    for i in range(24):
+        # adaln
+        converted_state_dict[f"layers.{i}.gate"] = all_sd[f"layers.{i}.attention.gate"]
+        converted_state_dict[f"layers.{i}.adaLN_modulation.1.weight"] = all_sd[f"layers.{i}.adaLN_modulation.1.weight"]
+        converted_state_dict[f"layers.{i}.adaLN_modulation.1.bias"] = all_sd[f"layers.{i}.adaLN_modulation.1.bias"]
+
+        # qkv
+        converted_state_dict[f"layers.{i}.attn1.to_q.weight"] = all_sd[f"layers.{i}.attention.wq.weight"]
+        converted_state_dict[f"layers.{i}.attn1.to_k.weight"] = all_sd[f"layers.{i}.attention.wk.weight"]
+        converted_state_dict[f"layers.{i}.attn1.to_v.weight"] = all_sd[f"layers.{i}.attention.wv.weight"]
+
+        # cap
+        converted_state_dict[f"layers.{i}.attn2.to_q.weight"] = all_sd[f"layers.{i}.attention.wq.weight"]
+        converted_state_dict[f"layers.{i}.attn2.to_k.weight"] = all_sd[f"layers.{i}.attention.wk_y.weight"]
+        converted_state_dict[f"layers.{i}.attn2.to_v.weight"] = all_sd[f"layers.{i}.attention.wv_y.weight"]
+
+        # output
+        converted_state_dict[f"layers.{i}.attn2.to_out.0.weight"] = all_sd[f"layers.{i}.attention.wo.weight"]
+
+        # attention
+        # qk norm
+        converted_state_dict[f"layers.{i}.attn1.norm_q.weight"] = all_sd[f"layers.{i}.attention.q_norm.weight"]
+        converted_state_dict[f"layers.{i}.attn1.norm_q.bias"] = all_sd[f"layers.{i}.attention.q_norm.bias"]
+
+        converted_state_dict[f"layers.{i}.attn1.norm_k.weight"] = all_sd[f"layers.{i}.attention.k_norm.weight"]
+        converted_state_dict[f"layers.{i}.attn1.norm_k.bias"] = all_sd[f"layers.{i}.attention.k_norm.bias"]
+
+        converted_state_dict[f"layers.{i}.attn2.norm_q.weight"] = all_sd[f"layers.{i}.attention.q_norm.weight"]
+        converted_state_dict[f"layers.{i}.attn2.norm_q.bias"] = all_sd[f"layers.{i}.attention.q_norm.bias"]
+
+        converted_state_dict[f"layers.{i}.attn2.norm_k.weight"] = all_sd[f"layers.{i}.attention.ky_norm.weight"]
+        converted_state_dict[f"layers.{i}.attn2.norm_k.bias"] = all_sd[f"layers.{i}.attention.ky_norm.bias"]
+
+        # attention norm
+        converted_state_dict[f"layers.{i}.attn_norm1.weight"] = all_sd[f"layers.{i}.attention_norm1.weight"]
+        converted_state_dict[f"layers.{i}.attn_norm2.weight"] = all_sd[f"layers.{i}.attention_norm2.weight"]
+        converted_state_dict[f"layers.{i}.norm1_context.weight"] = all_sd[f"layers.{i}.attention_y_norm.weight"]
+
+        # feed forward
+        converted_state_dict[f"layers.{i}.feed_forward.linear_1.weight"] = all_sd[f"layers.{i}.feed_forward.w1.weight"]
+        converted_state_dict[f"layers.{i}.feed_forward.linear_2.weight"] = all_sd[f"layers.{i}.feed_forward.w2.weight"]
+        converted_state_dict[f"layers.{i}.feed_forward.linear_3.weight"] = all_sd[f"layers.{i}.feed_forward.w3.weight"]
+
+        # feed forward norm
+        converted_state_dict[f"layers.{i}.ffn_norm1.weight"] = all_sd[f"layers.{i}.ffn_norm1.weight"]
+        converted_state_dict[f"layers.{i}.ffn_norm2.weight"] = all_sd[f"layers.{i}.ffn_norm2.weight"]
+
+    # final layer
+    converted_state_dict["final_layer.linear.weight"] = all_sd["final_layer.linear.weight"]
+    converted_state_dict["final_layer.linear.bias"] = all_sd["final_layer.linear.bias"]
+
+    converted_state_dict["final_layer.adaLN_modulation.1.weight"] = all_sd["final_layer.adaLN_modulation.1.weight"]
+    converted_state_dict["final_layer.adaLN_modulation.1.bias"] = all_sd["final_layer.adaLN_modulation.1.bias"]
+
+    # Lumina-Next-SFT 2B
+    transformer = LuminaNextDiT2DModel(
+        sample_size=128,
+        patch_size=2,
+        in_channels=4,
+        hidden_size=2304,
+        num_layers=24,
+        num_attention_heads=32,
+        num_kv_heads=8,
+        multiple_of=256,
+        ffn_dim_multiplier=None,
+        norm_eps=1e-5,
+        learn_sigma=True,
+        qk_norm=True,
+        cross_attention_dim=2048,
+        scaling_factor=1.0,
+    )
+    transformer.load_state_dict(converted_state_dict, strict=True)
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    if args.only_transformer:
+        transformer.save_pretrained(os.path.join(args.dump_path, "transformer"))
+    else:
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", torch_dtype=torch.float32)
+
+        tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
+        text_encoder = AutoModel.from_pretrained("google/gemma-2b")
+
+        pipeline = LuminaText2ImgPipeline(
+            tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, scheduler=scheduler
+        )
+        pipeline.save_pretrained(args.dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--origin_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[256, 512, 1024],
+        required=False,
+        help="Image size of pretrained model, either 512 or 1024.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--only_transformer", default=True, type=bool, required=True)
+
+    args = parser.parse_args()
+    main(args)

scripts/generate_logits.py

@@ -103,12 +103,12 @@ results["google_ddpm_ema_cat_256"] = torch.tensor([
 
 models = api.list_models(filter="diffusers")
 for mod in models:
-    if "google" in mod.author or mod.modelId == "CompVis/ldm-celebahq-256":
-        local_checkpoint = "/home/patrick/google_checkpoints/" + mod.modelId.split("/")[-1]
+    if "google" in mod.author or mod.id == "CompVis/ldm-celebahq-256":
+        local_checkpoint = "/home/patrick/google_checkpoints/" + mod.id.split("/")[-1]
 
-        print(f"Started running {mod.modelId}!!!")
+        print(f"Started running {mod.id}!!!")
 
-        if mod.modelId.startswith("CompVis"):
+        if mod.id.startswith("CompVis"):
             model = UNet2DModel.from_pretrained(local_checkpoint, subfolder="unet")
         else:
             model = UNet2DModel.from_pretrained(local_checkpoint)
@@ -122,6 +122,6 @@ for mod in models:
             logits = model(noise, time_step).sample
 
         assert torch.allclose(
-            logits[0, 0, 0, :30], results["_".join("_".join(mod.modelId.split("/")).split("-"))], atol=1e-3
+            logits[0, 0, 0, :30], results["_".join("_".join(mod.id.split("/")).split("-"))], atol=1e-3
         )
-        print(f"{mod.modelId} has passed successfully!!!")
+        print(f"{mod.id} has passed successfully!!!")

src/diffusers/__init__.py

@@ -76,6 +76,7 @@ else:
     _import_structure["models"].extend(
         [
             "AsymmetricAutoencoderKL",
+            "AuraFlowTransformer2DModel",
             "AutoencoderKL",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderTiny",
@@ -83,9 +84,13 @@ else:
             "ControlNetModel",
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
+            "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
+            "HunyuanDiT2DMultiControlNetModel",
             "I2VGenXLUNet",
             "Kandinsky3UNet",
+            "LatteTransformer3DModel",
+            "LuminaNextDiT2DModel",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
@@ -160,6 +165,7 @@ else:
             "EulerAncestralDiscreteScheduler",
             "EulerDiscreteScheduler",
             "FlowMatchEulerDiscreteScheduler",
+            "FlowMatchHeunDiscreteScheduler",
             "HeunDiscreteScheduler",
             "IPNDMScheduler",
             "KarrasVeScheduler",
@@ -230,10 +236,14 @@ else:
             "AudioLDM2ProjectionModel",
             "AudioLDM2UNet2DConditionModel",
             "AudioLDMPipeline",
+            "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
+            "ChatGLMModel",
+            "ChatGLMTokenizer",
             "CLIPImageProjection",
             "CycleDiffusionPipeline",
+            "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPipeline",
             "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
@@ -262,11 +272,15 @@ else:
             "KandinskyV22Pipeline",
             "KandinskyV22PriorEmb2EmbPipeline",
             "KandinskyV22PriorPipeline",
+            "KolorsImg2ImgPipeline",
+            "KolorsPipeline",
             "LatentConsistencyModelImg2ImgPipeline",
             "LatentConsistencyModelPipeline",
+            "LattePipeline",
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
+            "LuminaText2ImgPipeline",
             "MarigoldDepthPipeline",
             "MarigoldNormalsPipeline",
             "MusicLDMPipeline",
@@ -282,11 +296,13 @@ else:
             "StableCascadePriorPipeline",
             "StableDiffusion3ControlNetPipeline",
             "StableDiffusion3Img2ImgPipeline",
+            "StableDiffusion3InpaintPipeline",
             "StableDiffusion3Pipeline",
             "StableDiffusionAdapterPipeline",
             "StableDiffusionAttendAndExcitePipeline",
             "StableDiffusionControlNetImg2ImgPipeline",
             "StableDiffusionControlNetInpaintPipeline",
+            "StableDiffusionControlNetPAGPipeline",
             "StableDiffusionControlNetPipeline",
             "StableDiffusionControlNetXSPipeline",
             "StableDiffusionDepth2ImgPipeline",
@@ -301,6 +317,7 @@ else:
             "StableDiffusionLatentUpscalePipeline",
             "StableDiffusionLDM3DPipeline",
             "StableDiffusionModelEditingPipeline",
+            "StableDiffusionPAGPipeline",
             "StableDiffusionPanoramaPipeline",
             "StableDiffusionParadigmsPipeline",
             "StableDiffusionPipeline",
@@ -493,6 +510,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .models import (
             AsymmetricAutoencoderKL,
+            AuraFlowTransformer2DModel,
             AutoencoderKL,
             AutoencoderKLTemporalDecoder,
             AutoencoderTiny,
@@ -500,9 +518,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ControlNetModel,
             ControlNetXSAdapter,
             DiTTransformer2DModel,
+            HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
+            HunyuanDiT2DMultiControlNetModel,
             I2VGenXLUNet,
             Kandinsky3UNet,
+            LatteTransformer3DModel,
+            LuminaNextDiT2DModel,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,
@@ -574,6 +596,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             EulerAncestralDiscreteScheduler,
             EulerDiscreteScheduler,
             FlowMatchEulerDiscreteScheduler,
+            FlowMatchHeunDiscreteScheduler,
             HeunDiscreteScheduler,
             IPNDMScheduler,
             KarrasVeScheduler,
@@ -627,8 +650,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDM2ProjectionModel,
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
+            AuraFlowPipeline,
+            ChatGLMModel,
+            ChatGLMTokenizer,
             CLIPImageProjection,
             CycleDiffusionPipeline,
+            HunyuanDiTControlNetPipeline,
             HunyuanDiTPipeline,
             I2VGenXLPipeline,
             IFImg2ImgPipeline,
@@ -657,11 +684,15 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             KandinskyV22Pipeline,
             KandinskyV22PriorEmb2EmbPipeline,
             KandinskyV22PriorPipeline,
+            KolorsImg2ImgPipeline,
+            KolorsPipeline,
             LatentConsistencyModelImg2ImgPipeline,
             LatentConsistencyModelPipeline,
+            LattePipeline,
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            LuminaText2ImgPipeline,
             MarigoldDepthPipeline,
             MarigoldNormalsPipeline,
             MusicLDMPipeline,
@@ -677,11 +708,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableCascadePriorPipeline,
             StableDiffusion3ControlNetPipeline,
             StableDiffusion3Img2ImgPipeline,
+            StableDiffusion3InpaintPipeline,
             StableDiffusion3Pipeline,
             StableDiffusionAdapterPipeline,
             StableDiffusionAttendAndExcitePipeline,
             StableDiffusionControlNetImg2ImgPipeline,
             StableDiffusionControlNetInpaintPipeline,
+            StableDiffusionControlNetPAGPipeline,
             StableDiffusionControlNetPipeline,
             StableDiffusionControlNetXSPipeline,
             StableDiffusionDepth2ImgPipeline,
@@ -696,6 +729,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionLatentUpscalePipeline,
             StableDiffusionLDM3DPipeline,
             StableDiffusionModelEditingPipeline,
+            StableDiffusionPAGPipeline,
             StableDiffusionPanoramaPipeline,
             StableDiffusionParadigmsPipeline,
             StableDiffusionPipeline,

src/diffusers/commands/env.py

@@ -24,7 +24,6 @@ from ..utils import (
     is_bitsandbytes_available,
     is_flax_available,
     is_google_colab,
-    is_notebook,
     is_peft_available,
     is_safetensors_available,
     is_torch_available,
@@ -107,8 +106,6 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
 
         platform_info = platform.platform()
 
-        is_notebook_str = "Yes" if is_notebook() else "No"
-
         is_google_colab_str = "Yes" if is_google_colab() else "No"
 
         accelerator = "NA"
@@ -123,7 +120,7 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
                 out_str = out_str.decode("utf-8")
 
                 if len(out_str) > 0:
-                    accelerator = out_str.strip() + " VRAM"
+                    accelerator = out_str.strip()
             except FileNotFoundError:
                 pass
         elif platform.system() == "Darwin":  # Mac OS
@@ -155,7 +152,6 @@ class EnvironmentCommand(BaseDiffusersCLICommand):
         info = {
             "🤗 Diffusers version": version,
             "Platform": platform_info,
-            "Running on a notebook?": is_notebook_str,
             "Running on Google Colab?": is_google_colab_str,
             "Python version": platform.python_version(),
             "PyTorch version (GPU?)": f"{pt_version} ({pt_cuda_available})",

src/diffusers/configuration_utils.py

@@ -23,7 +23,7 @@ import json
 import os
 import re
 from collections import OrderedDict
-from pathlib import PosixPath
+from pathlib import Path
 from typing import Any, Dict, Tuple, Union
 
 import numpy as np
@@ -310,9 +310,6 @@ class ConfigMixin:
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                 cached versions if they exist.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -343,7 +340,6 @@ class ConfigMixin:
         local_dir = kwargs.pop("local_dir", None)
         local_dir_use_symlinks = kwargs.pop("local_dir_use_symlinks", "auto")
         force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_download", None)
         proxies = kwargs.pop("proxies", None)
         token = kwargs.pop("token", None)
         local_files_only = kwargs.pop("local_files_only", False)
@@ -386,7 +382,6 @@ class ConfigMixin:
                     cache_dir=cache_dir,
                     force_download=force_download,
                     proxies=proxies,
-                    resume_download=resume_download,
                     local_files_only=local_files_only,
                     token=token,
                     user_agent=user_agent,
@@ -587,8 +582,8 @@ class ConfigMixin:
         def to_json_saveable(value):
             if isinstance(value, np.ndarray):
                 value = value.tolist()
-            elif isinstance(value, PosixPath):
-                value = str(value)
+            elif isinstance(value, Path):
+                value = value.as_posix()
             return value
 
         config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}

src/diffusers/loaders/autoencoder.py

@@ -1,146 +0,0 @@
-# Copyright 2024 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.
-
-from huggingface_hub.utils import validate_hf_hub_args
-
-from .single_file_utils import (
-    create_diffusers_vae_model_from_ldm,
-    fetch_ldm_config_and_checkpoint,
-)
-
-
-class FromOriginalVAEMixin:
-    """
-    Load pretrained AutoencoderKL weights saved in the `.ckpt` or `.safetensors` format into a [`AutoencoderKL`].
-    """
-
-    @classmethod
-    @validate_hf_hub_args
-    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
-        r"""
-        Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
-        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
-
-        Parameters:
-            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
-                Can be either:
-                    - A link to the `.ckpt` file (for example
-                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
-                    - A path to a *file* containing all pipeline weights.
-            config_file (`str`, *optional*):
-                Filepath to the configuration YAML file associated with the model. If not provided it will default to:
-                https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml
-            torch_dtype (`str` or `torch.dtype`, *optional*):
-                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
-                dtype is automatically derived from the model's weights.
-            force_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
-                cached versions if they exist.
-            cache_dir (`Union[str, os.PathLike]`, *optional*):
-                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
-                is not used.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
-            proxies (`Dict[str, str]`, *optional*):
-                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
-                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
-            local_files_only (`bool`, *optional*, defaults to `False`):
-                Whether to only load local model weights and configuration files or not. If set to True, the model
-                won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
-                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
-                `diffusers-cli login` (stored in `~/.huggingface`) is used.
-            revision (`str`, *optional*, defaults to `"main"`):
-                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
-                allowed by Git.
-            image_size (`int`, *optional*, defaults to 512):
-                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
-                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
-            scaling_factor (`float`, *optional*, defaults to 0.18215):
-                The component-wise standard deviation of the trained latent space computed using the first batch of the
-                training set. This is used to scale the latent space to have unit variance when training the diffusion
-                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
-                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
-                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
-                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
-            kwargs (remaining dictionary of keyword arguments, *optional*):
-                Can be used to overwrite load and saveable variables (for example the pipeline components of the
-                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
-                method. See example below for more information.
-
-        <Tip warning={true}>
-
-            Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you're loading
-            a VAE from SDXL or a Stable Diffusion v2 model or higher.
-
-        </Tip>
-
-        Examples:
-
-        ```py
-        from diffusers import AutoencoderKL
-
-        url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"  # can also be local file
-        model = AutoencoderKL.from_single_file(url)
-        ```
-        """
-
-        original_config_file = kwargs.pop("original_config_file", None)
-        config_file = kwargs.pop("config_file", None)
-        resume_download = kwargs.pop("resume_download", None)
-        force_download = kwargs.pop("force_download", False)
-        proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
-        cache_dir = kwargs.pop("cache_dir", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        revision = kwargs.pop("revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", None)
-
-        class_name = cls.__name__
-
-        if (config_file is not None) and (original_config_file is not None):
-            raise ValueError(
-                "You cannot pass both `config_file` and `original_config_file` to `from_single_file`. Please use only one of these arguments."
-            )
-
-        original_config_file = original_config_file or config_file
-        original_config, checkpoint = fetch_ldm_config_and_checkpoint(
-            pretrained_model_link_or_path=pretrained_model_link_or_path,
-            class_name=class_name,
-            original_config_file=original_config_file,
-            resume_download=resume_download,
-            force_download=force_download,
-            proxies=proxies,
-            token=token,
-            revision=revision,
-            local_files_only=local_files_only,
-            cache_dir=cache_dir,
-        )
-
-        image_size = kwargs.pop("image_size", None)
-        scaling_factor = kwargs.pop("scaling_factor", None)
-        component = create_diffusers_vae_model_from_ldm(
-            class_name,
-            original_config,
-            checkpoint,
-            image_size=image_size,
-            scaling_factor=scaling_factor,
-            torch_dtype=torch_dtype,
-        )
-        vae = component["vae"]
-        if torch_dtype is not None:
-            vae = vae.to(torch_dtype)
-
-        return vae

src/diffusers/loaders/controlnet.py

@@ -1,136 +0,0 @@
-# Copyright 2024 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.
-
-from huggingface_hub.utils import validate_hf_hub_args
-
-from .single_file_utils import (
-    create_diffusers_controlnet_model_from_ldm,
-    fetch_ldm_config_and_checkpoint,
-)
-
-
-class FromOriginalControlNetMixin:
-    """
-    Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into a [`ControlNetModel`].
-    """
-
-    @classmethod
-    @validate_hf_hub_args
-    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
-        r"""
-        Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
-        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
-
-        Parameters:
-            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
-                Can be either:
-                    - A link to the `.ckpt` file (for example
-                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
-                    - A path to a *file* containing all pipeline weights.
-            config_file (`str`, *optional*):
-                Filepath to the configuration YAML file associated with the model. If not provided it will default to:
-                https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml
-            torch_dtype (`str` or `torch.dtype`, *optional*):
-                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
-                dtype is automatically derived from the model's weights.
-            force_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
-                cached versions if they exist.
-            cache_dir (`Union[str, os.PathLike]`, *optional*):
-                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
-                is not used.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
-            proxies (`Dict[str, str]`, *optional*):
-                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
-                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
-            local_files_only (`bool`, *optional*, defaults to `False`):
-                Whether to only load local model weights and configuration files or not. If set to True, the model
-                won't be downloaded from the Hub.
-            token (`str` or *bool*, *optional*):
-                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
-                `diffusers-cli login` (stored in `~/.huggingface`) is used.
-            revision (`str`, *optional*, defaults to `"main"`):
-                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
-                allowed by Git.
-            image_size (`int`, *optional*, defaults to 512):
-                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
-                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
-            upcast_attention (`bool`, *optional*, defaults to `None`):
-                Whether the attention computation should always be upcasted.
-            kwargs (remaining dictionary of keyword arguments, *optional*):
-                Can be used to overwrite load and saveable variables (for example the pipeline components of the
-                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
-                method. See example below for more information.
-
-        Examples:
-
-        ```py
-        from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-
-        url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
-        model = ControlNetModel.from_single_file(url)
-
-        url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors"  # can also be a local path
-        pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
-        ```
-        """
-        original_config_file = kwargs.pop("original_config_file", None)
-        config_file = kwargs.pop("config_file", None)
-        resume_download = kwargs.pop("resume_download", None)
-        force_download = kwargs.pop("force_download", False)
-        proxies = kwargs.pop("proxies", None)
-        token = kwargs.pop("token", None)
-        cache_dir = kwargs.pop("cache_dir", None)
-        local_files_only = kwargs.pop("local_files_only", None)
-        revision = kwargs.pop("revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", None)
-
-        class_name = cls.__name__
-        if (config_file is not None) and (original_config_file is not None):
-            raise ValueError(
-                "You cannot pass both `config_file` and `original_config_file` to `from_single_file`. Please use only one of these arguments."
-            )
-
-        original_config_file = config_file or original_config_file
-        original_config, checkpoint = fetch_ldm_config_and_checkpoint(
-            pretrained_model_link_or_path=pretrained_model_link_or_path,
-            class_name=class_name,
-            original_config_file=original_config_file,
-            resume_download=resume_download,
-            force_download=force_download,
-            proxies=proxies,
-            token=token,
-            revision=revision,
-            local_files_only=local_files_only,
-            cache_dir=cache_dir,
-        )
-
-        upcast_attention = kwargs.pop("upcast_attention", False)
-        image_size = kwargs.pop("image_size", None)
-
-        component = create_diffusers_controlnet_model_from_ldm(
-            class_name,
-            original_config,
-            checkpoint,
-            upcast_attention=upcast_attention,
-            image_size=image_size,
-            torch_dtype=torch_dtype,
-        )
-        controlnet = component["controlnet"]
-        if torch_dtype is not None:
-            controlnet = controlnet.to(torch_dtype)
-
-        return controlnet

src/diffusers/loaders/ip_adapter.py

@@ -90,9 +90,7 @@ class IPAdapterMixin:
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                 cached versions if they exist.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
+
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -135,7 +133,6 @@ class IPAdapterMixin:
         # Load the main state dict first.
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_download", None)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", None)
         token = kwargs.pop("token", None)
@@ -171,7 +168,6 @@ class IPAdapterMixin:
                     weights_name=weight_name,
                     cache_dir=cache_dir,
                     force_download=force_download,
-                    resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
                     token=token,

src/diffusers/loaders/lora.py

@@ -170,9 +170,7 @@ class LoraLoaderMixin:
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                 cached versions if they exist.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
+
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -194,7 +192,6 @@ class LoraLoaderMixin:
         # UNet and text encoder or both.
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_download", None)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", None)
         token = kwargs.pop("token", None)
@@ -235,7 +232,6 @@ class LoraLoaderMixin:
                         weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
                         cache_dir=cache_dir,
                         force_download=force_download,
-                        resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
                         token=token,
@@ -261,7 +257,6 @@ class LoraLoaderMixin:
                     weights_name=weight_name or LORA_WEIGHT_NAME,
                     cache_dir=cache_dir,
                     force_download=force_download,
-                    resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
                     token=token,
@@ -396,8 +391,7 @@ class LoraLoaderMixin:
         # their prefixes.
         keys = list(state_dict.keys())
         only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
-
-        if any(key.startswith(cls.unet_name) for key in keys) and not only_text_encoder:
+        if not only_text_encoder:
             # Load the layers corresponding to UNet.
             logger.info(f"Loading {cls.unet_name}.")
             unet.load_attn_procs(
@@ -1428,9 +1422,7 @@ class SD3LoraLoaderMixin:
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                 cached versions if they exist.
-            resume_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
-                incompletely downloaded files are deleted.
+
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -1451,7 +1443,6 @@ class SD3LoraLoaderMixin:
         # UNet and text encoder or both.
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
-        resume_download = kwargs.pop("resume_download", False)
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", None)
         token = kwargs.pop("token", None)
@@ -1482,7 +1473,6 @@ class SD3LoraLoaderMixin:
                         weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
                         cache_dir=cache_dir,
                         force_download=force_download,
-                        resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
                         token=token,
@@ -1504,7 +1494,6 @@ class SD3LoraLoaderMixin:
                     weights_name=weight_name or LORA_WEIGHT_NAME,
                     cache_dir=cache_dir,
                     force_download=force_download,
-                    resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
                     token=token,

src/diffusers/loaders/lora_conversion_utils.py

@@ -142,10 +142,10 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
     network_alphas = {}
 
     # Check for DoRA-enabled LoRAs.
-    if any(
-        "dora_scale" in k and ("lora_unet_" in k or "lora_te_" in k or "lora_te1_" in k or "lora_te2_" in k)
-        for k in state_dict
-    ):
+    dora_present_in_unet = any("dora_scale" in k and "lora_unet_" in k for k in state_dict)
+    dora_present_in_te = any("dora_scale" in k and ("lora_te_" in k or "lora_te1_" in k) for k in state_dict)
+    dora_present_in_te2 = any("dora_scale" in k and "lora_te2_" in k for k in state_dict)
+    if dora_present_in_unet or dora_present_in_te or dora_present_in_te2:
         if is_peft_version("<", "0.9.0"):
             raise ValueError(
                 "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
@@ -173,7 +173,7 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
             unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
 
             # Store DoRA scale if present.
-            if "dora_scale" in state_dict:
+            if dora_present_in_unet:
                 dora_scale_key_to_replace = "_lora.down." if "_lora.down." in diffusers_name else ".lora.down."
                 unet_state_dict[
                     diffusers_name.replace(dora_scale_key_to_replace, ".lora_magnitude_vector.")
@@ -192,7 +192,7 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
                 te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
 
             # Store DoRA scale if present.
-            if "dora_scale" in state_dict:
+            if dora_present_in_te or dora_present_in_te2:
                 dora_scale_key_to_replace_te = (
                     "_lora.down." if "_lora.down." in diffusers_name else ".lora_linear_layer."
                 )
@@ -214,7 +214,7 @@ def _convert_non_diffusers_lora_to_diffusers(state_dict, unet_name="unet", text_
     if len(state_dict) > 0:
         raise ValueError(f"The following keys have not been correctly renamed: \n\n {', '.join(state_dict.keys())}")
 
-    logger.info("Kohya-style checkpoint detected.")
+    logger.info("Non-diffusers checkpoint detected.")
 
     # Construct final state dict.
     unet_state_dict = {f"{unet_name}.{module_name}": params for module_name, params in unet_state_dict.items()}

src/diffusers/loaders/single_file.py

@@ -242,7 +242,6 @@ def _download_diffusers_model_config_from_hub(
     revision,
     proxies,
     force_download=None,
-    resume_download=None,
     local_files_only=None,
     token=None,
 ):
@@ -253,7 +252,6 @@ def _download_diffusers_model_config_from_hub(
         revision=revision,
         proxies=proxies,
         force_download=force_download,
-        resume_download=resume_download,
         local_files_only=local_files_only,
         token=token,
         allow_patterns=allow_patterns,
@@ -288,9 +286,7 @@ class FromSingleFileMixin:
             cache_dir (`Union[str, os.PathLike]`, *optional*):
                 Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                 is not used.
-            resume_download:
-                Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
-                of Diffusers.
+
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -352,7 +348,6 @@ class FromSingleFileMixin:
             deprecate("original_config_file", "1.0.0", deprecation_message)
             original_config = original_config_file
 
-        resume_download = kwargs.pop("resume_download", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
         token = kwargs.pop("token", None)
@@ -382,7 +377,6 @@ class FromSingleFileMixin:
 
         checkpoint = load_single_file_checkpoint(
             pretrained_model_link_or_path,
-            resume_download=resume_download,
             force_download=force_download,
             proxies=proxies,
             token=token,
@@ -412,7 +406,6 @@ class FromSingleFileMixin:
                     revision=revision,
                     proxies=proxies,
                     force_download=force_download,
-                    resume_download=resume_download,
                     local_files_only=local_files_only,
                     token=token,
                 )
@@ -435,7 +428,6 @@ class FromSingleFileMixin:
                         revision=revision,
                         proxies=proxies,
                         force_download=force_download,
-                        resume_download=resume_download,
                         local_files_only=False,
                         token=token,
                     )
@@ -555,7 +547,4 @@ class FromSingleFileMixin:
 
         pipe = pipeline_class(**init_kwargs)
 
-        if torch_dtype is not None:
-            pipe.to(dtype=torch_dtype)
-
         return pipe