Update docs/source/en/quantization/gguf.md

Co-authored-by: Aryan <aryan@huggingface.co>

.github/workflows/nightly_tests.yml

@@ -238,12 +238,13 @@ jobs:
 
   run_flax_tpu_tests:
     name: Nightly Flax TPU Tests
-    runs-on: docker-tpu
+    runs-on:
+      group: gcp-ct5lp-hightpu-8t
     if: github.event_name == 'schedule'
 
     container:
       image: diffusers/diffusers-flax-tpu
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --privileged
+      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache
     defaults:
       run:
         shell: bash
@@ -347,6 +348,66 @@ jobs:
         pip install slack_sdk tabulate
         python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
+  run_nightly_quantization_tests:
+    name: Torch quantization nightly tests
+    strategy:
+      fail-fast: false
+      max-parallel: 2
+      matrix: 
+        config:
+          - backend: "bitsandbytes"
+            test_location: "bnb"
+          - backend: "gguf"
+            test_location: "gguf"
+    runs-on:
+      group: aws-g6e-xlarge-plus
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "20gb" --ipc host --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          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 -U ${{ matrix.config.backend }}
+          python -m uv pip install pytest-reportlog
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: ${{ matrix.config.backend }} quantization tests on GPU
+        env:
+          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+          # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+          CUBLAS_WORKSPACE_CONFIG: :16:8
+          BIG_GPU_MEMORY: 40
+        run: |
+          python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+            --make-reports=tests_${{ matrix.config.backend }}_torch_cuda \
+            --report-log=tests_${{ matrix.config.backend }}_torch_cuda.log \
+            tests/quantization/${{ matrix.config.test_location }}
+      - name: Failure short reports
+        if: ${{ failure() }}
+        run: |
+          cat reports/tests_${{ matrix.config.backend }}_torch_cuda_stats.txt
+          cat reports/tests_${{ matrix.config.backend }}_torch_cuda_failures_short.txt
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: torch_cuda_${{ matrix.config.backend }}_reports
+          path: reports
+      - name: Generate Report and Notify Channel
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
 # M1 runner currently not well supported
 # TODO: (Dhruv) add these back when we setup better testing for Apple Silicon
 #  run_nightly_tests_apple_m1:
@@ -461,4 +522,4 @@ jobs:
 #        if: always()
 #        run: |
 #          pip install slack_sdk tabulate
-#          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+#          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/pr_test_peft_backend.yml

@@ -1,134 +0,0 @@
-name: Fast tests for PRs - PEFT backend
-
-on:
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "src/diffusers/**.py"
-      - "tests/**.py"
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  DIFFUSERS_IS_CI: yes
-  OMP_NUM_THREADS: 4
-  MKL_NUM_THREADS: 4
-  PYTEST_TIMEOUT: 60
-
-jobs:
-  check_code_quality:
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: make quality
-      - name: Check if failure
-        if: ${{ failure() }}
-        run: |
-          echo "Quality check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make style && make quality'" >> $GITHUB_STEP_SUMMARY
-
-  check_repository_consistency:
-    needs: check_code_quality
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check repo consistency
-        run: |
-          python utils/check_copies.py
-          python utils/check_dummies.py
-          make deps_table_check_updated
-      - name: Check if failure
-        if: ${{ failure() }}
-        run: |
-          echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
-
-  run_fast_tests:
-    needs: [check_code_quality, check_repository_consistency]
-    strategy:
-      fail-fast: false
-      matrix:
-        lib-versions: ["main", "latest"]
-
-
-    name: LoRA - ${{ matrix.lib-versions }}
-
-    runs-on:
-      group: aws-general-8-plus
-
-    container:
-      image: diffusers/diffusers-pytorch-cpu
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
-
-    defaults:
-      run:
-        shell: bash
-
-    steps:
-    - name: Checkout diffusers
-      uses: actions/checkout@v3
-      with:
-        fetch-depth: 2
-
-    - name: Install dependencies
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
-        # TODO (sayakpaul, DN6): revisit `--no-deps`
-        if [ "${{ matrix.lib-versions }}" == "main" ]; then
-            python -m pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
-            python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
-            pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
-        else
-            python -m uv pip install -U peft --no-deps
-            python -m uv pip install -U transformers accelerate --no-deps
-        fi
-
-    - name: Environment
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python utils/print_env.py
-
-    - name: Run fast PyTorch LoRA CPU tests with PEFT backend
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
-          --make-reports=tests_${{ matrix.lib-versions }} \
-          tests/lora/
-        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
-          --make-reports=tests_models_lora_${{ matrix.lib-versions }} \
-          tests/models/ -k "lora"
-
-
-    - name: Failure short reports
-      if: ${{ failure() }}
-      run: |
-        cat reports/tests_${{ matrix.lib-versions }}_failures_short.txt
-        cat reports/tests_models_lora_${{ matrix.lib-versions }}_failures_short.txt
-
-    - name: Test suite reports artifacts
-      if: ${{ always() }}
-      uses: actions/upload-artifact@v4
-      with:
-        name: pr_${{ matrix.lib-versions }}_test_reports
-        path: reports

.github/workflows/pr_tests.yml

@@ -234,3 +234,67 @@ jobs:
       with:
         name: pr_${{ matrix.config.report }}_test_reports
         path: reports
+
+  run_lora_tests:
+    needs: [check_code_quality, check_repository_consistency]
+    strategy:
+      fail-fast: false
+
+    name: LoRA tests with PEFT main
+
+    runs-on:
+      group: aws-general-8-plus
+
+    container:
+      image: diffusers/diffusers-pytorch-cpu
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
+
+    defaults:
+      run:
+        shell: bash
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        # TODO (sayakpaul, DN6): revisit `--no-deps`
+        python -m pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
+        python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+        pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
+
+    - name: Environment
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python utils/print_env.py
+
+    - name: Run fast PyTorch LoRA tests with PEFT
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
+          -s -v \
+          --make-reports=tests_peft_main \
+          tests/lora/
+        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
+          -s -v \
+          --make-reports=tests_models_lora_peft_main \
+          tests/models/ -k "lora"
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_lora_failures_short.txt
+        cat reports/tests_models_lora_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: pr_main_test_reports
+        path: reports
+

.github/workflows/push_tests.yml

@@ -161,10 +161,11 @@ jobs:
 
   flax_tpu_tests:
     name: Flax TPU Tests
-    runs-on: docker-tpu
+    runs-on:
+      group: gcp-ct5lp-hightpu-8t
     container:
       image: diffusers/diffusers-flax-tpu
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/ --privileged
+      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache 
     defaults:
       run:
         shell: bash

README.md

@@ -112,9 +112,9 @@ Check out the [Quickstart](https://huggingface.co/docs/diffusers/quicktour) to l
 | **Documentation**                                                   | **What can I learn?**                                                                                                                                                                           |
 |---------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
 | [Tutorial](https://huggingface.co/docs/diffusers/tutorials/tutorial_overview)                                                            | A basic crash course for learning how to use the library's most important features like using models and schedulers to build your own diffusion system, and training your own diffusion model.  |
-| [Loading](https://huggingface.co/docs/diffusers/using-diffusers/loading_overview)                                                             | Guides for how to load and configure all the components (pipelines, models, and schedulers) of the library, as well as how to use different schedulers.                                         |
-| [Pipelines for inference](https://huggingface.co/docs/diffusers/using-diffusers/pipeline_overview)                                             | Guides for how to use pipelines for different inference tasks, batched generation, controlling generated outputs and randomness, and how to contribute a pipeline to the library.               |
-| [Optimization](https://huggingface.co/docs/diffusers/optimization/opt_overview)                                                        | Guides for how to optimize your diffusion model to run faster and consume less memory.                                                                                                          |
+| [Loading](https://huggingface.co/docs/diffusers/using-diffusers/loading)                                                             | Guides for how to load and configure all the components (pipelines, models, and schedulers) of the library, as well as how to use different schedulers.                                         |
+| [Pipelines for inference](https://huggingface.co/docs/diffusers/using-diffusers/overview_techniques)                                             | Guides for how to use pipelines for different inference tasks, batched generation, controlling generated outputs and randomness, and how to contribute a pipeline to the library.               |
+| [Optimization](https://huggingface.co/docs/diffusers/optimization/fp16)                                                        | Guides for how to optimize your diffusion model to run faster and consume less memory.                                                                                                          |
 | [Training](https://huggingface.co/docs/diffusers/training/overview) | Guides for how to train a diffusion model for different tasks with different training techniques.                                                                                               |
 ## Contribution
 

docker/diffusers-onnxruntime-cuda/Dockerfile

@@ -28,7 +28,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     python3.10 -m uv pip install --no-cache-dir \
-        "torch<2.5.0" \
+        torch \
         torchvision \
         torchaudio \
         "onnxruntime-gpu>=1.13.1" \

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -29,7 +29,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     python3.10 -m uv pip install --no-cache-dir \
-    "torch<2.5.0" \
+    torch \
     torchvision \
     torchaudio \
     invisible_watermark && \

docker/diffusers-pytorch-cpu/Dockerfile

@@ -29,7 +29,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     python3.10 -m uv pip install --no-cache-dir \
-        "torch<2.5.0" \
+        torch \
         torchvision \
         torchaudio \
         invisible_watermark \

docker/diffusers-pytorch-cuda/Dockerfile

@@ -29,7 +29,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     python3.10 -m uv pip install --no-cache-dir \
-    "torch<2.5.0" \
+    torch \
     torchvision \
     torchaudio \
     invisible_watermark && \

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -29,7 +29,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
     python3.10 -m pip install --no-cache-dir \
-        "torch<2.5.0" \
+        torch \
         torchvision \
         torchaudio \
         invisible_watermark && \

docs/source/en/_toctree.yml

@@ -55,6 +55,8 @@
 - sections:
   - local: using-diffusers/overview_techniques
     title: Overview
+  - local: using-diffusers/create_a_server
+    title: Create a server
   - local: training/distributed_inference
     title: Distributed inference
   - local: using-diffusers/merge_loras
@@ -155,6 +157,10 @@
     title: Getting Started
   - local: quantization/bitsandbytes
     title: bitsandbytes
+  - local: quantization/gguf
+    title: gguf
+  - local: quantization/torchao
+    title: torchao
   title: Quantization Methods
 - sections:
   - local: optimization/fp16
@@ -250,6 +256,8 @@
         title: SD3ControlNetModel
       - local: api/models/controlnet_sparsectrl
         title: SparseControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
       title: ControlNets
     - sections:
       - local: api/models/allegro_transformer3d
@@ -266,10 +274,14 @@
         title: FluxTransformer2DModel
       - local: api/models/hunyuan_transformer2d
         title: HunyuanDiT2DModel
+      - local: api/models/hunyuan_video_transformer_3d
+        title: HunyuanVideoTransformer3DModel
       - local: api/models/latte_transformer3d
         title: LatteTransformer3DModel
       - local: api/models/lumina_nextdit2d
         title: LuminaNextDiT2DModel
+      - local: api/models/ltx_video_transformer3d
+        title: LTXVideoTransformer3DModel
       - local: api/models/mochi_transformer3d
         title: MochiTransformer3DModel
       - local: api/models/pixart_transformer2d
@@ -278,6 +290,8 @@
         title: PriorTransformer
       - local: api/models/sd3_transformer2d
         title: SD3Transformer2DModel
+      - local: api/models/sana_transformer2d
+        title: SanaTransformer2DModel
       - local: api/models/stable_audio_transformer
         title: StableAudioDiTModel
       - local: api/models/transformer2d
@@ -308,10 +322,16 @@
         title: AutoencoderKLAllegro
       - local: api/models/autoencoderkl_cogvideox
         title: AutoencoderKLCogVideoX
+      - local: api/models/autoencoder_kl_hunyuan_video
+        title: AutoencoderKLHunyuanVideo
+      - local: api/models/autoencoderkl_ltx_video
+        title: AutoencoderKLLTXVideo
       - local: api/models/autoencoderkl_mochi
         title: AutoencoderKLMochi
       - local: api/models/asymmetricautoencoderkl
         title: AsymmetricAutoencoderKL
+      - local: api/models/autoencoder_dc
+        title: AutoencoderDC
       - local: api/models/consistency_decoder_vae
         title: ConsistencyDecoderVAE
       - local: api/models/autoencoder_oobleck
@@ -364,6 +384,8 @@
       title: ControlNet-XS
     - local: api/pipelines/controlnetxs_sdxl
       title: ControlNet-XS with Stable Diffusion XL
+    - local: api/pipelines/controlnet_union
+      title: ControlNetUnion
     - local: api/pipelines/dance_diffusion
       title: Dance Diffusion
     - local: api/pipelines/ddim
@@ -380,6 +402,8 @@
       title: Flux
     - local: api/pipelines/hunyuandit
       title: Hunyuan-DiT
+    - local: api/pipelines/hunyuan_video
+      title: HunyuanVideo
     - local: api/pipelines/i2vgenxl
       title: I2VGen-XL
     - local: api/pipelines/pix2pix
@@ -400,6 +424,8 @@
       title: Latte
     - local: api/pipelines/ledits_pp
       title: LEDITS++
+    - local: api/pipelines/ltx_video
+      title: LTX
     - local: api/pipelines/lumina
       title: Lumina-T2X
     - local: api/pipelines/marigold
@@ -420,6 +446,8 @@
       title: PixArt-α
     - local: api/pipelines/pixart_sigma
       title: PixArt-Σ
+    - local: api/pipelines/sana
+      title: Sana
     - local: api/pipelines/self_attention_guidance
       title: Self-Attention Guidance
     - local: api/pipelines/semantic_stable_diffusion

docs/source/en/api/attnprocessor.md

@@ -15,40 +15,133 @@ specific language governing permissions and limitations under the License.
 An attention processor is a class for applying different types of attention mechanisms.
 
 ## AttnProcessor
+
 [[autodoc]] models.attention_processor.AttnProcessor
 
-## AttnProcessor2_0
 [[autodoc]] models.attention_processor.AttnProcessor2_0
 
-## AttnAddedKVProcessor
 [[autodoc]] models.attention_processor.AttnAddedKVProcessor
 
-## AttnAddedKVProcessor2_0
 [[autodoc]] models.attention_processor.AttnAddedKVProcessor2_0
 
-## CrossFrameAttnProcessor
-[[autodoc]] pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor
+[[autodoc]] models.attention_processor.AttnProcessorNPU
 
-## CustomDiffusionAttnProcessor
-[[autodoc]] models.attention_processor.CustomDiffusionAttnProcessor
-
-## CustomDiffusionAttnProcessor2_0
-[[autodoc]] models.attention_processor.CustomDiffusionAttnProcessor2_0
-
-## CustomDiffusionXFormersAttnProcessor
-[[autodoc]] models.attention_processor.CustomDiffusionXFormersAttnProcessor
-
-## FusedAttnProcessor2_0
 [[autodoc]] models.attention_processor.FusedAttnProcessor2_0
 
+## Allegro
+
+[[autodoc]] models.attention_processor.AllegroAttnProcessor2_0
+
+## AuraFlow
+
+[[autodoc]] models.attention_processor.AuraFlowAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FusedAuraFlowAttnProcessor2_0
+
+## CogVideoX
+
+[[autodoc]] models.attention_processor.CogVideoXAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FusedCogVideoXAttnProcessor2_0
+
+## CrossFrameAttnProcessor
+
+[[autodoc]] pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor
+
+## Custom Diffusion
+
+[[autodoc]] models.attention_processor.CustomDiffusionAttnProcessor
+
+[[autodoc]] models.attention_processor.CustomDiffusionAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.CustomDiffusionXFormersAttnProcessor
+
+## Flux
+
+[[autodoc]] models.attention_processor.FluxAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FusedFluxAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FluxSingleAttnProcessor2_0
+
+## Hunyuan
+
+[[autodoc]] models.attention_processor.HunyuanAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FusedHunyuanAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGHunyuanAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGCFGHunyuanAttnProcessor2_0
+
+## IdentitySelfAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGIdentitySelfAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGCFGIdentitySelfAttnProcessor2_0
+
+## IP-Adapter
+
+[[autodoc]] models.attention_processor.IPAdapterAttnProcessor
+
+[[autodoc]] models.attention_processor.IPAdapterAttnProcessor2_0
+
+## JointAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.JointAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGJointAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGCFGJointAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.FusedJointAttnProcessor2_0
+
+## LoRA
+
+[[autodoc]] models.attention_processor.LoRAAttnProcessor
+
+[[autodoc]] models.attention_processor.LoRAAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.LoRAAttnAddedKVProcessor
+
+[[autodoc]] models.attention_processor.LoRAXFormersAttnProcessor
+
+## Lumina-T2X
+
+[[autodoc]] models.attention_processor.LuminaAttnProcessor2_0
+
+## Mochi
+
+[[autodoc]] models.attention_processor.MochiAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.MochiVaeAttnProcessor2_0
+
+## Sana
+
+[[autodoc]] models.attention_processor.SanaLinearAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.SanaMultiscaleAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGCFGSanaLinearAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.PAGIdentitySanaLinearAttnProcessor2_0
+
+## Stable Audio
+
+[[autodoc]] models.attention_processor.StableAudioAttnProcessor2_0
+
 ## SlicedAttnProcessor
+
 [[autodoc]] models.attention_processor.SlicedAttnProcessor
 
-## SlicedAttnAddedKVProcessor
 [[autodoc]] models.attention_processor.SlicedAttnAddedKVProcessor
 
 ## XFormersAttnProcessor
+
 [[autodoc]] models.attention_processor.XFormersAttnProcessor
 
-## AttnProcessorNPU
-[[autodoc]] models.attention_processor.AttnProcessorNPU
+[[autodoc]] models.attention_processor.XFormersAttnAddedKVProcessor
+
+## XLAFlashAttnProcessor2_0
+
+[[autodoc]] models.attention_processor.XLAFlashAttnProcessor2_0

docs/source/en/api/loaders/lora.md

@@ -17,6 +17,9 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`StableDiffusionLoraLoaderMixin`] provides functions for loading and unloading, fusing and unfusing, enabling and disabling, and more functions for managing LoRA weights. This class can be used with any model.
 - [`StableDiffusionXLLoraLoaderMixin`] is a [Stable Diffusion (SDXL)](../../api/pipelines/stable_diffusion/stable_diffusion_xl) version of the [`StableDiffusionLoraLoaderMixin`] class for loading and saving LoRA weights. It can only be used with the SDXL model.
 - [`SD3LoraLoaderMixin`] provides similar functions for [Stable Diffusion 3](https://huggingface.co/blog/sd3).
+- [`FluxLoraLoaderMixin`] provides similar functions for [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux).
+- [`CogVideoXLoraLoaderMixin`] provides similar functions for [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox).
+- [`Mochi1LoraLoaderMixin`] provides similar functions for [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.
 
@@ -38,6 +41,18 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 
 [[autodoc]] loaders.lora_pipeline.SD3LoraLoaderMixin
 
+## FluxLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.FluxLoraLoaderMixin
+
+## CogVideoXLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.CogVideoXLoraLoaderMixin
+
+## Mochi1LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.Mochi1LoraLoaderMixin
+
 ## AmusedLoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin

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

@@ -0,0 +1,72 @@
+<!-- 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. -->
+
+# AutoencoderDC
+
+The 2D Autoencoder model used in [SANA](https://huggingface.co/papers/2410.10629) and introduced in [DCAE](https://huggingface.co/papers/2410.10733) by authors Junyu Chen\*, Han Cai\*, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, Song Han from MIT HAN Lab.
+
+The abstract from the paper is:
+
+*We present Deep Compression Autoencoder (DC-AE), a new family of autoencoder models for accelerating high-resolution diffusion models. Existing autoencoder models have demonstrated impressive results at a moderate spatial compression ratio (e.g., 8x), but fail to maintain satisfactory reconstruction accuracy for high spatial compression ratios (e.g., 64x). We address this challenge by introducing two key techniques: (1) Residual Autoencoding, where we design our models to learn residuals based on the space-to-channel transformed features to alleviate the optimization difficulty of high spatial-compression autoencoders; (2) Decoupled High-Resolution Adaptation, an efficient decoupled three-phases training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. With these designs, we improve the autoencoder's spatial compression ratio up to 128 while maintaining the reconstruction quality. Applying our DC-AE to latent diffusion models, we achieve significant speedup without accuracy drop. For example, on ImageNet 512x512, our DC-AE provides 19.1x inference speedup and 17.9x training speedup on H100 GPU for UViT-H while achieving a better FID, compared with the widely used SD-VAE-f8 autoencoder. Our code is available at [this https URL](https://github.com/mit-han-lab/efficientvit).*
+
+The following DCAE models are released and supported in Diffusers.
+
+| Diffusers format | Original format |
+|:----------------:|:---------------:|
+| [`mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers) | [`mit-han-lab/dc-ae-f32c32-sana-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.0)
+| [`mit-han-lab/dc-ae-f32c32-in-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-in-1.0-diffusers) | [`mit-han-lab/dc-ae-f32c32-in-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-in-1.0)
+| [`mit-han-lab/dc-ae-f32c32-mix-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-mix-1.0-diffusers) | [`mit-han-lab/dc-ae-f32c32-mix-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f32c32-mix-1.0)
+| [`mit-han-lab/dc-ae-f64c128-in-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f64c128-in-1.0-diffusers) | [`mit-han-lab/dc-ae-f64c128-in-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f64c128-in-1.0)
+| [`mit-han-lab/dc-ae-f64c128-mix-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f64c128-mix-1.0-diffusers) | [`mit-han-lab/dc-ae-f64c128-mix-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f64c128-mix-1.0)
+| [`mit-han-lab/dc-ae-f128c512-in-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f128c512-in-1.0-diffusers) | [`mit-han-lab/dc-ae-f128c512-in-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f128c512-in-1.0)
+| [`mit-han-lab/dc-ae-f128c512-mix-1.0-diffusers`](https://huggingface.co/mit-han-lab/dc-ae-f128c512-mix-1.0-diffusers) | [`mit-han-lab/dc-ae-f128c512-mix-1.0`](https://huggingface.co/mit-han-lab/dc-ae-f128c512-mix-1.0)
+
+This model was contributed by [lawrence-cj](https://github.com/lawrence-cj).
+
+Load a model in Diffusers format with [`~ModelMixin.from_pretrained`].
+
+```python
+from diffusers import AutoencoderDC
+
+ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", torch_dtype=torch.float32).to("cuda")
+```
+
+## Load a model in Diffusers via `from_single_file`
+
+```python
+from difusers import AutoencoderDC
+
+ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.0/blob/main/model.safetensors"
+model = AutoencoderDC.from_single_file(ckpt_path) 
+
+```
+
+The `AutoencoderDC` model has `in` and `mix` single file checkpoint variants that have matching checkpoint keys, but use different scaling factors. It is not possible for Diffusers to automatically infer the correct config file to use with the model based on just the checkpoint and will default to configuring the model using the `mix` variant config file. To override the automatically determined config, please use the `config` argument when using single file loading with `in` variant checkpoints. 
+
+```python
+from diffusers import AutoencoderDC
+
+ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f128c512-in-1.0/blob/main/model.safetensors"
+model = AutoencoderDC.from_single_file(ckpt_path, config="mit-han-lab/dc-ae-f128c512-in-1.0-diffusers")
+```
+
+
+## AutoencoderDC
+
+[[autodoc]] AutoencoderDC
+  - encode
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput
+

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

@@ -0,0 +1,32 @@
+<!-- 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. -->
+
+# AutoencoderKLHunyuanVideo
+
+The 3D variational autoencoder (VAE) model with KL loss used in [HunyuanVideo](https://github.com/Tencent/HunyuanVideo/), which was introduced in [HunyuanVideo: A Systematic Framework For Large Video Generative Models](https://huggingface.co/papers/2412.03603) by Tencent.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLHunyuanVideo
+
+vae = AutoencoderKLHunyuanVideo.from_pretrained("tencent/HunyuanVideo", torch_dtype=torch.float16)
+```
+
+## AutoencoderKLHunyuanVideo
+
+[[autodoc]] AutoencoderKLHunyuanVideo
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

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

@@ -0,0 +1,37 @@
+<!-- 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. -->
+
+# AutoencoderKLLTXVideo
+
+The 3D variational autoencoder (VAE) model with KL loss used in [LTX](https://huggingface.co/Lightricks/LTX-Video) was introduced by Lightricks.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLLTXVideo
+
+vae = AutoencoderKLLTXVideo.from_pretrained("TODO/TODO", subfolder="vae", torch_dtype=torch.float32).to("cuda")
+```
+
+## AutoencoderKLLTXVideo
+
+[[autodoc]] AutoencoderKLLTXVideo
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

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

@@ -0,0 +1,35 @@
+<!--Copyright 2024 The HuggingFace Team and The InstantX 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.
+-->
+
+# ControlNetUnionModel
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+## Loading
+
+By default the [`ControlNetUnionModel`] should be loaded with [`~ModelMixin.from_pretrained`].
+
+```py
+from diffusers import StableDiffusionXLControlNetUnionPipeline, ControlNetUnionModel
+
+controlnet = ControlNetUnionModel.from_pretrained("xinsir/controlnet-union-sdxl-1.0")
+pipe = StableDiffusionXLControlNetUnionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet)
+```
+
+## ControlNetUnionModel
+
+[[autodoc]] ControlNetUnionModel
+

docs/source/en/api/models/hunyuan_video_transformer_3d.md

@@ -0,0 +1,30 @@
+<!-- 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. -->
+
+# HunyuanVideoTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [HunyuanVideo: A Systematic Framework For Large Video Generative Models](https://huggingface.co/papers/2412.03603) by Tencent.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import HunyuanVideoTransformer3DModel
+
+transformer = HunyuanVideoTransformer3DModel.from_pretrained("tencent/HunyuanVideo", torch_dtype=torch.bfloat16)
+```
+
+## HunyuanVideoTransformer3DModel
+
+[[autodoc]] HunyuanVideoTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/models/ltx_video_transformer3d.md

@@ -0,0 +1,30 @@
+<!-- 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. -->
+
+# LTXVideoTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [LTX](https://huggingface.co/Lightricks/LTX-Video) was introduced by Lightricks.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import LTXVideoTransformer3DModel
+
+transformer = LTXVideoTransformer3DModel.from_pretrained("TODO/TODO", subfolder="transformer", torch_dtype=torch.bfloat16).to("cuda")
+```
+
+## LTXVideoTransformer3DModel
+
+[[autodoc]] LTXVideoTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/models/sana_transformer2d.md

@@ -0,0 +1,34 @@
+<!-- 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. -->
+
+# SanaTransformer2DModel
+
+A Diffusion Transformer model for 2D data from [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) was introduced from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
+
+The abstract from the paper is:
+
+*We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.*
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import SanaTransformer2DModel
+
+transformer = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_1600M_1024px_diffusers", subfolder="transformer", torch_dtype=torch.float16)
+```
+
+## SanaTransformer2DModel
+
+[[autodoc]] SanaTransformer2DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

@@ -29,16 +29,32 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
 
-There are two models available that can be used with the text-to-video and video-to-video CogVideoX pipelines:
-- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b): The recommended dtype for running this model is `fp16`.
-- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b): The recommended dtype for running this model is `bf16`.
+There are three official CogVideoX checkpoints for text-to-video and video-to-video.
 
-There is one model available that can be used with the image-to-video CogVideoX pipeline:
-- [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V): The recommended dtype for running this model is `bf16`.
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b) | torch.float16 |
+| [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b) | torch.bfloat16 |
+| [`THUDM/CogVideoX1.5-5b`](https://huggingface.co/THUDM/CogVideoX1.5-5b) | torch.bfloat16 |
 
-There are two models that support pose controllable generation (by the [Alibaba-PAI](https://huggingface.co/alibaba-pai) team):
-- [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose): The recommended dtype for running this model is `bf16`.
-- [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose): The recommended dtype for running this model is `bf16`.
+There are two official CogVideoX checkpoints available for image-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V) | torch.bfloat16 |
+| [`THUDM/CogVideoX-1.5-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-1.5-5b-I2V) | torch.bfloat16 |
+
+For the CogVideoX 1.5 series:
+- Text-to-video (T2V) works best at a resolution of 1360x768 because it was trained with that specific resolution.
+- Image-to-video (I2V) works for multiple resolutions. The width can vary from 768 to 1360, but the height must be 768. The height/width must be divisible by 16.
+- Both T2V and I2V models support generation with 81 and 161 frames and work best at this value. Exporting videos at 16 FPS is recommended.
+
+There are two official CogVideoX checkpoints that support pose controllable generation (by the [Alibaba-PAI](https://huggingface.co/alibaba-pai) team).
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose) | torch.bfloat16 |
+| [`alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose`](https://huggingface.co/alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose) | torch.bfloat16 |
 
 ## Inference
 

docs/source/en/api/pipelines/controlnet_sd3.md

@@ -28,6 +28,7 @@ This controlnet code is mainly implemented by [The InstantX Team](https://huggin
 | ControlNet type | Developer | Link |
 | -------- | ---------- | ---- |
 | Canny | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Canny) |
+| Depth | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Depth) |
 | Pose | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Pose) |
 | Tile | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/SD3-Controlnet-Tile) |
 | Inpainting | [The AlimamaCreative Team](https://huggingface.co/alimama-creative) | [link](https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting) |

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

@@ -0,0 +1,35 @@
+<!--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.
+-->
+
+# ControlNetUnion
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+
+## StableDiffusionXLControlNetUnionPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionImg2ImgPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionImg2ImgPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionInpaintPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionInpaintPipeline
+	- all
+	- __call__

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

@@ -22,12 +22,20 @@ Flux can be quite expensive to run on consumer hardware devices. However, you ca
 
 </Tip>
 
-Flux comes in two variants:
+Flux comes in the following variants:
 
-* Timestep-distilled (`black-forest-labs/FLUX.1-schnell`)
-* Guidance-distilled (`black-forest-labs/FLUX.1-dev`)
+| model type | model id |
+|:----------:|:--------:|
+| Timestep-distilled | [`black-forest-labs/FLUX.1-schnell`](https://huggingface.co/black-forest-labs/FLUX.1-schnell) |
+| Guidance-distilled | [`black-forest-labs/FLUX.1-dev`](https://huggingface.co/black-forest-labs/FLUX.1-dev) |
+| Fill Inpainting/Outpainting (Guidance-distilled) | [`black-forest-labs/FLUX.1-Fill-dev`](https://huggingface.co/black-forest-labs/FLUX.1-Fill-dev) |
+| Canny Control (Guidance-distilled) | [`black-forest-labs/FLUX.1-Canny-dev`](https://huggingface.co/black-forest-labs/FLUX.1-Canny-dev) |
+| Depth Control (Guidance-distilled) | [`black-forest-labs/FLUX.1-Depth-dev`](https://huggingface.co/black-forest-labs/FLUX.1-Depth-dev) |
+| Canny Control (LoRA) | [`black-forest-labs/FLUX.1-Canny-dev-lora`](https://huggingface.co/black-forest-labs/FLUX.1-Canny-dev-lora) |
+| Depth Control (LoRA) | [`black-forest-labs/FLUX.1-Depth-dev-lora`](https://huggingface.co/black-forest-labs/FLUX.1-Depth-dev-lora) |
+| Redux (Adapter) | [`black-forest-labs/FLUX.1-Redux-dev`](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev) |
 
-Both checkpoints have slightly difference usage which we detail below.
+All checkpoints have different usage which we detail below.
 
 ### Timestep-distilled
 
@@ -77,7 +85,191 @@ out = pipe(
 out.save("image.png")
 ```
 
+### Fill Inpainting/Outpainting
+
+* Flux Fill pipeline does not require `strength` as an input like regular inpainting pipelines.
+* It supports both inpainting and outpainting.
+
+```python
+import torch
+from diffusers import FluxFillPipeline
+from diffusers.utils import load_image
+
+image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/cup.png")
+mask = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/cup_mask.png")
+
+repo_id = "black-forest-labs/FLUX.1-Fill-dev"
+pipe = FluxFillPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16).to("cuda")
+
+image = pipe(
+    prompt="a white paper cup",
+    image=image,
+    mask_image=mask,
+    height=1632,
+    width=1232,
+    max_sequence_length=512,
+    generator=torch.Generator("cpu").manual_seed(0)
+).images[0]
+image.save(f"output.png")
+```
+
+### Canny Control
+
+**Note:** `black-forest-labs/Flux.1-Canny-dev` is _not_ a [`ControlNetModel`] model. ControlNet models are a separate component from the UNet/Transformer whose residuals are added to the actual underlying model. Canny Control is an alternate architecture that achieves effectively the same results as a ControlNet model would, by using channel-wise concatenation with input control condition and ensuring the transformer learns structure control by following the condition as closely as possible. 
+
+```python
+# !pip install -U controlnet-aux
+import torch
+from controlnet_aux import CannyDetector
+from diffusers import FluxControlPipeline
+from diffusers.utils import load_image
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-Canny-dev", torch_dtype=torch.bfloat16).to("cuda")
+
+prompt = "A robot made of exotic candies and chocolates of different kinds. The background is filled with confetti and celebratory gifts."
+control_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png")
+
+processor = CannyDetector()
+control_image = processor(control_image, low_threshold=50, high_threshold=200, detect_resolution=1024, image_resolution=1024)
+
+image = pipe(
+    prompt=prompt,
+    control_image=control_image,
+    height=1024,
+    width=1024,
+    num_inference_steps=50,
+    guidance_scale=30.0,
+).images[0]
+image.save("output.png")
+```
+
+Canny Control is also possible with a LoRA variant of this condition. The usage is as follows:
+
+```python
+# !pip install -U controlnet-aux
+import torch
+from controlnet_aux import CannyDetector
+from diffusers import FluxControlPipeline
+from diffusers.utils import load_image
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("black-forest-labs/FLUX.1-Canny-dev-lora")
+
+prompt = "A robot made of exotic candies and chocolates of different kinds. The background is filled with confetti and celebratory gifts."
+control_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png")
+
+processor = CannyDetector()
+control_image = processor(control_image, low_threshold=50, high_threshold=200, detect_resolution=1024, image_resolution=1024)
+
+image = pipe(
+    prompt=prompt,
+    control_image=control_image,
+    height=1024,
+    width=1024,
+    num_inference_steps=50,
+    guidance_scale=30.0,
+).images[0]
+image.save("output.png")
+```
+
+### Depth Control
+
+**Note:** `black-forest-labs/Flux.1-Depth-dev` is _not_ a ControlNet model. [`ControlNetModel`] models are a separate component from the UNet/Transformer whose residuals are added to the actual underlying model. Depth Control is an alternate architecture that achieves effectively the same results as a ControlNet model would, by using channel-wise concatenation with input control condition and ensuring the transformer learns structure control by following the condition as closely as possible.
+
+```python
+# !pip install git+https://github.com/huggingface/image_gen_aux
+import torch
+from diffusers import FluxControlPipeline, FluxTransformer2DModel
+from diffusers.utils import load_image
+from image_gen_aux import DepthPreprocessor
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-Depth-dev", torch_dtype=torch.bfloat16).to("cuda")
+
+prompt = "A robot made of exotic candies and chocolates of different kinds. The background is filled with confetti and celebratory gifts."
+control_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png")
+
+processor = DepthPreprocessor.from_pretrained("LiheYoung/depth-anything-large-hf")
+control_image = processor(control_image)[0].convert("RGB")
+
+image = pipe(
+    prompt=prompt,
+    control_image=control_image,
+    height=1024,
+    width=1024,
+    num_inference_steps=30,
+    guidance_scale=10.0,
+    generator=torch.Generator().manual_seed(42),
+).images[0]
+image.save("output.png")
+```
+
+Depth Control is also possible with a LoRA variant of this condition. The usage is as follows:
+
+```python
+# !pip install git+https://github.com/huggingface/image_gen_aux
+import torch
+from diffusers import FluxControlPipeline, FluxTransformer2DModel
+from diffusers.utils import load_image
+from image_gen_aux import DepthPreprocessor
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("black-forest-labs/FLUX.1-Depth-dev-lora")
+
+prompt = "A robot made of exotic candies and chocolates of different kinds. The background is filled with confetti and celebratory gifts."
+control_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png")
+
+processor = DepthPreprocessor.from_pretrained("LiheYoung/depth-anything-large-hf")
+control_image = processor(control_image)[0].convert("RGB")
+
+image = pipe(
+    prompt=prompt,
+    control_image=control_image,
+    height=1024,
+    width=1024,
+    num_inference_steps=30,
+    guidance_scale=10.0,
+    generator=torch.Generator().manual_seed(42),
+).images[0]
+image.save("output.png")
+```
+
+### Redux
+
+* Flux Redux pipeline is an adapter for FLUX.1 base models. It can be used with both flux-dev and flux-schnell, for image-to-image generation.
+* You can first use the `FluxPriorReduxPipeline` to get the `prompt_embeds` and `pooled_prompt_embeds`, and then feed them into the `FluxPipeline` for image-to-image generation.
+* When use `FluxPriorReduxPipeline` with a base pipeline, you can set `text_encoder=None` and `text_encoder_2=None` in the base pipeline, in order to save VRAM.
+
+```python
+import torch
+from diffusers import FluxPriorReduxPipeline, FluxPipeline
+from diffusers.utils import load_image
+device = "cuda"
+dtype = torch.bfloat16
+
+
+repo_redux = "black-forest-labs/FLUX.1-Redux-dev"
+repo_base = "black-forest-labs/FLUX.1-dev" 
+pipe_prior_redux = FluxPriorReduxPipeline.from_pretrained(repo_redux, torch_dtype=dtype).to(device)
+pipe = FluxPipeline.from_pretrained(
+    repo_base, 
+    text_encoder=None,
+    text_encoder_2=None,
+    torch_dtype=torch.bfloat16
+).to(device)
+
+image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/style_ziggy/img5.png")
+pipe_prior_output = pipe_prior_redux(image)
+images = pipe(
+    guidance_scale=2.5,
+    num_inference_steps=50,
+    generator=torch.Generator("cpu").manual_seed(0),
+    **pipe_prior_output,
+).images
+images[0].save("flux-redux.png")
+```
+
 ## Running FP16 inference
+
 Flux can generate high-quality images with FP16 (i.e. to accelerate inference on Turing/Volta GPUs) but produces different outputs compared to FP32/BF16. The issue is that some activations in the text encoders have to be clipped when running in FP16, which affects the overall image. Forcing text encoders to run with FP32 inference thus removes this output difference. See [here](https://github.com/huggingface/diffusers/pull/9097#issuecomment-2272292516) for details.
 
 FP16 inference code:
@@ -188,3 +380,27 @@ image.save("flux-fp8-dev.png")
 [[autodoc]] FluxControlNetImg2ImgPipeline
 	- all
 	- __call__
+
+## FluxControlPipeline
+
+[[autodoc]] FluxControlPipeline
+	- all
+	- __call__
+
+## FluxControlImg2ImgPipeline
+
+[[autodoc]] FluxControlImg2ImgPipeline
+	- all
+	- __call__
+
+## FluxPriorReduxPipeline
+
+[[autodoc]] FluxPriorReduxPipeline
+	- all
+	- __call__
+
+## FluxFillPipeline
+
+[[autodoc]] FluxFillPipeline
+	- all
+	- __call__

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

@@ -0,0 +1,43 @@
+<!-- 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. -->
+
+# HunyuanVideo
+
+[HunyuanVideo](https://www.arxiv.org/abs/2412.03603) by Tencent.
+
+*Recent advancements in video generation have significantly impacted daily life for both individuals and industries. However, the leading video generation models remain closed-source, resulting in a notable performance gap between industry capabilities and those available to the public. In this report, we introduce HunyuanVideo, an innovative open-source video foundation model that demonstrates performance in video generation comparable to, or even surpassing, that of leading closed-source models. HunyuanVideo encompasses a comprehensive framework that integrates several key elements, including data curation, advanced architectural design, progressive model scaling and training, and an efficient infrastructure tailored for large-scale model training and inference. As a result, we successfully trained a video generative model with over 13 billion parameters, making it the largest among all open-source models. We conducted extensive experiments and implemented a series of targeted designs to ensure high visual quality, motion dynamics, text-video alignment, and advanced filming techniques. According to evaluations by professionals, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models. By releasing the code for the foundation model and its applications, we aim to bridge the gap between closed-source and open-source communities. This initiative will empower individuals within the community to experiment with their ideas, fostering a more dynamic and vibrant video generation ecosystem. The code is publicly available at [this https URL](https://github.com/Tencent/HunyuanVideo).*
+
+<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>
+
+Recommendations for inference:
+- Both text encoders should be in `torch.float16`.
+- Transformer should be in `torch.bfloat16`.
+- VAE should be in `torch.float16`.
+- `num_frames` should be of the form `4 * k + 1`, for example `49` or `129`.
+- For smaller resolution images, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution images, try higher values (between `7.0` and `12.0`). The default value is `7.0` for HunyuanVideo.
+- For more information about supported resolutions and other details, please refer to the original repository [here](https://github.com/Tencent/HunyuanVideo/).
+
+## HunyuanVideoPipeline
+
+[[autodoc]] HunyuanVideoPipeline
+  - all
+  - __call__
+
+## HunyuanVideoPipelineOutput
+
+[[autodoc]] pipelines.hunyuan_video.pipeline_output.HunyuanVideoPipelineOutput

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

@@ -0,0 +1,80 @@
+<!-- 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. -->
+
+# LTX
+
+[LTX Video](https://huggingface.co/Lightricks/LTX-Video) is the first DiT-based video generation model capable of generating high-quality videos in real-time. It produces 24 FPS videos at a 768x512 resolution faster than they can be watched. Trained on a large-scale dataset of diverse videos, the model generates high-resolution videos with realistic and varied content. We provide a model for both text-to-video as well as image + text-to-video usecases.
+
+<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>
+
+## Loading Single Files
+
+Loading the original LTX Video checkpoints is also possible with [`~ModelMixin.from_single_file`].
+
+```python
+import torch
+from diffusers import AutoencoderKLLTXVideo, LTXImageToVideoPipeline, LTXVideoTransformer3DModel
+
+single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
+transformer = LTXVideoTransformer3DModel.from_single_file(
+  single_file_url, torch_dtype=torch.bfloat16
+)
+vae = AutoencoderKLLTXVideo.from_single_file(single_file_url, torch_dtype=torch.bfloat16)
+pipe = LTXImageToVideoPipeline.from_pretrained(
+  "Lightricks/LTX-Video", transformer=transformer, vae=vae, torch_dtype=torch.bfloat16
+)
+
+# ... inference code ...
+```
+
+Alternatively, the pipeline can be used to load the weights with [`~FromSingleFileMixin.from_single_file`].
+
+```python
+import torch
+from diffusers import LTXImageToVideoPipeline
+from transformers import T5EncoderModel, T5Tokenizer
+
+single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
+text_encoder = T5EncoderModel.from_pretrained(
+  "Lightricks/LTX-Video", subfolder="text_encoder", torch_dtype=torch.bfloat16
+)
+tokenizer = T5Tokenizer.from_pretrained(
+  "Lightricks/LTX-Video", subfolder="tokenizer", torch_dtype=torch.bfloat16
+)
+pipe = LTXImageToVideoPipeline.from_single_file(
+  single_file_url, text_encoder=text_encoder, tokenizer=tokenizer, torch_dtype=torch.bfloat16
+)
+```
+
+Refer to [this section](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox#memory-optimization) to learn more about optimizing memory consumption.
+
+## LTXPipeline
+
+[[autodoc]] LTXPipeline
+  - all
+  - __call__
+
+## LTXImageToVideoPipeline
+
+[[autodoc]] LTXImageToVideoPipeline
+  - all
+  - __call__
+
+## LTXPipelineOutput
+
+[[autodoc]] pipelines.ltx.pipeline_output.LTXPipelineOutput

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

@@ -48,6 +48,11 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
   - all
   - __call__
 
+## StableDiffusionPAGInpaintPipeline
+[[autodoc]] StableDiffusionPAGInpaintPipeline
+	- all
+	- __call__
+
 ## StableDiffusionPAGPipeline
 [[autodoc]] StableDiffusionPAGPipeline
 	- all
@@ -96,6 +101,10 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
 	- all
 	- __call__
 
+## StableDiffusion3PAGImg2ImgPipeline
+[[autodoc]] StableDiffusion3PAGImg2ImgPipeline
+	- all
+	- __call__
 
 ## PixArtSigmaPAGPipeline
 [[autodoc]] PixArtSigmaPAGPipeline

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

@@ -0,0 +1,67 @@
+<!-- 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. -->
+
+# SanaPipeline
+
+[SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
+
+The abstract from the paper is:
+
+*We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.*
+
+<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>
+
+This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj) and [chenjy2003](https://github.com/chenjy2003). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model).
+
+Available models:
+
+| Model | Recommended dtype |
+|:-----:|:-----------------:|
+| [`Efficient-Large-Model/Sana_1600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers) | `torch.bfloat16` |
+| [`Efficient-Large-Model/Sana_1600M_512px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_600M_1024px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_600M_512px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_600M_512px_diffusers) | `torch.float16` |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-673efba2a57ed99843f11f9e) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+<Tip>
+
+Make sure to pass the `variant` argument for downloaded checkpoints to use lower disk space. Set it to `"fp16"` for models with recommended dtype as `torch.float16`, and `"bf16"` for models with recommended dtype as `torch.bfloat16`. By default, `torch.float32` weights are downloaded, which use twice the amount of disk storage. Additionally, `torch.float32` weights can be downcasted on-the-fly by specifying the `torch_dtype` argument. Read about it in the [docs](https://huggingface.co/docs/diffusers/v0.31.0/en/api/pipelines/overview#diffusers.DiffusionPipeline.from_pretrained).
+
+</Tip>
+
+## SanaPipeline
+
+[[autodoc]] SanaPipeline
+  - all
+  - __call__
+
+## SanaPAGPipeline
+
+[[autodoc]] SanaPAGPipeline
+  - all
+  - __call__
+
+## SanaPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput

docs/source/en/api/quantization.md

@@ -28,6 +28,13 @@ Learn how to quantize models in the [Quantization](../quantization/overview) gui
 
 [[autodoc]] BitsAndBytesConfig
 
+## GGUFQuantizationConfig
+
+[[autodoc]] GGUFQuantizationConfig
+## TorchAoConfig
+
+[[autodoc]] TorchAoConfig
+
 ## DiffusersQuantizer
 
 [[autodoc]] quantizers.base.DiffusersQuantizer

docs/source/en/conceptual/evaluation.md

@@ -181,7 +181,7 @@ Then we load the [v1-5 checkpoint](https://huggingface.co/stable-diffusion-v1-5/
 
 ```python
 model_ckpt_1_5 = "stable-diffusion-v1-5/stable-diffusion-v1-5"
-sd_pipeline_1_5 = StableDiffusionPipeline.from_pretrained(model_ckpt_1_5, torch_dtype=weight_dtype).to(device)
+sd_pipeline_1_5 = StableDiffusionPipeline.from_pretrained(model_ckpt_1_5, torch_dtype=torch.float16).to("cuda")
 
 images_1_5 = sd_pipeline_1_5(prompts, num_images_per_prompt=1, generator=generator, output_type="np").images
 ```
@@ -280,7 +280,7 @@ from diffusers import StableDiffusionInstructPix2PixPipeline
 
 instruct_pix2pix_pipeline = StableDiffusionInstructPix2PixPipeline.from_pretrained(
     "timbrooks/instruct-pix2pix", torch_dtype=torch.float16
-).to(device)
+).to("cuda")
 ```
 
 Now, we perform the edits:
@@ -326,9 +326,9 @@ from transformers import (
 
 clip_id = "openai/clip-vit-large-patch14"
 tokenizer = CLIPTokenizer.from_pretrained(clip_id)
-text_encoder = CLIPTextModelWithProjection.from_pretrained(clip_id).to(device)
+text_encoder = CLIPTextModelWithProjection.from_pretrained(clip_id).to("cuda")
 image_processor = CLIPImageProcessor.from_pretrained(clip_id)
-image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_id).to(device)
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_id).to("cuda")
 ```
 
 Notice that we are using a particular CLIP checkpoint, i.e., `openai/clip-vit-large-patch14`. This is because the Stable Diffusion pre-training was performed with this CLIP variant. For more details, refer to the [documentation](https://huggingface.co/docs/transformers/model_doc/clip).
@@ -350,7 +350,7 @@ class DirectionalSimilarity(nn.Module):
 
     def preprocess_image(self, image):
         image = self.image_processor(image, return_tensors="pt")["pixel_values"]
-        return {"pixel_values": image.to(device)}
+        return {"pixel_values": image.to("cuda")}
 
     def tokenize_text(self, text):
         inputs = self.tokenizer(
@@ -360,7 +360,7 @@ class DirectionalSimilarity(nn.Module):
             truncation=True,
             return_tensors="pt",
         )
-        return {"input_ids": inputs.input_ids.to(device)}
+        return {"input_ids": inputs.input_ids.to("cuda")}
 
     def encode_image(self, image):
         preprocessed_image = self.preprocess_image(image)
@@ -459,6 +459,7 @@ with ZipFile(local_filepath, "r") as zipper:
 ```python
 from PIL import Image
 import os
+import numpy as np
 
 dataset_path = "sample-imagenet-images"
 image_paths = sorted([os.path.join(dataset_path, x) for x in os.listdir(dataset_path)])
@@ -477,6 +478,7 @@ Now that the images are loaded, let's apply some lightweight pre-processing on t
 
 ```python
 from torchvision.transforms import functional as F
+import torch
 
 
 def preprocess_image(image):
@@ -498,6 +500,10 @@ dit_pipeline = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", torch_dtype=
 dit_pipeline.scheduler = DPMSolverMultistepScheduler.from_config(dit_pipeline.scheduler.config)
 dit_pipeline = dit_pipeline.to("cuda")
 
+seed = 0
+generator = torch.manual_seed(seed)
+
+
 words = [
     "cassette player",
     "chainsaw",

docs/source/en/quantization/bitsandbytes.md

@@ -17,6 +17,12 @@ specific language governing permissions and limitations under the License.
 
 4-bit quantization compresses a model even further, and it is commonly used with [QLoRA](https://hf.co/papers/2305.14314) to finetune quantized LLMs.
 
+This guide demonstrates how quantization can enable running
+[FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)
+on less than 16GB of VRAM and even on a free Google
+Colab instance.
+
+![comparison image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/comparison.png)
 
 To use bitsandbytes, make sure you have the following libraries installed:
 
@@ -31,70 +37,167 @@ Now you can quantize a model by passing a [`BitsAndBytesConfig`] to [`~ModelMixi
 
 Quantizing a model in 8-bit halves the memory-usage:
 
+bitsandbytes is supported in both Transformers and Diffusers, so you can quantize both the
+[`FluxTransformer2DModel`] and [`~transformers.T5EncoderModel`].
+
+For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bfloat16`.
+
+> [!TIP]
+> The [`CLIPTextModel`] and [`AutoencoderKL`] aren't quantized because they're already small in size and because [`AutoencoderKL`] only has a few `torch.nn.Linear` layers.
+
 ```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
+from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
 
-quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+from diffusers import FluxTransformer2DModel
+from transformers import T5EncoderModel
 
-model_8bit = FluxTransformer2DModel.from_pretrained(
-    "black-forest-labs/FLUX.1-dev", 
+quant_config = TransformersBitsAndBytesConfig(load_in_8bit=True,)
+
+text_encoder_2_8bit = T5EncoderModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="text_encoder_2",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True,)
+
+transformer_8bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=quantization_config
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
 )
 ```
 
-By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter if you want:
+By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.
 
-```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
-
-quantization_config = BitsAndBytesConfig(load_in_8bit=True)
-
-model_8bit = FluxTransformer2DModel.from_pretrained(
-    "black-forest-labs/FLUX.1-dev", 
+```diff
+transformer_8bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=quantization_config,
-    torch_dtype=torch.float32
+    quantization_config=quant_config,
++   torch_dtype=torch.float32,
 )
-model_8bit.transformer_blocks.layers[-1].norm2.weight.dtype
 ```
 
-Once a model is quantized, you can push the model to the Hub with the [`~ModelMixin.push_to_hub`] method. The quantization `config.json` file is pushed first, followed by the quantized model weights. You can also save the serialized 4-bit models locally with [`~ModelMixin.save_pretrained`].
+Let's generate an image using our quantized models.
+
+Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the
+CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory.
+
+```py
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    transformer=transformer_8bit,
+    text_encoder_2=text_encoder_2_8bit,
+    torch_dtype=torch.float16,
+    device_map="auto",
+)
+
+pipe_kwargs = {
+    "prompt": "A cat holding a sign that says hello world",
+    "height": 1024,
+    "width": 1024,
+    "guidance_scale": 3.5,
+    "num_inference_steps": 50,
+    "max_sequence_length": 512,
+}
+
+image = pipe(**pipe_kwargs, generator=torch.manual_seed(0),).images[0]
+```
+
+<div class="flex justify-center">
+   <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/8bit.png"/>
+</div>
+
+When there is enough memory, you can also directly move the pipeline to the GPU with `.to("cuda")` and apply [`~DiffusionPipeline.enable_model_cpu_offload`] to optimize GPU memory usage.
+
+Once a model is quantized, you can push the model to the Hub with the [`~ModelMixin.push_to_hub`] method. The quantization `config.json` file is pushed first, followed by the quantized model weights. You can also save the serialized 8-bit models locally with [`~ModelMixin.save_pretrained`].
 
 </hfoption>
 <hfoption id="4-bit">
 
 Quantizing a model in 4-bit reduces your memory-usage by 4x:
 
+bitsandbytes is supported in both Transformers and Diffusers, so you can can quantize both the
+[`FluxTransformer2DModel`] and [`~transformers.T5EncoderModel`].
+
+For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bfloat16`.
+
+> [!TIP]
+> The [`CLIPTextModel`] and [`AutoencoderKL`] aren't quantized because they're already small in size and because [`AutoencoderKL`] only has a few `torch.nn.Linear` layers.
+
 ```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
+from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
 
-quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+from diffusers import FluxTransformer2DModel
+from transformers import T5EncoderModel
 
-model_4bit = FluxTransformer2DModel.from_pretrained(
-    "black-forest-labs/FLUX.1-dev", 
+quant_config = TransformersBitsAndBytesConfig(load_in_4bit=True,)
+
+text_encoder_2_4bit = T5EncoderModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="text_encoder_2",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True,)
+
+transformer_4bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=quantization_config
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
 )
 ```
 
-By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter if you want:
+By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.
 
-```py
-from diffusers import FluxTransformer2DModel, BitsAndBytesConfig
-
-quantization_config = BitsAndBytesConfig(load_in_4bit=True)
-
-model_4bit = FluxTransformer2DModel.from_pretrained(
-    "black-forest-labs/FLUX.1-dev", 
+```diff
+transformer_4bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=quantization_config,
-    torch_dtype=torch.float32
+    quantization_config=quant_config,
++   torch_dtype=torch.float32,
 )
-model_4bit.transformer_blocks.layers[-1].norm2.weight.dtype
 ```
 
-Call [`~ModelMixin.push_to_hub`] after loading it in 4-bit precision. You can also save the serialized 4-bit models locally with [`~ModelMixin.save_pretrained`].  
+Let's generate an image using our quantized models.
+
+Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory.
+
+```py
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    transformer=transformer_4bit,
+    text_encoder_2=text_encoder_2_4bit,
+    torch_dtype=torch.float16,
+    device_map="auto",
+)
+
+pipe_kwargs = {
+    "prompt": "A cat holding a sign that says hello world",
+    "height": 1024,
+    "width": 1024,
+    "guidance_scale": 3.5,
+    "num_inference_steps": 50,
+    "max_sequence_length": 512,
+}
+
+image = pipe(**pipe_kwargs, generator=torch.manual_seed(0),).images[0]
+```
+
+<div class="flex justify-center">
+   <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/4bit.png"/>
+</div>
+
+When there is enough memory, you can also directly move the pipeline to the GPU with `.to("cuda")` and apply [`~DiffusionPipeline.enable_model_cpu_offload`] to optimize GPU memory usage.
+
+Once a model is quantized, you can push the model to the Hub with the [`~ModelMixin.push_to_hub`] method. The quantization `config.json` file is pushed first, followed by the quantized model weights. You can also save the serialized 4-bit models locally with [`~ModelMixin.save_pretrained`].
 
 </hfoption>
 </hfoptions>
@@ -199,17 +302,34 @@ quantization_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dty
 NF4 is a 4-bit data type from the [QLoRA](https://hf.co/papers/2305.14314) paper, adapted for weights initialized from a normal distribution. You should use NF4 for training 4-bit base models. This can be configured with the `bnb_4bit_quant_type` parameter in the [`BitsAndBytesConfig`]:
 
 ```py
-from diffusers import BitsAndBytesConfig
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
+from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
 
-nf4_config = BitsAndBytesConfig(
+from diffusers import FluxTransformer2DModel
+from transformers import T5EncoderModel
+
+quant_config = TransformersBitsAndBytesConfig(
     load_in_4bit=True,
     bnb_4bit_quant_type="nf4",
 )
 
-model_nf4 = SD3Transformer2DModel.from_pretrained(
-    "stabilityai/stable-diffusion-3-medium-diffusers",
+text_encoder_2_4bit = T5EncoderModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="text_encoder_2",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+)
+
+transformer_4bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=nf4_config,
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
 )
 ```
 
@@ -220,38 +340,74 @@ For inference, the `bnb_4bit_quant_type` does not have a huge impact on performa
 Nested quantization is a technique that can save additional memory at no additional performance cost. This feature performs a second quantization of the already quantized weights to save an additional 0.4 bits/parameter. 
 
 ```py
-from diffusers import BitsAndBytesConfig
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
+from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
 
-double_quant_config = BitsAndBytesConfig(
+from diffusers import FluxTransformer2DModel
+from transformers import T5EncoderModel
+
+quant_config = TransformersBitsAndBytesConfig(
     load_in_4bit=True,
     bnb_4bit_use_double_quant=True,
 )
 
-double_quant_model = SD3Transformer2DModel.from_pretrained(
-    "stabilityai/stable-diffusion-3-medium-diffusers",
+text_encoder_2_4bit = T5EncoderModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="text_encoder_2",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+)
+
+transformer_4bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
     subfolder="transformer",
-    quantization_config=double_quant_config,
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
 )
 ```
 
 ## Dequantizing `bitsandbytes` models
 
-Once quantized, you can dequantize the model to the original precision but this might result in a small quality loss of the model. Make sure you have enough GPU RAM to fit the dequantized model. 
+Once quantized, you can dequantize a model to its original precision, but this might result in a small loss of quality. Make sure you have enough GPU RAM to fit the dequantized model. 
 
 ```python
-from diffusers import BitsAndBytesConfig
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
+from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
 
-double_quant_config = BitsAndBytesConfig(
+from diffusers import FluxTransformer2DModel
+from transformers import T5EncoderModel
+
+quant_config = TransformersBitsAndBytesConfig(
     load_in_4bit=True,
     bnb_4bit_use_double_quant=True,
 )
 
-double_quant_model = SD3Transformer2DModel.from_pretrained(
-    "stabilityai/stable-diffusion-3-medium-diffusers",
-    subfolder="transformer",
-    quantization_config=double_quant_config,
+text_encoder_2_4bit = T5EncoderModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="text_encoder_2",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
 )
-model.dequantize()
+
+quant_config = DiffusersBitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+)
+
+transformer_4bit = FluxTransformer2DModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+text_encoder_2_4bit.dequantize()
+transformer_4bit.dequantize()
 ```
 
 ## Resources

docs/source/en/quantization/gguf.md

@@ -0,0 +1,70 @@
+<!--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.
+
+-->
+
+# GGUF
+
+The GGUF file format is typically used to store models for inference with [GGML](https://github.com/ggerganov/ggml) and supports a variety of block wise quantization options. Diffusers supports loading checkpoints prequantized and saved in the GGUF format via `from_single_file` loading with Model classes. Loading GGUF checkpoints via Pipelines is currently not supported.
+
+The following example will load the [FLUX.1 DEV](https://huggingface.co/black-forest-labs/FLUX.1-dev) transformer model using the GGUF Q2_K quantization variant.
+
+Before starting please install gguf in your environment
+
+```shell
+pip install -U gguf
+```
+
+Since GGUF is a single file format, use [`~FromSingleFileMixin.from_single_file`] to load the model and pass in the [`GGUFQuantizationConfig`].
+
+When using GGUF checkpoints, the quantized weights remain in a low memory `dtype`(typically `torch.uint8`) and are dynamically dequantized and cast to the configured `compute_dtype` during each module's forward pass through the model. The `GGUFQuantizationConfig` allows you to set the `compute_dtype`.
+
+The functions used for dynamic dequantizatation are based on the great work done by [city96](https://github.com/city96/ComfyUI-GGUF), who created the Pytorch ports of the original [`numpy`](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/quants.py) implementation by [compilade](https://github.com/compilade).
+
+```python
+import torch
+
+from diffusers import FluxPipeline, FluxTransformer2DModel, GGUFQuantizationConfig
+
+ckpt_path = (
+    "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q2_K.gguf"
+)
+transformer = FluxTransformer2DModel.from_single_file(
+    ckpt_path,
+    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
+    torch_dtype=torch.bfloat16,
+)
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    transformer=transformer,
+    generator=torch.manual_seed(0),
+    torch_dtype=torch.bfloat16,
+)
+pipe.enable_model_cpu_offload()
+prompt = "A cat holding a sign that says hello world"
+image = pipe(prompt).images[0]
+image.save("flux-gguf.png")
+```
+
+## Supported Quantization Types
+
+- BF16
+- Q4_0
+- Q4_1
+- Q5_0
+- Q5_1
+- Q8_0
+- Q2_K
+- Q3_K
+- Q4_K
+- Q5_K
+- Q6_K
+

docs/source/en/quantization/overview.md

@@ -17,7 +17,7 @@ Quantization techniques focus on representing data with less information while a
 
 <Tip>
 
-Interested in adding a new quantization method to Transformers? Refer to the [Contribute new quantization method guide](https://huggingface.co/docs/transformers/main/en/quantization/contribute) to learn more about adding a new quantization method.
+Interested in adding a new quantization method to Diffusers? Refer to the [Contribute new quantization method guide](https://huggingface.co/docs/transformers/main/en/quantization/contribute) to learn more about adding a new quantization method.
 
 </Tip>
 
@@ -32,4 +32,9 @@ If you are new to the quantization field, we recommend you to check out these be
 
 ## When to use what?
 
-This section will be expanded once Diffusers has multiple quantization backends. Currently, we only support `bitsandbytes`. [This resource](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) provides a good overview of the pros and cons of different quantization techniques. 
+Diffusers currently supports the following quantization methods.
+- [BitsandBytes](./bitsandbytes.md)
+- [TorchAO](./torchao.md)
+- [GGUF](./gguf.md)
+
+[This resource](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) provides a good overview of the pros and cons of different quantization techniques.

docs/source/en/quantization/torchao.md

@@ -0,0 +1,92 @@
+<!-- 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. -->
+
+# torchao
+
+[TorchAO](https://github.com/pytorch/ao) is an architecture optimization library for PyTorch. It provides high-performance dtypes, optimization techniques, and kernels for inference and training, featuring composability with native PyTorch features like [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html), FullyShardedDataParallel (FSDP), and more.
+
+Before you begin, make sure you have Pytorch 2.5+ and TorchAO installed.
+
+```bash
+pip install -U torch torchao
+```
+
+
+Quantize a model by passing [`TorchAoConfig`] to [`~ModelMixin.from_pretrained`] (you can also load pre-quantized models). This works for any model in any modality, as long as it supports loading with [Accelerate](https://hf.co/docs/accelerate/index) and contains `torch.nn.Linear` layers.
+
+The example below only quantizes the weights to int8.
+
+```python
+from diffusers import FluxPipeline, FluxTransformer2DModel, TorchAoConfig
+
+model_id = "black-forest-labs/Flux.1-Dev"
+dtype = torch.bfloat16
+
+quantization_config = TorchAoConfig("int8wo")
+transformer = FluxTransformer2DModel.from_pretrained(
+    model_id,
+    subfolder="transformer",
+    quantization_config=quantization_config,
+    torch_dtype=dtype,
+)
+pipe = FluxPipeline.from_pretrained(
+    model_id,
+    transformer=transformer,
+    torch_dtype=dtype,
+)
+pipe.to("cuda")
+
+prompt = "A cat holding a sign that says hello world"
+image = pipe(prompt, num_inference_steps=28, guidance_scale=0.0).images[0]
+image.save("output.png")
+```
+
+TorchAO is fully compatible with [torch.compile](./optimization/torch2.0#torchcompile), setting it apart from other quantization methods. This makes it easy to speed up inference with just one line of code.
+
+```python
+# In the above code, add the following after initializing the transformer
+transformer = torch.compile(transformer, mode="max-autotune", fullgraph=True)
+```
+
+For speed and memory benchmarks on Flux and CogVideoX, please refer to the table [here](https://github.com/huggingface/diffusers/pull/10009#issue-2688781450). You can also find some torchao [benchmarks](https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks) numbers for various hardware.
+
+torchao also supports an automatic quantization API through [autoquant](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md#autoquantization). Autoquantization determines the best quantization strategy applicable to a model by comparing the performance of each technique on chosen input types and shapes. Currently, this can be used directly on the underlying modeling components. Diffusers will also expose an autoquant configuration option in the future.
+
+The `TorchAoConfig` class accepts three parameters:
+- `quant_type`: A string value mentioning one of the quantization types below.
+- `modules_to_not_convert`: A list of module full/partial module names for which quantization should not be performed. For example, to not perform any quantization of the [`FluxTransformer2DModel`]'s first block, one would specify: `modules_to_not_convert=["single_transformer_blocks.0"]`.
+- `kwargs`: A dict of keyword arguments to pass to the underlying quantization method which will be invoked based on `quant_type`.
+
+## Supported quantization types
+
+torchao supports weight-only quantization and weight and dynamic-activation quantization for int8, float3-float8, and uint1-uint7.
+
+Weight-only quantization stores the model weights in a specific low-bit data type but performs computation with a higher-precision data type, like `bfloat16`. This lowers the memory requirements from model weights but retains the memory peaks for activation computation.
+
+Dynamic activation quantization stores the model weights in a low-bit dtype, while also quantizing the activations on-the-fly to save additional memory. This lowers the memory requirements from model weights, while also lowering the memory overhead from activation computations. However, this may come at a quality tradeoff at times, so it is recommended to test different models thoroughly.
+
+The quantization methods supported are as follows:
+
+| **Category** | **Full Function Names** | **Shorthands** |
+|--------------|-------------------------|----------------|
+| **Integer quantization** | `int4_weight_only`, `int8_dynamic_activation_int4_weight`, `int8_weight_only`, `int8_dynamic_activation_int8_weight` | `int4wo`, `int4dq`, `int8wo`, `int8dq` |
+| **Floating point 8-bit quantization** | `float8_weight_only`, `float8_dynamic_activation_float8_weight`, `float8_static_activation_float8_weight` | `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`, `float8_e4m3_tensor`, `float8_e4m3_row` |
+| **Floating point X-bit quantization** | `fpx_weight_only` | `fpX_eAwB` where `X` is the number of bits (1-7), `A` is exponent bits, and `B` is mantissa bits. Constraint: `X == A + B + 1` |
+| **Unsigned Integer quantization** | `uintx_weight_only` | `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo` |
+
+Some quantization methods are aliases (for example, `int8wo` is the commonly used shorthand for `int8_weight_only`). This allows using the quantization methods described in the torchao docs as-is, while also making it convenient to remember their shorthand notations.
+
+Refer to the official torchao documentation for a better understanding of the available quantization methods and the exhaustive list of configuration options available.
+
+## Resources
+
+- [TorchAO Quantization API](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md)
+- [Diffusers-TorchAO examples](https://github.com/sayakpaul/diffusers-torchao)

docs/source/en/training/create_dataset.md

@@ -1,6 +1,6 @@
 # Create a dataset for training
 
-There are many datasets on the [Hub](https://huggingface.co/datasets?task_categories=task_categories:text-to-image&sort=downloads) to train a model on, but if you can't find one you're interested in or want to use your own, you can create a dataset with the 🤗 [Datasets](hf.co/docs/datasets) library. The dataset structure depends on the task you want to train your model on. The most basic dataset structure is a directory of images for tasks like unconditional image generation. Another dataset structure may be a directory of images and a text file containing their corresponding text captions for tasks like text-to-image generation.
+There are many datasets on the [Hub](https://huggingface.co/datasets?task_categories=task_categories:text-to-image&sort=downloads) to train a model on, but if you can't find one you're interested in or want to use your own, you can create a dataset with the 🤗 [Datasets](https://huggingface.co/docs/datasets) library. The dataset structure depends on the task you want to train your model on. The most basic dataset structure is a directory of images for tasks like unconditional image generation. Another dataset structure may be a directory of images and a text file containing their corresponding text captions for tasks like text-to-image generation.
 
 This guide will show you two ways to create a dataset to finetune on:
 
@@ -87,4 +87,4 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image.py \
 
 Now that you've created a dataset, you can plug it into the `train_data_dir` (if your dataset is local) or `dataset_name` (if your dataset is on the Hub) arguments of a training script.
 
-For your next steps, feel free to try and use your dataset to train a model for [unconditional generation](unconditional_training) or [text-to-image generation](text2image)!
+For your next steps, feel free to try and use your dataset to train a model for [unconditional generation](unconditional_training) or [text-to-image generation](text2image)!

docs/source/en/tutorials/basic_training.md

@@ -75,7 +75,7 @@ For convenience, create a `TrainingConfig` class containing the training hyperpa
 
 ...     push_to_hub = True  # whether to upload the saved model to the HF Hub
 ...     hub_model_id = "<your-username>/<my-awesome-model>"  # the name of the repository to create on the HF Hub
-...     hub_private_repo = False
+...     hub_private_repo = None
 ...     overwrite_output_dir = True  # overwrite the old model when re-running the notebook
 ...     seed = 0
 

docs/source/en/tutorials/using_peft_for_inference.md

@@ -56,7 +56,7 @@ image
 
 With the `adapter_name` parameter, it is really easy to use another adapter for inference! Load the [nerijs/pixel-art-xl](https://huggingface.co/nerijs/pixel-art-xl) adapter that has been fine-tuned to generate pixel art images and call it `"pixel"`.
 
-The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method:
+The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~PeftAdapterMixin.set_adapters`] method:
 
 ```python
 pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
@@ -85,7 +85,7 @@ By default, if the most up-to-date versions of PEFT and Transformers are detecte
 
 You can also merge different adapter checkpoints for inference to blend their styles together.
 
-Once again, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
+Once again, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
 
 ```python
 pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
@@ -114,7 +114,7 @@ Impressive! As you can see, the model generated an image that mixed the characte
 > [!TIP]
 > Through its PEFT integration, Diffusers also offers more efficient merging methods which you can learn about in the [Merge LoRAs](../using-diffusers/merge_loras) guide!
 
-To return to only using one adapter, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `"toy"` adapter:
+To return to only using one adapter, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `"toy"` adapter:
 
 ```python
 pipe.set_adapters("toy")
@@ -127,7 +127,7 @@ image = pipe(
 image
 ```
 
-Or to disable all adapters entirely, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.disable_lora`] method to return the base model.
+Or to disable all adapters entirely, use the [`~PeftAdapterMixin.disable_lora`] method to return the base model.
 
 ```python
 pipe.disable_lora()
@@ -140,7 +140,8 @@ image
 ![no-lora](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_20_1.png)
 
 ### Customize adapters strength
-For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`].
+
+For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~PeftAdapterMixin.set_adapters`].
 
 For example, here's how you can turn on the adapter for the `down` parts, but turn it off for the `mid` and `up` parts:
 ```python
@@ -195,7 +196,7 @@ image
 
 ![block-lora-mixed](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mixed.png)
 
-## Manage active adapters
+## Manage adapters
 
 You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.StableDiffusionLoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:
 
@@ -212,3 +213,11 @@ list_adapters_component_wise = pipe.get_list_adapters()
 list_adapters_component_wise
 {"text_encoder": ["toy", "pixel"], "unet": ["toy", "pixel"], "text_encoder_2": ["toy", "pixel"]}
 ```
+
+The [`~PeftAdapterMixin.delete_adapters`] function completely removes an adapter and their LoRA layers from a model.
+
+```py
+pipe.delete_adapters("toy")
+pipe.get_active_adapters()
+["pixel"]
+```

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

@@ -0,0 +1,61 @@
+
+# Create a server
+
+Diffusers' pipelines can be used as an inference engine for a server. It supports concurrent and multithreaded requests to generate images that may be requested by multiple users at the same time.
+
+This guide will show you how to use the [`StableDiffusion3Pipeline`] in a server, but feel free to use any pipeline you want.
+
+
+Start by navigating to the `examples/server` folder and installing all of the dependencies.
+
+```py
+pip install .
+pip install -f requirements.txt
+```
+
+Launch the server with the following command.
+
+```py
+python server.py
+```
+
+The server is accessed at http://localhost:8000. You can curl this model with the following command.
+```
+curl -X POST -H "Content-Type: application/json" --data '{"model": "something", "prompt": "a kitten in front of a fireplace"}' http://localhost:8000/v1/images/generations
+```
+
+If you need to upgrade some dependencies, you can use either [pip-tools](https://github.com/jazzband/pip-tools) or [uv](https://github.com/astral-sh/uv). For example, upgrade the dependencies with `uv` using the following command.
+
+```
+uv pip compile requirements.in -o requirements.txt
+```
+
+
+The server is built with [FastAPI](https://fastapi.tiangolo.com/async/). The endpoint for `v1/images/generations` is shown below.
+```py
+@app.post("/v1/images/generations")
+async def generate_image(image_input: TextToImageInput):
+    try:
+        loop = asyncio.get_event_loop()
+        scheduler = shared_pipeline.pipeline.scheduler.from_config(shared_pipeline.pipeline.scheduler.config)
+        pipeline = StableDiffusion3Pipeline.from_pipe(shared_pipeline.pipeline, scheduler=scheduler)
+        generator = torch.Generator(device="cuda")
+        generator.manual_seed(random.randint(0, 10000000))
+        output = await loop.run_in_executor(None, lambda: pipeline(image_input.prompt, generator = generator))
+        logger.info(f"output: {output}")
+        image_url = save_image(output.images[0])
+        return {"data": [{"url": image_url}]}
+    except Exception as e:
+        if isinstance(e, HTTPException):
+            raise e
+        elif hasattr(e, 'message'):
+            raise HTTPException(status_code=500, detail=e.message + traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e) + traceback.format_exc())
+```
+The `generate_image` function is defined as asynchronous with the [async](https://fastapi.tiangolo.com/async/) keyword so that FastAPI knows that whatever is happening in this function won't necessarily return a result right away. Once it hits some point in the function that it needs to await some other [Task](https://docs.python.org/3/library/asyncio-task.html#asyncio.Task), the main thread goes back to answering other HTTP requests. This is shown in the code below with the [await](https://fastapi.tiangolo.com/async/#async-and-await) keyword.
+```py
+output = await loop.run_in_executor(None, lambda: pipeline(image_input.prompt, generator = generator))
+```
+At this point, the execution of the pipeline function is placed onto a [new thread](https://docs.python.org/3/library/asyncio-eventloop.html#asyncio.loop.run_in_executor), and the main thread performs other things until a result is returned from the `pipeline`.
+
+Another important aspect of this implementation is creating a `pipeline` from `shared_pipeline`. The goal behind this is to avoid loading the underlying model more than once onto the GPU while still allowing for each new request that is running on a separate thread to have its own generator and scheduler. The scheduler, in particular, is not thread-safe, and it will cause errors like: `IndexError: index 21 is out of bounds for dimension 0 with size 21` if you try to use the same scheduler across multiple threads.

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

@@ -134,14 +134,16 @@ The [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method loads L
 - the LoRA weights don't have separate identifiers for the UNet and text encoder
 - the LoRA weights have separate identifiers for the UNet and text encoder
 
-But if you only need to load LoRA weights into the UNet, then you can use the [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method. Let's load the [jbilcke-hf/sdxl-cinematic-1](https://huggingface.co/jbilcke-hf/sdxl-cinematic-1) LoRA:
+To directly load (and save) a LoRA adapter at the *model-level*, use [`~PeftAdapterMixin.load_lora_adapter`], which builds and prepares the necessary model configuration for the adapter. Like [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`PeftAdapterMixin.load_lora_adapter`] can load LoRAs for both the UNet and text encoder. For example, if you're loading a LoRA for the UNet, [`PeftAdapterMixin.load_lora_adapter`] ignores the keys for the text encoder.
+
+Use the `weight_name` parameter to specify the specific weight file and the `prefix` parameter to filter for the appropriate state dicts (`"unet"` in this case) to load.
 
 ```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")
+pipeline.unet.load_lora_adapter("jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", prefix="unet")
 
 # use cnmt in the prompt to trigger the LoRA
 prompt = "A cute cnmt eating a slice of pizza, stunning color scheme, masterpiece, illustration"
@@ -153,6 +155,8 @@ image
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
 </div>
 
+Save an adapter with [`~PeftAdapterMixin.save_lora_adapter`].
+
 To unload the LoRA weights, use the [`~loaders.StableDiffusionLoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
 
 ```py

docs/source/ko/api/pipelines/stable_diffusion/stable_diffusion_xl.md

@@ -121,7 +121,7 @@ image = pipe(prompt=prompt, image=init_image, mask_image=mask_image, num_inferen
 
 ### 이미지 결과물을 정제하기
 
-[base 모델 체크포인트](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)에서, StableDiffusion-XL 또한 고주파 품질을 향상시키는 이미지를 생성하기 위해 낮은 노이즈 단계 이미지를 제거하는데 특화된 [refiner 체크포인트](huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 포함하고 있습니다. 이 refiner 체크포인트는 이미지 품질을 향상시키기 위해 base 체크포인트를 실행한 후 "두 번째 단계" 파이프라인에 사용될 수 있습니다.
+[base 모델 체크포인트](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)에서, StableDiffusion-XL 또한 고주파 품질을 향상시키는 이미지를 생성하기 위해 낮은 노이즈 단계 이미지를 제거하는데 특화된 [refiner 체크포인트](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 포함하고 있습니다. 이 refiner 체크포인트는 이미지 품질을 향상시키기 위해 base 체크포인트를 실행한 후 "두 번째 단계" 파이프라인에 사용될 수 있습니다.
 
 refiner를 사용할 때, 쉽게 사용할 수 있습니다
 - 1.) base 모델과 refiner을 사용하는데, 이는 *Denoisers의 앙상블*을 위한 첫 번째 제안된 [eDiff-I](https://research.nvidia.com/labs/dir/eDiff-I/)를 사용하거나
@@ -215,7 +215,7 @@ image = refiner(
 
 #### 2.) 노이즈가 완전히 제거된 기본 이미지에서 이미지 출력을 정제하기
 
-일반적인 [`StableDiffusionImg2ImgPipeline`] 방식에서, 기본 모델에서 생성된 완전히 노이즈가 제거된 이미지는 [refiner checkpoint](huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 사용해 더 향상시킬 수 있습니다.
+일반적인 [`StableDiffusionImg2ImgPipeline`] 방식에서, 기본 모델에서 생성된 완전히 노이즈가 제거된 이미지는 [refiner checkpoint](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 사용해 더 향상시킬 수 있습니다.
 
 이를 위해, 보통의 "base" text-to-image 파이프라인을 수행 후에 image-to-image 파이프라인으로써 refiner를 실행시킬 수 있습니다. base 모델의 출력을 잠재 공간에 남겨둘 수 있습니다.
 

docs/source/ko/training/create_dataset.md

@@ -1,7 +1,7 @@
 # 학습을 위한 데이터셋 만들기
 
 [Hub](https://huggingface.co/datasets?task_categories=task_categories:text-to-image&sort=downloads) 에는 모델 교육을 위한 많은 데이터셋이 있지만,
-관심이 있거나 사용하고 싶은 데이터셋을 찾을 수 없는 경우 🤗 [Datasets](hf.co/docs/datasets) 라이브러리를 사용하여 데이터셋을 만들 수 있습니다.
+관심이 있거나 사용하고 싶은 데이터셋을 찾을 수 없는 경우 🤗 [Datasets](https://huggingface.co/docs/datasets) 라이브러리를 사용하여 데이터셋을 만들 수 있습니다.
 데이터셋 구조는 모델을 학습하려는 작업에 따라 달라집니다.
 가장 기본적인 데이터셋 구조는 unconditional 이미지 생성과 같은 작업을 위한 이미지 디렉토리입니다.
 또 다른 데이터셋 구조는 이미지 디렉토리와 text-to-image 생성과 같은 작업에 해당하는 텍스트 캡션이 포함된 텍스트 파일일 수 있습니다.

docs/source/ko/training/lora.md

@@ -36,7 +36,7 @@ specific language governing permissions and limitations under the License.
 
 [cloneofsimo](https://github.com/cloneofsimo)는 인기 있는 [lora](https://github.com/cloneofsimo/lora) GitHub 리포지토리에서 Stable Diffusion을 위한 LoRA 학습을 최초로 시도했습니다. 🧨 Diffusers는 [text-to-image 생성](https://github.com/huggingface/diffusers/tree/main/examples/text_to_image#training-with-lora) 및 [DreamBooth](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth#training-with-low-rank-adaptation-of-large-language-models-lora)을 지원합니다. 이 가이드는 두 가지를 모두 수행하는 방법을 보여줍니다.
 
-모델을 저장하거나 커뮤니티와 공유하려면 Hugging Face 계정에 로그인하세요(아직 계정이 없는 경우 [생성](hf.co/join)하세요):
+모델을 저장하거나 커뮤니티와 공유하려면 Hugging Face 계정에 로그인하세요(아직 계정이 없는 경우 [생성](https://huggingface.co/join)하세요):
 
 ```bash
 huggingface-cli login

docs/source/ko/tutorials/basic_training.md

@@ -76,7 +76,7 @@ huggingface-cli login
 ...     output_dir = "ddpm-butterflies-128"  # 로컬 및 HF Hub에 저장되는 모델명
 
 ...     push_to_hub = True  # 저장된 모델을 HF Hub에 업로드할지 여부
-...     hub_private_repo = False
+...     hub_private_repo = None
 ...     overwrite_output_dir = True  # 노트북을 다시 실행할 때 이전 모델에 덮어씌울지
 ...     seed = 0
 

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -2154,6 +2154,7 @@ def main(args):
 
                 # encode batch prompts when custom prompts are provided for each image -
                 if train_dataset.custom_instance_prompts:
+                    elems_to_repeat = 1
                     if freeze_text_encoder:
                         prompt_embeds, pooled_prompt_embeds, text_ids = compute_text_embeddings(
                             prompts, text_encoders, tokenizers
@@ -2168,17 +2169,21 @@ def main(args):
                             max_sequence_length=args.max_sequence_length,
                             add_special_tokens=add_special_tokens_t5,
                         )
+                else:
+                    elems_to_repeat = len(prompts)
 
                 if not freeze_text_encoder:
                     prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
                         text_encoders=[text_encoder_one, text_encoder_two],
                         tokenizers=[None, None],
-                        text_input_ids_list=[tokens_one, tokens_two],
+                        text_input_ids_list=[
+                            tokens_one.repeat(elems_to_repeat, 1),
+                            tokens_two.repeat(elems_to_repeat, 1),
+                        ],
                         max_sequence_length=args.max_sequence_length,
                         device=accelerator.device,
                         prompt=prompts,
                     )
-
                 # Convert images to latent space
                 if args.cache_latents:
                     model_input = latents_cache[step].sample()
@@ -2371,6 +2376,9 @@ def main(args):
                     epoch=epoch,
                     torch_dtype=weight_dtype,
                 )
+                images = None
+                del pipeline
+
                 if freeze_text_encoder:
                     del text_encoder_one, text_encoder_two
                     free_memory()
@@ -2448,6 +2456,8 @@ def main(args):
                 commit_message="End of training",
                 ignore_patterns=["step_*", "epoch_*"],
             )
+        images = None
+        del pipeline
 
     accelerator.end_training()
 

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -872,10 +872,9 @@ def prepare_rotary_positional_embeddings(
         crops_coords=grid_crops_coords,
         grid_size=(grid_height, grid_width),
         temporal_size=num_frames,
+        device=device,
     )
 
-    freqs_cos = freqs_cos.to(device=device)
-    freqs_sin = freqs_sin.to(device=device)
     return freqs_cos, freqs_sin
 
 

examples/cogvideo/train_cogvideox_lora.py

@@ -894,10 +894,9 @@ def prepare_rotary_positional_embeddings(
         crops_coords=grid_crops_coords,
         grid_size=(grid_height, grid_width),
         temporal_size=num_frames,
+        device=device,
     )
 
-    freqs_cos = freqs_cos.to(device=device)
-    freqs_sin = freqs_sin.to(device=device)
     return freqs_cos, freqs_sin
 
 

examples/community/README.md

@@ -11,22 +11,22 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
 |Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
-|Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|NA|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
+|Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
 |Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
 | HD-Painter                                                                                                                            | [HD-Painter](https://github.com/Picsart-AI-Research/HD-Painter) enables prompt-faithfull and high resolution (up to 2k) image inpainting upon any diffusion-based image inpainting method.                                                                                                                                                                                                                                                                                                               | [HD-Painter](#hd-painter)                                                                 | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/PAIR/HD-Painter)                                                                              | [Manukyan Hayk](https://github.com/haikmanukyan) and [Sargsyan Andranik](https://github.com/AndranikSargsyan) |
 | Marigold Monocular Depth Estimation                                                                                                   | A universal monocular depth estimator, utilizing Stable Diffusion, delivering sharp predictions in the wild. (See the [project page](https://marigoldmonodepth.github.io) and [full codebase](https://github.com/prs-eth/marigold) for more details.)                                                                                                                                                                                                                                                        | [Marigold Depth Estimation](#marigold-depth-estimation)                                   | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/toshas/marigold) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/12G8reD13DdpMie5ZQlaFNo2WCGeNUH-u?usp=sharing) | [Bingxin Ke](https://github.com/markkua) and [Anton Obukhov](https://github.com/toshas) |
 | LLM-grounded Diffusion (LMD+)                                                                                                         | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion)                                                                                                                                                                                                                                                                                                                                                                                                                                   | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion)                             | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) |                [Long (Tony) Lian](https://tonylian.com/) |
 | CLIP Guided Stable Diffusion                                                                                                          | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)                             | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |                [Suraj Patil](https://github.com/patil-suraj/) |
 | One Step U-Net (Dummy)                                                                                                                | Example showcasing of how to use Community Pipelines (see <https://github.com/huggingface/diffusers/issues/841>)                                                                                                                                                                                                                                                                                                                                                                                           | [One Step U-Net](#one-step-unet)                                                          | -                                                                                                                                                                                                                  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
-| Stable Diffusion Interpolation                                                                                                        | Interpolate the latent space of Stable Diffusion between different prompts/seeds                                                                                                                                                                                                                                                                                                                                                                                                                         | [Stable Diffusion Interpolation](#stable-diffusion-interpolation)                         | -                                                                                                                                                                                                                  |                       [Nate Raw](https://github.com/nateraw/) |
-| Stable Diffusion Mega                                                                                                                 | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega)                                           | -                                                                                                                                                                                                                  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
-| Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | -                                                                                                                                                                                                                  |                           [SkyTNT](https://github.com/SkyTNT) |
-| Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       | -                                                                                                                                                                                                                  |             [Mikail Duzenli](https://github.com/MikailINTech)
-| Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | -                                                                                                                                                                                                                  |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
+| Stable Diffusion Interpolation                                                                                                        | Interpolate the latent space of Stable Diffusion between different prompts/seeds                                                                                                                                                                                                                                                                                                                                                                                                                         | [Stable Diffusion Interpolation](#stable-diffusion-interpolation)                         | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_interpolation.ipynb)                                                                                                                                                           |                       [Nate Raw](https://github.com/nateraw/) |
+| Stable Diffusion Mega                                                                                                                 | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega)                                           | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_mega.ipynb)                                                                                                                                                                             |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
+| Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/long_prompt_weighting_stable_diffusion.ipynb)                                                                                        |                           [SkyTNT](https://github.com/SkyTNT) |
+| Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/speech_to_image.ipynb)                                                                                                                                                                                                   |             [Mikail Duzenli](https://github.com/MikailINTech)
+| Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/wildcard_stable_diffusion.ipynb)                                                                                                                                                                                 |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
 | [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "&#124;" in prompts (as an AND condition) and weights (separated by "&#124;" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion                                                                                                        | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                                                                                                       | [Seed Resizing](#seed-resizing)                                                           | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Imagic Stable Diffusion                                                                                                               | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image                                                                                                                                                                                                                                                                                                                                                                                                                   | [Imagic Stable Diffusion](#imagic-stable-diffusion)                                       | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
-| Multilingual Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                                                                                                                                               | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                  | -                                                                                                                                                                                                                  |          [Juan Carlos Piñeros](https://github.com/juancopi81) |
+| Multilingual Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                                                                                                                                               | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/multilingual_stable_diffusion.ipynb)                                                                                                                                                                             |          [Juan Carlos Piñeros](https://github.com/juancopi81) |
 | GlueGen Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter.                                                                                                                                                                                                                                                                                                                                                                                                                               | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline)                  | -                                                                                                                                                                                                                  |          [Phạm Hồng Vinh](https://github.com/rootonchair) |
 | Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | -                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
 | Text Based Inpainting Stable Diffusion                                                                                                | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting                                                                                                                                                                                                                                                                                                                                                                                              | [Text Based Inpainting Stable Diffusion](#text-based-inpainting-stable-diffusion)     | -                                                                                                                                                                                                                  |                   [Dhruv Karan](https://github.com/unography) |
@@ -41,8 +41,8 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | DDIM Noise Comparative Analysis Pipeline                                                                                              | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227))                                                                                                                                                                                                                                                                                                                             | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline)     | - |              [Aengus (Duc-Anh)](https://github.com/aengusng8) |
 | CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | - |               [Nipun Jindal](https://github.com/nipunjindal/) |
 | TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
-| EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | - |                    [Joqsan Azocar](https://github.com/Joqsan) |
-| Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )                                    | - |                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
+| EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/edict_image_pipeline.ipynb) |                    [Joqsan Azocar](https://github.com/Joqsan) |
+| Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_repaint.ipynb)|                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
 | TensorRT Stable Diffusion Image to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
 | Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) |
 | CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
@@ -61,13 +61,13 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 |   Regional Prompting Pipeline                                                                                               | Assign multiple prompts for different regions                                                                                                                                                                                                                                                                                                                                                    |  [Regional Prompting Pipeline](#regional-prompting-pipeline) | - | [hako-mikan](https://github.com/hako-mikan) |
 | LDM3D-sr (LDM3D upscaler)                                                                                                             | Upscale low resolution RGB and depth inputs to high resolution                                                                                                                                                                                                                                                                                                                                                                                                                              | [StableDiffusionUpscaleLDM3D Pipeline](https://github.com/estelleafl/diffusers/tree/ldm3d_upscaler_community/examples/community#stablediffusionupscaleldm3d-pipeline)                                                                             | -                                                                                                                                                                                                             |                                                        [Estelle Aflalo](https://github.com/estelleafl) |
 | AnimateDiff ControlNet Pipeline                                                                                                    | Combines AnimateDiff with precise motion control using ControlNets                                                                                                                                                                                                                                                                                                                                                                                                                                    | [AnimateDiff ControlNet Pipeline](#animatediff-controlnet-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SKboYeGjEQmQPWoFC0aLYpBlYdHXkvAu?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) and [Edoardo Botta](https://github.com/EdoardoBotta) |
-|   DemoFusion Pipeline                                                                                                    | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [DemoFusion Pipeline](#demofusion)      | - |              [Ruoyi Du](https://github.com/RuoyiDu) |
-|   Instaflow Pipeline                                                                                                    | Implementation of [InstaFlow! One-Step Stable Diffusion with Rectified Flow](https://arxiv.org/abs/2309.06380)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Instaflow Pipeline](#instaflow-pipeline)      | - |              [Ayush Mangal](https://github.com/ayushtues) |
+|   DemoFusion Pipeline                                                                                                    | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [DemoFusion Pipeline](#demofusion)      | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/demo_fusion.ipynb) |              [Ruoyi Du](https://github.com/RuoyiDu) |
+|   Instaflow Pipeline                                                                                                    | Implementation of [InstaFlow! One-Step Stable Diffusion with Rectified Flow](https://arxiv.org/abs/2309.06380)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Instaflow Pipeline](#instaflow-pipeline)      | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/insta_flow.ipynb) |              [Ayush Mangal](https://github.com/ayushtues) |
 |   Null-Text Inversion Pipeline  | Implement [Null-text Inversion for Editing Real Images using Guided Diffusion Models](https://arxiv.org/abs/2211.09794) as a pipeline.                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Null-Text Inversion](https://github.com/google/prompt-to-prompt/)      | - |              [Junsheng Luan](https://github.com/Junsheng121) |
 |   Rerender A Video Pipeline                                                                                                    | Implementation of [[SIGGRAPH Asia 2023] Rerender A Video: Zero-Shot Text-Guided Video-to-Video Translation](https://arxiv.org/abs/2306.07954)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Rerender A Video Pipeline](#rerender-a-video)      | - |              [Yifan Zhou](https://github.com/SingleZombie) |
 | StyleAligned Pipeline                                                                                                    | Implementation of [Style Aligned Image Generation via Shared Attention](https://arxiv.org/abs/2312.02133)                                                                                                                                                                                                                                                                                                                                                                                                                                   | [StyleAligned Pipeline](#stylealigned-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://drive.google.com/file/d/15X2E0jFPTajUIjS0FzX50OaHsCbP2lQ0/view?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
 | AnimateDiff Image-To-Video Pipeline | Experimental Image-To-Video support for AnimateDiff (open to improvements) | [AnimateDiff Image To Video Pipeline](#animatediff-image-to-video-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://drive.google.com/file/d/1TvzCDPHhfFtdcJZe4RLloAwyoLKuttWK/view?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
-|   IP Adapter FaceID Stable Diffusion                                                                                               | Stable Diffusion Pipeline that supports IP Adapter Face ID                                                                                                                                                                                                                                                                                                                                                  |  [IP Adapter Face ID](#ip-adapter-face-id) | - | [Fabio Rigano](https://github.com/fabiorigano) |
+|   IP Adapter FaceID Stable Diffusion                                                                                               | Stable Diffusion Pipeline that supports IP Adapter Face ID                                                                                                                                                                                                                                                                                                                                                  |  [IP Adapter Face ID](#ip-adapter-face-id) |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/ip_adapter_face_id.ipynb)| [Fabio Rigano](https://github.com/fabiorigano) |
 |   InstantID Pipeline                                                                                               | Stable Diffusion XL Pipeline that supports InstantID                                                                                                                                                                                                                                                                                                                                                 |  [InstantID Pipeline](#instantid-pipeline) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/InstantX/InstantID) | [Haofan Wang](https://github.com/haofanwang) |
 |   UFOGen Scheduler                                                                                               | Scheduler for UFOGen Model (compatible with Stable Diffusion pipelines)                                                                                                                                                                                                                                                                                                                                                 |  [UFOGen Scheduler](#ufogen-scheduler) | - | [dg845](https://github.com/dg845) |
 | 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/) |
@@ -251,24 +251,30 @@ Example usage:
 from diffusers import DiffusionPipeline
 import torch
 
+model_name = "black-forest-labs/FLUX.1-dev"
+prompt = "a watercolor painting of a unicorn"
+negative_prompt = "pink"
+
+# Load the diffusion pipeline
 pipeline = DiffusionPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-dev",
+    model_name,
     torch_dtype=torch.bfloat16,
     custom_pipeline="pipeline_flux_with_cfg"
 )
 pipeline.enable_model_cpu_offload()
-prompt = "a watercolor painting of a unicorn"
-negative_prompt = "pink"
 
+# Generate the image
 img = pipeline(
     prompt=prompt,
     negative_prompt=negative_prompt,
     true_cfg=1.5,
     guidance_scale=3.5,
-    num_images_per_prompt=1,
     generator=torch.manual_seed(0)
 ).images[0]
+
+# Save the generated image
 img.save("cfg_flux.png")
+print("Image generated and saved successfully.")
 ```
 
 ### Differential Diffusion
@@ -841,6 +847,8 @@ out = pipe(
     wildcard_files=["object.txt", "animal.txt"],
     num_prompt_samples=1
 )
+out.images[0].save("image.png")
+torch.cuda.empty_cache()
 ```
 
 ### Composable Stable diffusion
@@ -2617,16 +2625,17 @@ for obj in range(bs):
 
 ### Stable Diffusion XL Reference
 
-This pipeline uses the Reference. Refer to the [stable_diffusion_reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference).
+This pipeline uses the Reference. Refer to the [Stable Diffusion Reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference) section for more information.
 
 ```py
 import torch
-from PIL import Image
+# from diffusers import DiffusionPipeline
 from diffusers.utils import load_image
-from diffusers import DiffusionPipeline
 from diffusers.schedulers import UniPCMultistepScheduler
 
-input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
+from .stable_diffusion_xl_reference import StableDiffusionXLReferencePipeline
+
+input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_input_cat.jpg")
 
 # pipe = DiffusionPipeline.from_pretrained(
 #     "stabilityai/stable-diffusion-xl-base-1.0",
@@ -2644,7 +2653,7 @@ pipe = StableDiffusionXLReferencePipeline.from_pretrained(
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
 
 result_img = pipe(ref_image=input_image,
-      prompt="1girl",
+      prompt="a dog",
       num_inference_steps=20,
       reference_attn=True,
       reference_adain=True).images[0]
@@ -2652,14 +2661,14 @@ result_img = pipe(ref_image=input_image,
 
 Reference Image
 
-![reference_image](https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png)
+![reference_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_input_cat.jpg)
 
 Output Image
 
-`prompt: 1 girl`
+`prompt: a dog`
 
-`reference_attn=True, reference_adain=True, num_inference_steps=20`
-![Output_image](https://github.com/zideliu/diffusers/assets/34944964/743848da-a215-48f9-ae39-b5e2ae49fb13)
+`reference_attn=False, reference_adain=True, num_inference_steps=20`
+![Output_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_adain_dog.png)
 
 Reference Image
 ![reference_image](https://github.com/huggingface/diffusers/assets/34944964/449bdab6-e744-4fb2-9620-d4068d9a741b)
@@ -2681,6 +2690,88 @@ Output Image
 `reference_attn=True, reference_adain=True, num_inference_steps=20`
 ![output_image](https://github.com/huggingface/diffusers/assets/34944964/9b2f1aca-886f-49c3-89ec-d2031c8e3670)
 
+### Stable Diffusion XL ControlNet Reference
+
+This pipeline uses the Reference Control and with ControlNet. Refer to the [Stable Diffusion ControlNet Reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-controlnet-reference) and [Stable Diffusion XL Reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-xl-reference) sections for more information.
+
+```py
+from diffusers import ControlNetModel, AutoencoderKL
+from diffusers.schedulers import UniPCMultistepScheduler
+from diffusers.utils import load_image
+import numpy as np
+import torch
+
+import cv2
+from PIL import Image
+
+from .stable_diffusion_xl_controlnet_reference import StableDiffusionXLControlNetReferencePipeline
+
+# download an image
+canny_image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_input_cat.jpg"
+)
+
+ref_image = load_image(
+    "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png"
+)
+
+# initialize the models and pipeline
+controlnet_conditioning_scale = 0.5  # recommended for good generalization
+controlnet = ControlNetModel.from_pretrained(
+    "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16
+)
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+pipe = StableDiffusionXLControlNetReferencePipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16
+).to("cuda:0")
+
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+# get canny image
+image = np.array(canny_image)
+image = cv2.Canny(image, 100, 200)
+image = image[:, :, None]
+image = np.concatenate([image, image, image], axis=2)
+canny_image = Image.fromarray(image)
+
+# generate image
+image = pipe(
+    prompt="a cat",
+    num_inference_steps=20,
+    controlnet_conditioning_scale=controlnet_conditioning_scale,
+    image=canny_image,
+    ref_image=ref_image,
+    reference_attn=False,
+    reference_adain=True,
+    style_fidelity=1.0,
+    generator=torch.Generator("cuda").manual_seed(42)
+).images[0]
+```
+
+Canny ControlNet Image
+
+![canny_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_input_cat.jpg)
+
+Reference Image
+
+![ref_image](https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png)
+
+Output Image
+
+`prompt: a cat`
+
+`reference_attn=True, reference_adain=True, num_inference_steps=20, style_fidelity=1.0`
+
+![Output_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_attn_adain_canny_cat.png)
+
+`reference_attn=False, reference_adain=True, num_inference_steps=20, style_fidelity=1.0`
+
+![Output_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_adain_canny_cat.png)
+
+`reference_attn=True, reference_adain=False, num_inference_steps=20, style_fidelity=1.0`
+
+![Output_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_attn_canny_cat.png)
+
 ### Stable diffusion fabric pipeline
 
 FABRIC approach applicable to a wide range of popular diffusion models, which exploits
@@ -3376,6 +3467,20 @@ best quality, 3persons in garden, a boy blue shirt BREAK
 best quality, 3persons in garden, an old man red suit
 ```
 
+### Use base prompt
+
+You can use a base prompt to apply the prompt to all areas. You can set a base prompt by adding `ADDBASE` at the end. Base prompts can also be combined with common prompts, but the base prompt must be specified first.
+
+```
+2d animation style ADDBASE
+masterpiece, high quality ADDCOMM
+(blue sky)++ BREAK
+green hair twintail BREAK
+book shelf BREAK
+messy desk BREAK
+orange++ dress and sofa
+```
+
 ### Negative prompt
 
 Negative prompts are equally effective across all regions, but it is possible to set region-specific prompts for negative prompts as well. The number of BREAKs must be the same as the number of prompts. If the number of prompts does not match, the negative prompts will be used without being divided into regions.
@@ -3406,6 +3511,7 @@ pipe(prompt=prompt, rp_args=rp_args)
 ### Optional Parameters
 
 - `save_mask`: In `Prompt` mode, choose whether to output the generated mask along with the image. The default is `False`.
+- `base_ratio`: Used with `ADDBASE`. Sets the ratio of the base prompt; if base ratio is set to 0.2, then resulting images will consist of `20%*BASE_PROMPT + 80%*REGION_PROMPT`
 
 The Pipeline supports `compel` syntax. Input prompts using the `compel` structure will be automatically applied and processed.
 
@@ -3734,6 +3840,7 @@ The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
 
 ```py
 from diffusers import DiffusionPipeline
+import torch
 
 pipe = DiffusionPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-base-1.0",
@@ -3857,9 +3964,10 @@ You can also combine it with LORA out of the box, like <https://huggingface.co/a
 from diffusers import DiffusionPipeline
 import torch
 
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 pipe = DiffusionPipeline.from_pretrained("XCLIU/instaflow_0_9B_from_sd_1_5", torch_dtype=torch.float16, custom_pipeline="instaflow_one_step")
-pipe.to("cuda")  ### if GPU is not available, comment this line
+pipe.to(device)  ### if GPU is not available, comment this line
 pipe.load_lora_weights("artificialguybr/logo-redmond-1-5v-logo-lora-for-liberteredmond-sd-1-5")
 prompt = "logo, A logo for a fitness app, dynamic running figure, energetic colors (red, orange) ),LogoRedAF ,"
 images = pipe(prompt=prompt,
@@ -4692,4 +4800,4 @@ with torch.no_grad():
 ```
 
 In the folder examples/pixart there is also a script that can be used to train new models.
-Please check the script `train_controlnet_hf_diffusers.sh` on how to start the training.
+Please check the script `train_controlnet_hf_diffusers.sh` on how to start the training.

examples/community/README_community_scripts.md

@@ -6,9 +6,9 @@ If a community script doesn't work as expected, please open an issue and ping th
 
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
-| Using IP-Adapter with Negative Noise                                                                                                  | Using negative noise with IP-adapter to better control the generation (see the [original post](https://github.com/huggingface/diffusers/discussions/7167) on the forum for more details)                                                                                                                                                                                                                                                    | [IP-Adapter Negative Noise](#ip-adapter-negative-noise)                                   | https://github.com/huggingface/notebooks/blob/main/diffusers/ip_adapter_negative_noise.ipynb | [Álvaro Somoza](https://github.com/asomoza)|
-| Asymmetric Tiling                                                                                                  |configure seamless image tiling independently for the X and Y axes                                                                                                                                                                                                      | [Asymmetric Tiling](#Asymmetric-Tiling )                                   |https://github.com/huggingface/notebooks/blob/main/diffusers/asymetric_tiling.ipynb | [alexisrolland](https://github.com/alexisrolland)|
-| Prompt Scheduling Callback                                                                                                  |Allows changing prompts during a generation                                                                                                                                                                                                      | [Prompt Scheduling-Callback](#Prompt-Scheduling-Callback )                                   |https://github.com/huggingface/notebooks/blob/main/diffusers/prompt_scheduling_callback.ipynb | [hlky](https://github.com/hlky)|
+| Using IP-Adapter with Negative Noise                                                                                                  | Using negative noise with IP-adapter to better control the generation (see the [original post](https://github.com/huggingface/diffusers/discussions/7167) on the forum for more details)                                                                                                                                                                                                                                                    | [IP-Adapter Negative Noise](#ip-adapter-negative-noise)                                   |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/ip_adapter_negative_noise.ipynb) | [Álvaro Somoza](https://github.com/asomoza)|
+| Asymmetric Tiling                                                                                                  |configure seamless image tiling independently for the X and Y axes                                                                                                                                                                                                      | [Asymmetric Tiling](#Asymmetric-Tiling )                                   |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/asymetric_tiling.ipynb) | [alexisrolland](https://github.com/alexisrolland)|
+| Prompt Scheduling Callback                                                                                                  |Allows changing prompts during a generation                                                                                                                                                                                                      | [Prompt Scheduling-Callback](#Prompt-Scheduling-Callback )                                   |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/prompt_scheduling_callback.ipynb) | [hlky](https://github.com/hlky)|
 
 
 ## Example usages
@@ -241,27 +241,15 @@ from diffusers import StableDiffusionPipeline
 from diffusers.callbacks import PipelineCallback, MultiPipelineCallbacks
 from diffusers.configuration_utils import register_to_config
 import torch
-from typing import Any, Dict, Optional
+from typing import Any, Dict, Tuple, Union
 
 
-pipeline: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
-    "stable-diffusion-v1-5/stable-diffusion-v1-5",
-    torch_dtype=torch.float16,
-    variant="fp16",
-    use_safetensors=True,
-).to("cuda")
-pipeline.safety_checker = None
-pipeline.requires_safety_checker = False
-
-
-class SDPromptScheduleCallback(PipelineCallback):
+class SDPromptSchedulingCallback(PipelineCallback):
     @register_to_config
     def __init__(
         self,
-        prompt: str,
-        negative_prompt: Optional[str] = None,
-        num_images_per_prompt: int = 1,
-        cutoff_step_ratio=1.0,
+        encoded_prompt: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        cutoff_step_ratio=None,
         cutoff_step_index=None,
     ):
         super().__init__(
@@ -275,6 +263,10 @@ class SDPromptScheduleCallback(PipelineCallback):
     ) -> Dict[str, Any]:
         cutoff_step_ratio = self.config.cutoff_step_ratio
         cutoff_step_index = self.config.cutoff_step_index
+        if isinstance(self.config.encoded_prompt, tuple):
+            prompt_embeds, negative_prompt_embeds = self.config.encoded_prompt
+        else:
+            prompt_embeds = self.config.encoded_prompt
 
         # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
         cutoff_step = (
@@ -284,34 +276,164 @@ class SDPromptScheduleCallback(PipelineCallback):
         )
 
         if step_index == cutoff_step:
-            prompt_embeds, negative_prompt_embeds = pipeline.encode_prompt(
-                prompt=self.config.prompt,
-                negative_prompt=self.config.negative_prompt,
-                device=pipeline._execution_device,
-                num_images_per_prompt=self.config.num_images_per_prompt,
-                do_classifier_free_guidance=pipeline.do_classifier_free_guidance,
-            )
             if pipeline.do_classifier_free_guidance:
                 prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
             callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
         return callback_kwargs
 
+
+pipeline: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+pipeline.safety_checker = None
+pipeline.requires_safety_checker = False
+
 callback = MultiPipelineCallbacks(
     [
-        SDPromptScheduleCallback(
-            prompt="Official portrait of a smiling world war ii general, female, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
-            negative_prompt="Deformed, ugly, bad anatomy",
-            cutoff_step_ratio=0.25,
-        )
+        SDPromptSchedulingCallback(
+            encoded_prompt=pipeline.encode_prompt(
+                prompt=f"prompt {index}",
+                negative_prompt=f"negative prompt {index}",
+                device=pipeline._execution_device,
+                num_images_per_prompt=1,
+                # pipeline.do_classifier_free_guidance can't be accessed until after pipeline is ran
+                do_classifier_free_guidance=True,
+            ),
+            cutoff_step_index=index,
+        ) for index in range(1, 20)
     ]
 )
 
 image = pipeline(
-    prompt="Official portrait of a smiling world war ii general, male, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
-    negative_prompt="Deformed, ugly, bad anatomy",
+    prompt="prompt"
+    negative_prompt="negative prompt",
     callback_on_step_end=callback,
     callback_on_step_end_tensor_inputs=["prompt_embeds"],
 ).images[0]
 torch.cuda.empty_cache()
 image.save('image.png')
 ```
+
+```python
+from diffusers import StableDiffusionXLPipeline
+from diffusers.callbacks import PipelineCallback, MultiPipelineCallbacks
+from diffusers.configuration_utils import register_to_config
+import torch
+from typing import Any, Dict, Tuple, Union
+
+
+class SDXLPromptSchedulingCallback(PipelineCallback):
+    @register_to_config
+    def __init__(
+        self,
+        encoded_prompt: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        add_text_embeds: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        add_time_ids: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        cutoff_step_ratio=None,
+        cutoff_step_index=None,
+    ):
+        super().__init__(
+            cutoff_step_ratio=cutoff_step_ratio, cutoff_step_index=cutoff_step_index
+        )
+
+    tensor_inputs = ["prompt_embeds", "add_text_embeds", "add_time_ids"]
+
+    def callback_fn(
+        self, pipeline, step_index, timestep, callback_kwargs
+    ) -> Dict[str, Any]:
+        cutoff_step_ratio = self.config.cutoff_step_ratio
+        cutoff_step_index = self.config.cutoff_step_index
+        if isinstance(self.config.encoded_prompt, tuple):
+            prompt_embeds, negative_prompt_embeds = self.config.encoded_prompt
+        else:
+            prompt_embeds = self.config.encoded_prompt
+        if isinstance(self.config.add_text_embeds, tuple):
+            add_text_embeds, negative_add_text_embeds = self.config.add_text_embeds
+        else:
+            add_text_embeds = self.config.add_text_embeds
+        if isinstance(self.config.add_time_ids, tuple):
+            add_time_ids, negative_add_time_ids = self.config.add_time_ids
+        else:
+            add_time_ids = self.config.add_time_ids
+
+        # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
+        cutoff_step = (
+            cutoff_step_index
+            if cutoff_step_index is not None
+            else int(pipeline.num_timesteps * cutoff_step_ratio)
+        )
+
+        if step_index == cutoff_step:
+            if pipeline.do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+                add_text_embeds = torch.cat([negative_add_text_embeds, add_text_embeds])
+                add_time_ids = torch.cat([negative_add_time_ids, add_time_ids])
+            callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
+            callback_kwargs[self.tensor_inputs[1]] = add_text_embeds
+            callback_kwargs[self.tensor_inputs[2]] = add_time_ids
+        return callback_kwargs
+
+
+pipeline: StableDiffusionXLPipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+
+callbacks = []
+for index in range(1, 20):
+    (
+        prompt_embeds,
+        negative_prompt_embeds,
+        pooled_prompt_embeds,
+        negative_pooled_prompt_embeds,
+    ) = pipeline.encode_prompt(
+        prompt=f"prompt {index}",
+        negative_prompt=f"prompt {index}",
+        device=pipeline._execution_device,
+        num_images_per_prompt=1,
+        # pipeline.do_classifier_free_guidance can't be accessed until after pipeline is ran
+        do_classifier_free_guidance=True,
+    )
+    text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+    add_time_ids = pipeline._get_add_time_ids(
+        (1024, 1024),
+        (0, 0),
+        (1024, 1024),
+        dtype=prompt_embeds.dtype,
+        text_encoder_projection_dim=text_encoder_projection_dim,
+    )
+    negative_add_time_ids = pipeline._get_add_time_ids(
+        (1024, 1024),
+        (0, 0),
+        (1024, 1024),
+        dtype=prompt_embeds.dtype,
+        text_encoder_projection_dim=text_encoder_projection_dim,
+    )
+    callbacks.append(
+        SDXLPromptSchedulingCallback(
+            encoded_prompt=(prompt_embeds, negative_prompt_embeds),
+            add_text_embeds=(pooled_prompt_embeds, negative_pooled_prompt_embeds),
+            add_time_ids=(add_time_ids, negative_add_time_ids),
+            cutoff_step_index=index,
+        )
+    )
+
+
+callback = MultiPipelineCallbacks(callbacks)
+
+image = pipeline(
+    prompt="prompt",
+    negative_prompt="negative prompt",
+    callback_on_step_end=callback,
+    callback_on_step_end_tensor_inputs=[
+        "prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+    ],
+).images[0]
+```

examples/community/pipeline_hunyuandit_differential_img2img.py

@@ -1008,6 +1008,8 @@ class HunyuanDiTDifferentialImg2ImgPipeline(DiffusionPipeline):
             self.transformer.inner_dim // self.transformer.num_heads,
             grid_crops_coords,
             (grid_height, grid_width),
+            device=device,
+            output_type="pt",
         )
 
         style = torch.tensor([0], device=device)

examples/community/regional_prompting_stable_diffusion.py

@@ -3,13 +3,12 @@ from typing import Dict, Optional
 
 import torch
 import torchvision.transforms.functional as FF
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from diffusers import StableDiffusionPipeline
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
-from diffusers.utils import USE_PEFT_BACKEND
 
 
 try:
@@ -17,6 +16,7 @@ try:
 except ImportError:
     Compel = None
 
+KBASE = "ADDBASE"
 KCOMM = "ADDCOMM"
 KBRK = "BREAK"
 
@@ -34,6 +34,11 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
 
         Optional
             rp_args["save_mask"]: True/False (save masks in prompt mode)
+            rp_args["power"]: int (power for attention maps in prompt mode)
+            rp_args["base_ratio"]:
+                float (Sets the ratio of the base prompt)
+                ex) 0.2 (20%*BASE_PROMPT + 80%*REGION_PROMPT)
+                [Use base prompt](https://github.com/hako-mikan/sd-webui-regional-prompter?tab=readme-ov-file#use-base-prompt)
 
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -70,6 +75,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__(
@@ -80,6 +86,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             scheduler,
             safety_checker,
             feature_extractor,
+            image_encoder,
             requires_safety_checker,
         )
         self.register_modules(
@@ -90,6 +97,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
 
     @torch.no_grad()
@@ -110,17 +118,40 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         rp_args: Dict[str, str] = None,
     ):
         active = KBRK in prompt[0] if isinstance(prompt, list) else KBRK in prompt
+        use_base = KBASE in prompt[0] if isinstance(prompt, list) else KBASE in prompt
         if negative_prompt is None:
             negative_prompt = "" if isinstance(prompt, str) else [""] * len(prompt)
 
         device = self._execution_device
         regions = 0
 
+        self.base_ratio = float(rp_args["base_ratio"]) if "base_ratio" in rp_args else 0.0
         self.power = int(rp_args["power"]) if "power" in rp_args else 1
 
         prompts = prompt if isinstance(prompt, list) else [prompt]
-        n_prompts = negative_prompt if isinstance(prompt, str) else [negative_prompt]
+        n_prompts = negative_prompt if isinstance(negative_prompt, list) else [negative_prompt]
         self.batch = batch = num_images_per_prompt * len(prompts)
+
+        if use_base:
+            bases = prompts.copy()
+            n_bases = n_prompts.copy()
+
+            for i, prompt in enumerate(prompts):
+                parts = prompt.split(KBASE)
+                if len(parts) == 2:
+                    bases[i], prompts[i] = parts
+                elif len(parts) > 2:
+                    raise ValueError(f"Multiple instances of {KBASE} found in prompt: {prompt}")
+            for i, prompt in enumerate(n_prompts):
+                n_parts = prompt.split(KBASE)
+                if len(n_parts) == 2:
+                    n_bases[i], n_prompts[i] = n_parts
+                elif len(n_parts) > 2:
+                    raise ValueError(f"Multiple instances of {KBASE} found in negative prompt: {prompt}")
+
+            all_bases_cn, _ = promptsmaker(bases, num_images_per_prompt)
+            all_n_bases_cn, _ = promptsmaker(n_bases, num_images_per_prompt)
+
         all_prompts_cn, all_prompts_p = promptsmaker(prompts, num_images_per_prompt)
         all_n_prompts_cn, _ = promptsmaker(n_prompts, num_images_per_prompt)
 
@@ -137,8 +168,16 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
 
             conds = getcompelembs(all_prompts_cn)
             unconds = getcompelembs(all_n_prompts_cn)
-            embs = getcompelembs(prompts)
-            n_embs = getcompelembs(n_prompts)
+            base_embs = getcompelembs(all_bases_cn) if use_base else None
+            base_n_embs = getcompelembs(all_n_bases_cn) if use_base else None
+            # When using base, it seems more reasonable to use base prompts as prompt_embeddings rather than regional prompts
+            embs = getcompelembs(prompts) if not use_base else base_embs
+            n_embs = getcompelembs(n_prompts) if not use_base else base_n_embs
+
+            if use_base and self.base_ratio > 0:
+                conds = self.base_ratio * base_embs + (1 - self.base_ratio) * conds
+                unconds = self.base_ratio * base_n_embs + (1 - self.base_ratio) * unconds
+
             prompt = negative_prompt = None
         else:
             conds = self.encode_prompt(prompts, device, 1, True)[0]
@@ -147,6 +186,18 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                 if equal
                 else self.encode_prompt(all_n_prompts_cn, device, 1, True)[0]
             )
+
+            if use_base and self.base_ratio > 0:
+                base_embs = self.encode_prompt(bases, device, 1, True)[0]
+                base_n_embs = (
+                    self.encode_prompt(n_bases, device, 1, True)[0]
+                    if equal
+                    else self.encode_prompt(all_n_bases_cn, device, 1, True)[0]
+                )
+
+                conds = self.base_ratio * base_embs + (1 - self.base_ratio) * conds
+                unconds = self.base_ratio * base_n_embs + (1 - self.base_ratio) * unconds
+
             embs = n_embs = None
 
         if not active:
@@ -225,8 +276,6 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
 
                     residual = hidden_states
 
-                    args = () if USE_PEFT_BACKEND else (scale,)
-
                     if attn.spatial_norm is not None:
                         hidden_states = attn.spatial_norm(hidden_states, temb)
 
@@ -247,16 +296,15 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     if attn.group_norm is not None:
                         hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
 
-                    args = () if USE_PEFT_BACKEND else (scale,)
-                    query = attn.to_q(hidden_states, *args)
+                    query = attn.to_q(hidden_states)
 
                     if encoder_hidden_states is None:
                         encoder_hidden_states = hidden_states
                     elif attn.norm_cross:
                         encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
 
-                    key = attn.to_k(encoder_hidden_states, *args)
-                    value = attn.to_v(encoder_hidden_states, *args)
+                    key = attn.to_k(encoder_hidden_states)
+                    value = attn.to_v(encoder_hidden_states)
 
                     inner_dim = key.shape[-1]
                     head_dim = inner_dim // attn.heads
@@ -283,7 +331,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     hidden_states = hidden_states.to(query.dtype)
 
                     # linear proj
-                    hidden_states = attn.to_out[0](hidden_states, *args)
+                    hidden_states = attn.to_out[0](hidden_states)
                     # dropout
                     hidden_states = attn.to_out[1](hidden_states)
 
@@ -410,9 +458,9 @@ def promptsmaker(prompts, batch):
         add = ""
         if KCOMM in prompt:
             add, prompt = prompt.split(KCOMM)
-            add = add + " "
-        prompts = prompt.split(KBRK)
-        out_p.append([add + p for p in prompts])
+            add = add.strip() + " "
+        prompts = [p.strip() for p in prompt.split(KBRK)]
+        out_p.append([add + p for i, p in enumerate(prompts)])
     out = [None] * batch * len(out_p[0]) * len(out_p)
     for p, prs in enumerate(out_p):  # inputs prompts
         for r, pr in enumerate(prs):  # prompts for regions
@@ -449,7 +497,6 @@ def make_cells(ratios):
             add = []
             startend(add, inratios[1:])
             icells.append(add)
-
     return ocells, icells, sum(len(cell) for cell in icells)
 
 

examples/community/stable_diffusion_xl_reference.py

@@ -1,5 +1,6 @@
 # Based on stable_diffusion_reference.py
 
+import inspect
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
@@ -7,28 +8,33 @@ import PIL.Image
 import torch
 
 from diffusers import StableDiffusionXLPipeline
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
 from diffusers.models.attention import BasicTransformerBlock
-from diffusers.models.unets.unet_2d_blocks import (
-    CrossAttnDownBlock2D,
-    CrossAttnUpBlock2D,
-    DownBlock2D,
-    UpBlock2D,
-)
-from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
-from diffusers.utils import PIL_INTERPOLATION, logging
+from diffusers.models.unets.unet_2d_blocks import CrossAttnDownBlock2D, CrossAttnUpBlock2D, DownBlock2D, UpBlock2D
+from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+from diffusers.utils import PIL_INTERPOLATION, deprecate, 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  # type: ignore
+
+    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 UniPCMultistepScheduler
+        >>> from diffusers.schedulers import UniPCMultistepScheduler
         >>> from diffusers.utils import load_image
 
-        >>> input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
+        >>> input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl_reference_input_cat.jpg")
 
         >>> pipe = StableDiffusionXLReferencePipeline.from_pretrained(
             "stabilityai/stable-diffusion-xl-base-1.0",
@@ -38,7 +44,7 @@ EXAMPLE_DOC_STRING = """
 
         >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
         >>> result_img = pipe(ref_image=input_image,
-                        prompt="1girl",
+                        prompt="a dog",
                         num_inference_steps=20,
                         reference_attn=True,
                         reference_adain=True).images[0]
@@ -56,8 +62,6 @@ def torch_dfs(model: torch.nn.Module):
 
 
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
-
-
 def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     """
     Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
@@ -72,33 +76,102 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
+# 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,
+):
+    r"""
+    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 StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
-    def _default_height_width(self, height, width, image):
-        # NOTE: It is possible that a list of images have different
-        # dimensions for each image, so just checking the first image
-        # is not _exactly_ correct, but it is simple.
-        while isinstance(image, list):
-            image = image[0]
+    def prepare_ref_latents(self, refimage, batch_size, dtype, device, generator, do_classifier_free_guidance):
+        refimage = refimage.to(device=device)
+        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
+            self.upcast_vae()
+            refimage = refimage.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+        if refimage.dtype != self.vae.dtype:
+            refimage = refimage.to(dtype=self.vae.dtype)
+        # encode the mask image into latents space so we can concatenate it to the latents
+        if isinstance(generator, list):
+            ref_image_latents = [
+                self.vae.encode(refimage[i : i + 1]).latent_dist.sample(generator=generator[i])
+                for i in range(batch_size)
+            ]
+            ref_image_latents = torch.cat(ref_image_latents, dim=0)
+        else:
+            ref_image_latents = self.vae.encode(refimage).latent_dist.sample(generator=generator)
+        ref_image_latents = self.vae.config.scaling_factor * ref_image_latents
 
-        if height is None:
-            if isinstance(image, PIL.Image.Image):
-                height = image.height
-            elif isinstance(image, torch.Tensor):
-                height = image.shape[2]
+        # duplicate mask and ref_image_latents for each generation per prompt, using mps friendly method
+        if ref_image_latents.shape[0] < batch_size:
+            if not batch_size % ref_image_latents.shape[0] == 0:
+                raise ValueError(
+                    "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+                    f" to a total batch size of {batch_size}, but {ref_image_latents.shape[0]} images were passed."
+                    " Make sure the number of images that you pass is divisible by the total requested batch size."
+                )
+            ref_image_latents = ref_image_latents.repeat(batch_size // ref_image_latents.shape[0], 1, 1, 1)
 
-            height = (height // 8) * 8  # round down to nearest multiple of 8
+        ref_image_latents = torch.cat([ref_image_latents] * 2) if do_classifier_free_guidance else ref_image_latents
 
-        if width is None:
-            if isinstance(image, PIL.Image.Image):
-                width = image.width
-            elif isinstance(image, torch.Tensor):
-                width = image.shape[3]
+        # aligning device to prevent device errors when concating it with the latent model input
+        ref_image_latents = ref_image_latents.to(device=device, dtype=dtype)
+        return ref_image_latents
 
-            width = (width // 8) * 8
-
-        return height, width
-
-    def prepare_image(
+    def prepare_ref_image(
         self,
         image,
         width,
@@ -151,41 +224,42 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
 
         return image
 
-    def prepare_ref_latents(self, refimage, batch_size, dtype, device, generator, do_classifier_free_guidance):
-        refimage = refimage.to(device=device)
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            refimage = refimage.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-        if refimage.dtype != self.vae.dtype:
-            refimage = refimage.to(dtype=self.vae.dtype)
-        # encode the mask image into latents space so we can concatenate it to the latents
-        if isinstance(generator, list):
-            ref_image_latents = [
-                self.vae.encode(refimage[i : i + 1]).latent_dist.sample(generator=generator[i])
-                for i in range(batch_size)
-            ]
-            ref_image_latents = torch.cat(ref_image_latents, dim=0)
-        else:
-            ref_image_latents = self.vae.encode(refimage).latent_dist.sample(generator=generator)
-        ref_image_latents = self.vae.config.scaling_factor * ref_image_latents
+    def check_ref_inputs(
+        self,
+        ref_image,
+        reference_guidance_start,
+        reference_guidance_end,
+        style_fidelity,
+        reference_attn,
+        reference_adain,
+    ):
+        ref_image_is_pil = isinstance(ref_image, PIL.Image.Image)
+        ref_image_is_tensor = isinstance(ref_image, torch.Tensor)
 
-        # duplicate mask and ref_image_latents for each generation per prompt, using mps friendly method
-        if ref_image_latents.shape[0] < batch_size:
-            if not batch_size % ref_image_latents.shape[0] == 0:
-                raise ValueError(
-                    "The passed images and the required batch size don't match. Images are supposed to be duplicated"
-                    f" to a total batch size of {batch_size}, but {ref_image_latents.shape[0]} images were passed."
-                    " Make sure the number of images that you pass is divisible by the total requested batch size."
-                )
-            ref_image_latents = ref_image_latents.repeat(batch_size // ref_image_latents.shape[0], 1, 1, 1)
+        if not ref_image_is_pil and not ref_image_is_tensor:
+            raise TypeError(
+                f"ref image must be passed and be one of PIL image or torch tensor, but is {type(ref_image)}"
+            )
 
-        ref_image_latents = torch.cat([ref_image_latents] * 2) if do_classifier_free_guidance else ref_image_latents
+        if not reference_attn and not reference_adain:
+            raise ValueError("`reference_attn` or `reference_adain` must be True.")
 
-        # aligning device to prevent device errors when concating it with the latent model input
-        ref_image_latents = ref_image_latents.to(device=device, dtype=dtype)
-        return ref_image_latents
+        if style_fidelity < 0.0:
+            raise ValueError(f"style fidelity: {style_fidelity} can't be smaller than 0.")
+        if style_fidelity > 1.0:
+            raise ValueError(f"style fidelity: {style_fidelity} can't be larger than 1.0.")
+
+        if reference_guidance_start >= reference_guidance_end:
+            raise ValueError(
+                f"reference guidance start: {reference_guidance_start} cannot be larger or equal to reference guidance end: {reference_guidance_end}."
+            )
+        if reference_guidance_start < 0.0:
+            raise ValueError(f"reference guidance start: {reference_guidance_start} can't be smaller than 0.")
+        if reference_guidance_end > 1.0:
+            raise ValueError(f"reference guidance end: {reference_guidance_end} can't be larger than 1.0.")
 
     @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
         self,
         prompt: Union[str, List[str]] = None,
@@ -194,6 +268,8 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
         height: Optional[int] = None,
         width: Optional[int] = None,
         num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
         denoising_end: Optional[float] = None,
         guidance_scale: float = 5.0,
         negative_prompt: Optional[Union[str, List[str]]] = None,
@@ -206,28 +282,220 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
         negative_prompt_embeds: Optional[torch.Tensor] = None,
         pooled_prompt_embeds: Optional[torch.Tensor] = None,
         negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
-        callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
-        callback_steps: int = 1,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
         guidance_rescale: float = 0.0,
         original_size: Optional[Tuple[int, int]] = None,
         crops_coords_top_left: Tuple[int, int] = (0, 0),
         target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         attention_auto_machine_weight: float = 1.0,
         gn_auto_machine_weight: float = 1.0,
+        reference_guidance_start: float = 0.0,
+        reference_guidance_end: float = 1.0,
         style_fidelity: float = 0.5,
         reference_attn: bool = True,
         reference_adain: bool = True,
+        **kwargs,
     ):
-        assert reference_attn or reference_adain, "`reference_attn` or `reference_adain` must be True."
+        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 the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            ref_image (`torch.Tensor`, `PIL.Image.Image`):
+                The Reference Control input condition. Reference Control uses this input condition to generate guidance to Unet. If
+                the type is specified as `Torch.Tensor`, it is passed to Reference Control as is. `PIL.Image.Image` can
+                also be accepted as an image.
+            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.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            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.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            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.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            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 in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` 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.Tensor`, *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.Tensor`, *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.Tensor`, *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.Tensor`, *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.Tensor`, *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.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+                provided, embeddings are computed from the `ip_adapter_image` 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.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. 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.
+            attention_auto_machine_weight (`float`):
+                Weight of using reference query for self attention's context.
+                If attention_auto_machine_weight=1.0, use reference query for all self attention's context.
+            gn_auto_machine_weight (`float`):
+                Weight of using reference adain. If gn_auto_machine_weight=2.0, use all reference adain plugins.
+            reference_guidance_start (`float`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the reference ControlNet starts applying.
+            reference_guidance_end (`float`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the reference ControlNet stops applying.
+            style_fidelity (`float`):
+                style fidelity of ref_uncond_xt. If style_fidelity=1.0, control more important,
+                elif style_fidelity=0.0, prompt more important, else balanced.
+            reference_attn (`bool`):
+                Whether to use reference query for self attention's context.
+            reference_adain (`bool`):
+                Whether to use reference adain.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
 
         # 0. Default height and width to unet
-        # height, width = self._default_height_width(height, width, ref_image)
-
         height = height or self.default_sample_size * self.vae_scale_factor
         width = width or self.default_sample_size * self.vae_scale_factor
+
         original_size = original_size or (height, width)
         target_size = target_size or (height, width)
 
@@ -244,8 +512,27 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             negative_prompt_embeds,
             pooled_prompt_embeds,
             negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
         )
 
+        self.check_ref_inputs(
+            ref_image,
+            reference_guidance_start,
+            reference_guidance_end,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._interrupt = False
+
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
             batch_size = 1
@@ -256,15 +543,11 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
 
         device = self._execution_device
 
-        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
-        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
-        # corresponds to doing no classifier free guidance.
-        do_classifier_free_guidance = guidance_scale > 1.0
-
         # 3. Encode input prompt
-        text_encoder_lora_scale = (
-            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
         )
+
         (
             prompt_embeds,
             negative_prompt_embeds,
@@ -275,17 +558,19 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             prompt_2=prompt_2,
             device=device,
             num_images_per_prompt=num_images_per_prompt,
-            do_classifier_free_guidance=do_classifier_free_guidance,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
             negative_prompt=negative_prompt,
             negative_prompt_2=negative_prompt_2,
             prompt_embeds=prompt_embeds,
             negative_prompt_embeds=negative_prompt_embeds,
             pooled_prompt_embeds=pooled_prompt_embeds,
             negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
-            lora_scale=text_encoder_lora_scale,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
         )
+
         # 4. Preprocess reference image
-        ref_image = self.prepare_image(
+        ref_image = self.prepare_ref_image(
             image=ref_image,
             width=width,
             height=height,
@@ -296,9 +581,9 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
         )
 
         # 5. Prepare timesteps
-        self.scheduler.set_timesteps(num_inference_steps, device=device)
-
-        timesteps = self.scheduler.timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas
+        )
 
         # 6. Prepare latent variables
         num_channels_latents = self.unet.config.in_channels
@@ -312,6 +597,7 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             generator,
             latents,
         )
+
         # 7. Prepare reference latent variables
         ref_image_latents = self.prepare_ref_latents(
             ref_image,
@@ -319,13 +605,21 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             prompt_embeds.dtype,
             device,
             generator,
-            do_classifier_free_guidance,
+            self.do_classifier_free_guidance,
         )
 
         # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
-        # 9. Modify self attebtion and group norm
+        # 8.1 Create tensor stating which reference controlnets to keep
+        reference_keeps = []
+        for i in range(len(timesteps)):
+            reference_keep = 1.0 - float(
+                i / len(timesteps) < reference_guidance_start or (i + 1) / len(timesteps) > reference_guidance_end
+            )
+            reference_keeps.append(reference_keep)
+
+        # 8.2 Modify self attention and group norm
         MODE = "write"
         uc_mask = (
             torch.Tensor([1] * batch_size * num_images_per_prompt + [0] * batch_size * num_images_per_prompt)
@@ -333,6 +627,8 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             .bool()
         )
 
+        do_classifier_free_guidance = self.do_classifier_free_guidance
+
         def hacked_basic_transformer_inner_forward(
             self,
             hidden_states: torch.Tensor,
@@ -604,7 +900,7 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             return hidden_states
 
         def hacked_UpBlock2D_forward(
-            self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, **kwargs
+            self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None, *args, **kwargs
         ):
             eps = 1e-6
             for i, resnet in enumerate(self.resnets):
@@ -684,7 +980,7 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
                 module.var_bank = []
                 module.gn_weight *= 2
 
-        # 10. Prepare added time ids & embeddings
+        # 9. Prepare added time ids & embeddings
         add_text_embeds = pooled_prompt_embeds
         if self.text_encoder_2 is None:
             text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
@@ -698,62 +994,101 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             dtype=prompt_embeds.dtype,
             text_encoder_projection_dim=text_encoder_projection_dim,
         )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
 
-        if do_classifier_free_guidance:
+        if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
             add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
-            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
 
         prompt_embeds = prompt_embeds.to(device)
         add_text_embeds = add_text_embeds.to(device)
         add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
 
-        # 11. Denoising loop
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 10. Denoising loop
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
 
         # 10.1 Apply denoising_end
-        if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
             discrete_timestep_cutoff = int(
                 round(
                     self.scheduler.config.num_train_timesteps
-                    - (denoising_end * self.scheduler.config.num_train_timesteps)
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
                 )
             )
             num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
             timesteps = timesteps[:num_inference_steps]
 
+        # 11. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
                 # expand the latents if we are doing classifier free guidance
-                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
 
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
+                # predict the noise residual
                 added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
 
                 # ref only part
-                noise = randn_tensor(
-                    ref_image_latents.shape, generator=generator, device=device, dtype=ref_image_latents.dtype
-                )
-                ref_xt = self.scheduler.add_noise(
-                    ref_image_latents,
-                    noise,
-                    t.reshape(
-                        1,
-                    ),
-                )
-                ref_xt = self.scheduler.scale_model_input(ref_xt, t)
+                if reference_keeps[i] > 0:
+                    noise = randn_tensor(
+                        ref_image_latents.shape, generator=generator, device=device, dtype=ref_image_latents.dtype
+                    )
+                    ref_xt = self.scheduler.add_noise(
+                        ref_image_latents,
+                        noise,
+                        t.reshape(
+                            1,
+                        ),
+                    )
+                    ref_xt = self.scheduler.scale_model_input(ref_xt, t)
 
-                MODE = "write"
-
-                self.unet(
-                    ref_xt,
-                    t,
-                    encoder_hidden_states=prompt_embeds,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    added_cond_kwargs=added_cond_kwargs,
-                    return_dict=False,
-                )
+                    MODE = "write"
+                    self.unet(
+                        ref_xt,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )
 
                 # predict the noise residual
                 MODE = "read"
@@ -761,22 +1096,44 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
-                    cross_attention_kwargs=cross_attention_kwargs,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
                     added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 
                 # perform guidance
-                if do_classifier_free_guidance:
+                if self.do_classifier_free_guidance:
                     noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
 
-                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
                     # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
-                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
 
                 # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, 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)
+                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
 
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
@@ -785,6 +1142,9 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
         if not output_type == "latent":
             # make sure the VAE is in float32 mode, as it overflows in float16
             needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
@@ -792,25 +1152,43 @@ class StableDiffusionXLReferencePipeline(StableDiffusionXLPipeline):
             if needs_upcasting:
                 self.upcast_vae()
                 latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+            elif latents.dtype != self.vae.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
+                    self.vae = self.vae.to(latents.dtype)
 
-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
 
             # cast back to fp16 if needed
             if needs_upcasting:
                 self.vae.to(dtype=torch.float16)
         else:
             image = latents
-            return StableDiffusionXLPipelineOutput(images=image)
 
-        # apply watermark if available
-        if self.watermark is not None:
-            image = self.watermark.apply_watermark(image)
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
 
-        image = self.image_processor.postprocess(image, output_type=output_type)
+            image = self.image_processor.postprocess(image, output_type=output_type)
 
-        # Offload last model to CPU
-        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
-            self.final_offload_hook.offload()
+        # Offload all models
+        self.maybe_free_model_hooks()
 
         if not return_dict:
             return (image,)

examples/controlnet/README_sd3.md

@@ -1,6 +1,6 @@
-# ControlNet training example for Stable Diffusion 3 (SD3)
+# ControlNet training example for Stable Diffusion 3/3.5 (SD3/3.5)
 
-The `train_controlnet_sd3.py` script shows how to implement the ControlNet training procedure and adapt it for [Stable Diffusion 3](https://arxiv.org/abs/2403.03206).
+The `train_controlnet_sd3.py` script shows how to implement the ControlNet training procedure and adapt it for [Stable Diffusion 3](https://arxiv.org/abs/2403.03206) and [Stable Diffusion 3.5](https://stability.ai/news/introducing-stable-diffusion-3-5).
 
 ## Running locally with PyTorch
 
@@ -51,9 +51,9 @@ Please download the dataset and unzip it in the directory `fill50k` in the `exam
 
 ## Training
 
-First download the SD3 model from [Hugging Face Hub](https://huggingface.co/stabilityai/stable-diffusion-3-medium). We will use it as a base model for the ControlNet training.
+First download the SD3 model from [Hugging Face Hub](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) or the SD3.5 model from [Hugging Face Hub](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium). We will use it as a base model for the ControlNet training.
 > [!NOTE]
-> As the model is gated, before using it with diffusers you first need to go to the [Stable Diffusion 3 Medium Hugging Face page](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+> As the model is gated, before using it with diffusers you first need to go to the [Stable Diffusion 3 Medium Hugging Face page](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) or [Stable Diffusion 3.5 Large Hugging Face page](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
 
 ```bash
 huggingface-cli login
@@ -90,6 +90,8 @@ accelerate launch train_controlnet_sd3.py \
     --gradient_accumulation_steps=4
 ```
 
+To train a ControlNet model for Stable Diffusion 3.5, replace the `MODEL_DIR` with `stabilityai/stable-diffusion-3.5-medium`.
+
 To better track our training experiments, we're using flags `validation_image`, `validation_prompt`, and `validation_steps` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 Our experiments were conducted on a single 40GB A100 GPU.
@@ -124,6 +126,8 @@ image = pipe(
 image.save("./output.png")
 ```
 
+Similarly, for SD3.5, replace the `base_model_path` with `stabilityai/stable-diffusion-3.5-medium` and controlnet_path `DavyMorgan/sd35-controlnet-out'.
+
 ## Notes
 
 ### GPU usage
@@ -135,6 +139,8 @@ Make sure to use the right GPU when configuring the [accelerator](https://huggin
 
 ## Example results
 
+### SD3
+
 #### After 500 steps with batch size 8
 
 | |  |
@@ -150,3 +156,20 @@ Make sure to use the right GPU when configuring the [accelerator](https://huggin
 || pale golden rod circle with old lace background |
  ![conditioning image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/controlnet_training/conditioning_image_1.png) | ![pale golden rod circle with old lace background](https://huggingface.co/datasets/DavyMorgan/sd3-controlnet-results/resolve/main/step-6500.png) |
 
+### SD3.5
+
+#### After 500 steps with batch size 8
+
+| |                                                                                                                                                     |
+|-------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------:|
+||                                                   pale golden rod circle with old lace background                                                   |
+ ![conditioning image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/controlnet_training/conditioning_image_1.png) | ![pale golden rod circle with old lace background](https://huggingface.co/datasets/DavyMorgan/sd3-controlnet-results/resolve/main/step-500-3.5.png) |
+
+
+#### After 3000 steps with batch size 8:
+
+| |                                                                                                                                                      |
+|-------------------|:----------------------------------------------------------------------------------------------------------------------------------------------------:|
+||                                                   pale golden rod circle with old lace background                                                    |
+ ![conditioning image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/controlnet_training/conditioning_image_1.png) | ![pale golden rod circle with old lace background](https://huggingface.co/datasets/DavyMorgan/sd3-controlnet-results/resolve/main/step-3000-3.5.png) |
+

examples/controlnet/test_controlnet.py

@@ -138,6 +138,27 @@ class ControlNetSD3(ExamplesTestsAccelerate):
             self.assertTrue(os.path.isfile(os.path.join(tmpdir, "diffusion_pytorch_model.safetensors")))
 
 
+class ControlNetSD35(ExamplesTestsAccelerate):
+    def test_controlnet_sd3(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            examples/controlnet/train_controlnet_sd3.py
+            --pretrained_model_name_or_path=hf-internal-testing/tiny-sd35-pipe
+            --dataset_name=hf-internal-testing/fill10
+            --output_dir={tmpdir}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --controlnet_model_name_or_path=DavyMorgan/tiny-controlnet-sd35
+            --max_train_steps=4
+            --checkpointing_steps=2
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "diffusion_pytorch_model.safetensors")))
+
+
 class ControlNetflux(ExamplesTestsAccelerate):
     def test_controlnet_flux(self):
         with tempfile.TemporaryDirectory() as tmpdir:

examples/controlnet/train_controlnet.py

@@ -571,9 +571,6 @@ def parse_args(input_args=None):
     if args.dataset_name is None and args.train_data_dir is None:
         raise ValueError("Specify either `--dataset_name` or `--train_data_dir`")
 
-    if args.dataset_name is not None and args.train_data_dir is not None:
-        raise ValueError("Specify only one of `--dataset_name` or `--train_data_dir`")
-
     if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
         raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
 
@@ -615,6 +612,7 @@ def make_train_dataset(args, tokenizer, accelerator):
             args.dataset_name,
             args.dataset_config_name,
             cache_dir=args.cache_dir,
+            data_dir=args.train_data_dir,
         )
     else:
         if args.train_data_dir is not None:

examples/controlnet/train_controlnet_sd3.py

@@ -263,6 +263,12 @@ def parse_args(input_args=None):
         help="Path to pretrained controlnet model or model identifier from huggingface.co/models."
         " If not specified controlnet weights are initialized from unet.",
     )
+    parser.add_argument(
+        "--num_extra_conditioning_channels",
+        type=int,
+        default=0,
+        help="Number of extra conditioning channels for controlnet.",
+    )
     parser.add_argument(
         "--revision",
         type=str,
@@ -539,6 +545,9 @@ def parse_args(input_args=None):
         default=77,
         help="Maximum sequence length to use with with the T5 text encoder",
     )
+    parser.add_argument(
+        "--dataset_preprocess_batch_size", type=int, default=1000, help="Batch size for preprocessing dataset."
+    )
     parser.add_argument(
         "--validation_prompt",
         type=str,
@@ -986,7 +995,9 @@ def main(args):
         controlnet = SD3ControlNetModel.from_pretrained(args.controlnet_model_name_or_path)
     else:
         logger.info("Initializing controlnet weights from transformer")
-        controlnet = SD3ControlNetModel.from_transformer(transformer)
+        controlnet = SD3ControlNetModel.from_transformer(
+            transformer, num_extra_conditioning_channels=args.num_extra_conditioning_channels
+        )
 
     transformer.requires_grad_(False)
     vae.requires_grad_(False)
@@ -1123,7 +1134,12 @@ def main(args):
         # fingerprint used by the cache for the other processes to load the result
         # details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401
         new_fingerprint = Hasher.hash(args)
-        train_dataset = train_dataset.map(compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint)
+        train_dataset = train_dataset.map(
+            compute_embeddings_fn,
+            batched=True,
+            batch_size=args.dataset_preprocess_batch_size,
+            new_fingerprint=new_fingerprint,
+        )
 
     del text_encoder_one, text_encoder_two, text_encoder_three
     del tokenizer_one, tokenizer_two, tokenizer_three

examples/controlnet/train_controlnet_sdxl.py

@@ -598,9 +598,6 @@ def parse_args(input_args=None):
     if args.dataset_name is None and args.train_data_dir is None:
         raise ValueError("Specify either `--dataset_name` or `--train_data_dir`")
 
-    if args.dataset_name is not None and args.train_data_dir is not None:
-        raise ValueError("Specify only one of `--dataset_name` or `--train_data_dir`")
-
     if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
         raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
 
@@ -642,6 +639,7 @@ def get_train_dataset(args, accelerator):
             args.dataset_name,
             args.dataset_config_name,
             cache_dir=args.cache_dir,
+            data_dir=args.train_data_dir,
         )
     else:
         if args.train_data_dir is not None:

examples/dreambooth/README_flux.md

@@ -118,7 +118,7 @@ accelerate launch train_dreambooth_flux.py \
 
 To better track our training experiments, we're using the following flags in the command above:
 
-* `report_to="wandb` will ensure the training runs are tracked on Weights and Biases. To use it, be sure to install `wandb` with `pip install wandb`.
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
 * `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 > [!NOTE]

examples/dreambooth/README_sd3.md

@@ -105,7 +105,7 @@ accelerate launch train_dreambooth_sd3.py \
 
 To better track our training experiments, we're using the following flags in the command above:
 
-* `report_to="wandb` will ensure the training runs are tracked on Weights and Biases. To use it, be sure to install `wandb` with `pip install wandb`.
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
 * `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 > [!NOTE]

examples/dreambooth/README_sdxl.md

@@ -99,7 +99,7 @@ accelerate launch train_dreambooth_lora_sdxl.py \
 
 To better track our training experiments, we're using the following flags in the command above:
 
-* `report_to="wandb` will ensure the training runs are tracked on Weights and Biases. To use it, be sure to install `wandb` with `pip install wandb`.
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
 * `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 Our experiments were conducted on a single 40GB A100 GPU.

examples/dreambooth/REAMDE_sana.md

@@ -0,0 +1,127 @@
+# DreamBooth training example for SANA
+
+[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few (3~5) images of a subject.
+
+The `train_dreambooth_lora_sana.py` script shows how to implement the training procedure with [LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) and adapt it for [SANA](https://arxiv.org/abs/2410.10629). 
+
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+## Running locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the `examples/dreambooth` folder and run
+```bash
+pip install -r requirements_sana.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell (e.g., a notebook)
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.14.0` installed in your environment.
+
+
+### Dog toy example
+
+Now let's get our dataset. For this example we will use some dog images: https://huggingface.co/datasets/diffusers/dog-example.
+
+Let's first download it locally:
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = "./dog"
+snapshot_download(
+    "diffusers/dog-example",
+    local_dir=local_dir, repo_type="dataset",
+    ignore_patterns=".gitattributes",
+)
+```
+
+This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
+
+Now, we can launch training using:
+
+```bash
+export MODEL_NAME="Efficient-Large-Model/Sana_1600M_1024px_diffusers"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-sana-lora"
+
+accelerate launch train_dreambooth_lora_sana.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+For using `push_to_hub`, make you're logged into your Hugging Face account:
+
+```bash
+huggingface-cli login
+```
+
+To better track our training experiments, we're using the following flags in the command above:
+
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
+* `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
+
+## Notes
+
+Additionally, we welcome you to explore the following CLI arguments:
+
+* `--lora_layers`: The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. E.g. - "to_k,to_q,to_v" will result in lora training of attention layers only.
+* `--complex_human_instruction`: Instructions for complex human attention as shown in [here](https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55).
+* `--max_sequence_length`: Maximum sequence length to use for text embeddings.
+
+
+We provide several options for optimizing memory optimization:
+
+* `--offload`: When enabled, we will offload the text encoder and VAE to CPU, when they are not used.
+* `cache_latents`: When enabled, we will pre-compute the latents from the input images with the VAE and remove the VAE from memory once done.
+* `--use_8bit_adam`: When enabled, we will use the 8bit version of AdamW provided by the `bitsandbytes` library.
+
+Refer to the [official documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana) of the `SanaPipeline` to know more about the models available under the SANA family and their preferred dtypes during inference.

examples/dreambooth/requirements_sana.txt

@@ -0,0 +1,8 @@
+accelerate>=1.0.0
+torchvision
+transformers>=4.47.0
+ftfy
+tensorboard
+Jinja2
+peft>=0.14.0
+sentencepiece

examples/dreambooth/train_dreambooth.py

@@ -1300,16 +1300,17 @@ def main(args):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    base_weight = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
 
                     if noise_scheduler.config.prediction_type == "v_prediction":
                         # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
+                        divisor = snr + 1
                     else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
+                        divisor = snr
+
+                    mse_loss_weights = (
+                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / divisor
+                    )
+
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -1648,11 +1648,15 @@ def main(args):
                             prompt=prompts,
                         )
                 else:
+                    elems_to_repeat = len(prompts)
                     if args.train_text_encoder:
                         prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
                             text_encoders=[text_encoder_one, text_encoder_two],
                             tokenizers=[None, None],
-                            text_input_ids_list=[tokens_one, tokens_two],
+                            text_input_ids_list=[
+                                tokens_one.repeat(elems_to_repeat, 1),
+                                tokens_two.repeat(elems_to_repeat, 1),
+                            ],
                             max_sequence_length=args.max_sequence_length,
                             device=accelerator.device,
                             prompt=args.instance_prompt,

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1294,10 +1294,13 @@ def main(args):
             for model in models:
                 if isinstance(model, type(unwrap_model(transformer))):
                     transformer_lora_layers_to_save = get_peft_model_state_dict(model)
-                elif isinstance(model, type(unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
-                elif isinstance(model, type(unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
+                elif isinstance(model, type(unwrap_model(text_encoder_one))):  # or text_encoder_two
+                    # both text encoders are of the same class, so we check hidden size to distinguish between the two
+                    hidden_size = unwrap_model(model).config.hidden_size
+                    if hidden_size == 768:
+                        text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
+                    elif hidden_size == 1280:
+                        text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 

examples/flux-control/README.md

@@ -0,0 +1,204 @@
+# Training Flux Control
+
+This (experimental) example shows how to train Control LoRAs with [Flux](https://huggingface.co/black-forest-labs/FLUX.1-dev) by conditioning it with additional structural controls (like depth maps, poses, etc.). We provide a script for full fine-tuning, too, refer to [this section](#full-fine-tuning). To know more about Flux Control family, refer to the following resources:
+
+* [Docs](https://github.com/black-forest-labs/flux/blob/main/docs/structural-conditioning.md) by Black Forest Labs
+* Diffusers docs ([1](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux#canny-control), [2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux#depth-control))
+
+To incorporate additional condition latents, we expand the input features of Flux.1-Dev from 64 to 128. The first 64 channels correspond to the original input latents to be denoised, while the latter 64 channels correspond to control latents. This expansion happens on the `x_embedder` layer, where the combined latents are projected to the expected feature dimension of rest of the network. Inference is performed using the `FluxControlPipeline`.
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [FLUX.1 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.1-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+The example command below shows how to launch fine-tuning for pose conditions. The dataset ([`raulc0399/open_pose_controlnet`](https://huggingface.co/datasets/raulc0399/open_pose_controlnet)) being used here already has the pose conditions of the original images, so we don't have to compute them.
+
+```bash
+accelerate launch train_control_lora_flux.py \
+  --pretrained_model_name_or_path="black-forest-labs/FLUX.1-dev" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control-lora" \
+  --mixed_precision="bf16" \
+  --train_batch_size=1 \
+  --rank=64 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=5000 \
+  --validation_image="openpose.png" \
+  --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+`openpose.png` comes from [here](https://huggingface.co/Adapter/t2iadapter/resolve/main/openpose.png).
+
+You need to install `diffusers` from the branch of [this PR](https://github.com/huggingface/diffusers/pull/9999). When it's merged, you should install `diffusers` from the `main`.
+
+The training script exposes additional CLI args that might be useful to experiment with:
+
+* `use_lora_bias`: When set, additionally trains the biases of the `lora_B` layer. 
+* `train_norm_layers`: When set, additionally trains the normalization scales. Takes care of saving and loading.
+* `lora_layers`: Specify the layers you want to apply LoRA to. If you specify "all-linear", all the linear layers will be LoRA-attached.
+
+### Training with DeepSpeed
+
+It's possible to train with [DeepSpeed](https://github.com/microsoft/DeepSpeed), specifically leveraging the Zero2 system optimization. To use it, save the following config to an YAML file (feel free to modify as needed):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+And then while launching training, pass the config file:
+
+```bash
+accelerate launch --config_file=CONFIG_FILE.yaml ...
+```
+
+### Inference
+
+The pose images in our dataset were computed using the [`controlnet_aux`](https://github.com/huggingface/controlnet_aux) library. Let's install it first:
+
+```bash
+pip install controlnet_aux
+```
+
+And then we are ready:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import FluxControlPipeline
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("...") # change this.
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  condition_image=image,
+  num_inference_steps=50,
+  joint_attention_kwargs={"scale": 0.9},
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Full fine-tuning
+
+We provide a non-LoRA version of the training script `train_control_flux.py`. Here is an example command:
+
+```bash
+accelerate launch --config_file=accelerate_ds2.yaml train_control_flux.py \
+  --pretrained_model_name_or_path="black-forest-labs/FLUX.1-dev" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control" \
+  --mixed_precision="bf16" \
+  --train_batch_size=2 \
+  --dataloader_num_workers=4 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --proportion_empty_prompts=0.2 \
+  --learning_rate=5e-5 \
+  --adam_weight_decay=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="cosine" \
+  --lr_warmup_steps=1000 \
+  --checkpointing_steps=1000 \
+  --max_train_steps=10000 \
+  --validation_steps=200 \
+  --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
+  --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+Change the `validation_image` and `validation_prompt` as needed.
+
+For inference, this time, we will run:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import FluxControlPipeline, FluxTransformer2DModel
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+transformer = FluxTransformer2DModel.from_pretrained("...") # change this.
+pipe = FluxControlPipeline.from_pretrained(
+  "black-forest-labs/FLUX.1-dev",  transformer=transformer, torch_dtype=torch.bfloat16
+).to("cuda")
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  condition_image=image,
+  num_inference_steps=50,
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Things to note
+
+* The scripts provided in this directory are experimental and educational. This means we may have to tweak things around to get good results on a given condition. We believe this is best done with the community 🤗
+* The scripts are not memory-optimized but we offload the VAE and the text encoders to CPU when they are not used. 
+* We can extract LoRAs from the fully fine-tuned model. While we currently don't provide any utilities for that, users are welcome to refer to [this script](https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/master/control_lora_create.py) that provides a similar functionality. 

examples/flux-control/requirements.txt

@@ -0,0 +1,6 @@
+transformers==4.47.0
+wandb
+torch
+torchvision
+accelerate==1.2.0
+peft>=0.14.0

examples/model_search/README.md

@@ -0,0 +1,175 @@
+# Search models on Civitai and Hugging Face
+
+The [auto_diffusers](https://github.com/suzukimain/auto_diffusers) library provides additional functionalities to Diffusers such as searching for models on Civitai and the Hugging Face Hub.
+Please refer to the original library [here](https://pypi.org/project/auto-diffusers/)
+
+## Installation
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+> [!IMPORTANT]
+> To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the installation up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment.
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install .
+```
+Set up the pipeline. You can also cd to this folder and run it.
+```bash
+!wget https://raw.githubusercontent.com/suzukimain/auto_diffusers/refs/heads/master/src/auto_diffusers/pipeline_easy.py
+```
+
+## Load from Civitai
+```python
+from pipeline_easy import (
+    EasyPipelineForText2Image,
+    EasyPipelineForImage2Image,
+    EasyPipelineForInpainting,
+)
+
+# Text-to-Image
+pipeline = EasyPipelineForText2Image.from_civitai(
+    "search_word",
+    base_model="SD 1.5",
+).to("cuda")
+
+
+# Image-to-Image
+pipeline = EasyPipelineForImage2Image.from_civitai(
+    "search_word",
+    base_model="SD 1.5",
+).to("cuda")
+
+
+# Inpainting
+pipeline = EasyPipelineForInpainting.from_civitai(
+    "search_word",
+    base_model="SD 1.5",
+).to("cuda")
+```
+
+## Load from Hugging Face
+```python
+from pipeline_easy import (
+    EasyPipelineForText2Image,
+    EasyPipelineForImage2Image,
+    EasyPipelineForInpainting,
+)
+
+# Text-to-Image
+pipeline = EasyPipelineForText2Image.from_huggingface(
+    "search_word",
+    checkpoint_format="diffusers",
+).to("cuda")
+
+
+# Image-to-Image
+pipeline = EasyPipelineForImage2Image.from_huggingface(
+    "search_word",
+    checkpoint_format="diffusers",
+).to("cuda")
+
+
+# Inpainting
+pipeline = EasyPipelineForInpainting.from_huggingface(
+    "search_word",
+    checkpoint_format="diffusers",
+).to("cuda")
+```
+
+
+## Search Civitai and Huggingface
+
+```python
+from pipeline_easy import (
+    search_huggingface,
+    search_civitai,
+) 
+
+# Search Lora
+Lora = search_civitai(
+    "Keyword_to_search_Lora",
+    model_type="LORA",
+    base_model = "SD 1.5",
+    download=True,
+    )
+# Load Lora into the pipeline.
+pipeline.load_lora_weights(Lora)
+
+
+# Search TextualInversion
+TextualInversion = search_civitai(
+    "EasyNegative",
+    model_type="TextualInversion",
+    base_model = "SD 1.5",
+    download=True
+)
+# Load TextualInversion into the pipeline.
+pipeline.load_textual_inversion(TextualInversion, token="EasyNegative")
+```
+
+### Search Civitai
+
+> [!TIP]
+> **If an error occurs, insert the `token` and run again.**
+
+#### `EasyPipeline.from_civitai` parameters
+
+| Name            | Type                   | Default       | Description                                                                    |
+|:---------------:|:----------------------:|:-------------:|:-----------------------------------------------------------------------------------:|
+| search_word     | string, Path           | ー            | The search query string. Can be a keyword, Civitai URL, local directory or file path. |
+| model_type      | string                 | `Checkpoint`  | The type of model to search for.  <br>(for example `Checkpoint`, `TextualInversion`, `Controlnet`, `LORA`, `Hypernetwork`, `AestheticGradient`, `Poses`)      |
+| base_model      | string                 | None          | Trained model tag (for example  `SD 1.5`, `SD 3.5`, `SDXL 1.0`) |
+| torch_dtype     | string, torch.dtype    | None          | Override the default `torch.dtype` and load the model with another dtype.     |
+| force_download  | bool                   | 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       | string, Path | None    | Path to the folder where cached files are stored. |
+| resume          | bool   | False         | Whether to resume an incomplete download. |
+| token           | string | None          | API token for Civitai authentication. |
+
+
+#### `search_civitai` parameters
+
+| Name            | Type           | Default       | Description                                                                    |
+|:---------------:|:--------------:|:-------------:|:-----------------------------------------------------------------------------------:|
+| search_word     | string, Path   | ー            | The search query string. Can be a keyword, Civitai URL, local directory or file path. |
+| model_type      | string         | `Checkpoint`  | The type of model to search for. <br>(for example `Checkpoint`, `TextualInversion`, `Controlnet`, `LORA`, `Hypernetwork`, `AestheticGradient`, `Poses`)   |
+| base_model      | string         | None          | Trained model tag (for example  `SD 1.5`, `SD 3.5`, `SDXL 1.0`)                        |
+| download        | bool           | False         | Whether to download the model.                                   |
+| force_download  | bool           | False         | Whether to force the download if the model already exists.                          |
+| cache_dir       | string, Path   | None          | Path to the folder where cached files are stored.                              |
+| resume          | bool           | False         | Whether to resume an incomplete download.                                           |
+| token           | string         | None          | API token for Civitai authentication.                                               |
+| include_params  | bool           | False         | Whether to include parameters in the returned data.           |
+| skip_error      | bool           | False         | Whether to skip errors and return None.                                             |
+
+### Search Huggingface
+
+> [!TIP]
+> **If an error occurs, insert the `token` and run again.**
+
+#### `EasyPipeline.from_huggingface` parameters
+
+| Name                  | Type                | Default        | Description                                                      |
+|:---------------------:|:-------------------:|:--------------:|:----------------------------------------------------------------:|
+| search_word           | string, Path        | ー             | The search query string. Can be a keyword, Hugging Face URL, local directory or file path, or a Hugging Face path (`<creator>/<repo>`). |
+| checkpoint_format     | string              | `single_file`  | The format of the model checkpoint.<br>● `single_file` to search for `single file checkpoint` <br>●`diffusers` to search for `multifolder diffusers format checkpoint` |
+| torch_dtype           | string, torch.dtype | None           | Override the default `torch.dtype` and load the model with another dtype. |
+| force_download        | bool                | 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             | string, Path        | None           | Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.   |
+| token                 | string, bool        | None           | The token to use as HTTP bearer authorization for remote files.  |
+
+
+#### `search_huggingface` parameters
+
+| Name                  | Type                | Default        | Description                                                      |
+|:---------------------:|:-------------------:|:--------------:|:----------------------------------------------------------------:|
+| search_word           | string, Path        | ー             | The search query string. Can be a keyword, Hugging Face URL, local directory or file path, or a Hugging Face path (`<creator>/<repo>`). |
+| checkpoint_format     | string              | `single_file`  | The format of the model checkpoint. <br>● `single_file` to search for `single file checkpoint` <br>●`diffusers` to search for `multifolder diffusers format checkpoint` |
+| pipeline_tag          | string              | None           | Tag to filter models by pipeline.                                |
+| download              | bool                | False          | Whether to download the model.                                   |
+| force_download        | bool                | 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             | string, Path        | None           | Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.   |
+| token                 | string, bool        | None           | The token to use as HTTP bearer authorization for remote files.  |
+| include_params        | bool                | False         | Whether to include parameters in the returned data.               |
+| skip_error            | bool                | False         | Whether to skip errors and return None.                           |

examples/model_search/requirements.txt

@@ -0,0 +1 @@
+huggingface-hub>=0.26.2

examples/research_projects/ip_adapter/README.md

@@ -0,0 +1,226 @@
+# IP Adapter Training Example 
+
+[IP Adapter](https://arxiv.org/abs/2308.06721) is a novel approach designed to enhance text-to-image models such as Stable Diffusion by enabling them to generate images based on image prompts rather than text prompts alone. Unlike traditional methods that rely solely on complex text prompts, IP Adapter introduces the concept of using image prompts, leveraging the idea that "an image is worth a thousand words." By decoupling cross-attention layers for text and image features, IP Adapter effectively integrates image prompts into the generation process without the need for extensive fine-tuning or large computing resources.
+
+## Training locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the example folder and run
+
+```bash
+pip install -r requirements.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell e.g. a notebook
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+Certainly! Below is the documentation in pure Markdown format:
+
+### Accelerate Launch Command Documentation
+
+#### Description:
+The Accelerate launch command is used to train a model using multiple GPUs and mixed precision training. It launches the training script `tutorial_train_ip-adapter.py` with specified parameters and configurations.
+
+#### Usage Example:
+
+```
+accelerate launch --mixed_precision "fp16" \
+tutorial_train_ip-adapter.py \
+--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5/" \
+--image_encoder_path="{image_encoder_path}" \
+--data_json_file="{data.json}" \
+--data_root_path="{image_path}" \
+--mixed_precision="fp16" \
+--resolution=512 \
+--train_batch_size=8 \
+--dataloader_num_workers=4 \
+--learning_rate=1e-04 \
+--weight_decay=0.01 \
+--output_dir="{output_dir}" \
+--save_steps=10000
+```
+
+### Multi-GPU Script:
+```
+accelerate launch --num_processes 8 --multi_gpu --mixed_precision "fp16" \
+  tutorial_train_ip-adapter.py \
+  --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5/" \
+  --image_encoder_path="{image_encoder_path}" \
+  --data_json_file="{data.json}" \
+  --data_root_path="{image_path}" \
+  --mixed_precision="fp16" \
+  --resolution=512 \
+  --train_batch_size=8 \
+  --dataloader_num_workers=4 \
+  --learning_rate=1e-04 \
+  --weight_decay=0.01 \
+  --output_dir="{output_dir}" \
+  --save_steps=10000
+```
+
+#### Parameters:
+- `--num_processes`: Number of processes to launch for distributed training (in this example, 8 processes).
+- `--multi_gpu`: Flag indicating the usage of multiple GPUs for training.
+- `--mixed_precision "fp16"`: Enables mixed precision training with 16-bit floating-point precision.
+- `tutorial_train_ip-adapter.py`: Name of the training script to be executed.
+- `--pretrained_model_name_or_path`: Path or identifier for a pretrained model.
+- `--image_encoder_path`: Path to the CLIP image encoder.
+- `--data_json_file`: Path to the training data in JSON format.
+- `--data_root_path`: Root path where training images are located.
+- `--resolution`: Resolution of input images (512x512 in this example).
+- `--train_batch_size`: Batch size for training data (8 in this example).
+- `--dataloader_num_workers`: Number of subprocesses for data loading (4 in this example).
+- `--learning_rate`: Learning rate for training (1e-04 in this example).
+- `--weight_decay`: Weight decay for regularization (0.01 in this example).
+- `--output_dir`: Directory to save model checkpoints and predictions.
+- `--save_steps`: Frequency of saving checkpoints during training (10000 in this example).
+
+### Inference
+
+#### Description:
+The provided inference code is used to load a trained model checkpoint and extract the components related to image projection and IP (Image Processing) adapter. These components are then saved into a binary file for later use in inference.
+
+#### Usage Example:
+```python
+from safetensors.torch import load_file, save_file
+
+# Load the trained model checkpoint in safetensors format
+ckpt = "checkpoint-50000/pytorch_model.safetensors"
+sd = load_file(ckpt)  # Using safetensors load function
+
+# Extract image projection and IP adapter components
+image_proj_sd = {}
+ip_sd = {}
+
+for k in sd:
+    if k.startswith("unet"):
+        pass  # Skip unet-related keys
+    elif k.startswith("image_proj_model"):
+        image_proj_sd[k.replace("image_proj_model.", "")] = sd[k]
+    elif k.startswith("adapter_modules"):
+        ip_sd[k.replace("adapter_modules.", "")] = sd[k]
+
+# Save the components into separate safetensors files
+save_file(image_proj_sd, "image_proj.safetensors")
+save_file(ip_sd, "ip_adapter.safetensors")
+```
+
+### Sample Inference Script using the CLIP Model
+
+```python
+
+import torch
+from safetensors.torch import load_file
+from transformers import CLIPProcessor, CLIPModel  # Using the Hugging Face CLIP model 
+
+# Load model components from safetensors
+image_proj_ckpt = "image_proj.safetensors"
+ip_adapter_ckpt = "ip_adapter.safetensors"
+
+# Load the saved weights
+image_proj_sd = load_file(image_proj_ckpt)
+ip_adapter_sd = load_file(ip_adapter_ckpt)
+
+# Define the model Parameters
+class ImageProjectionModel(torch.nn.Module):
+    def __init__(self, input_dim=768, output_dim=512):  # CLIP's default embedding size is 768
+        super().__init__()
+        self.model = torch.nn.Linear(input_dim, output_dim)
+
+    def forward(self, x):
+        return self.model(x)
+
+class IPAdapterModel(torch.nn.Module):
+    def __init__(self, input_dim=512, output_dim=10):  # Example for 10 classes
+        super().__init__()
+        self.model = torch.nn.Linear(input_dim, output_dim)
+
+    def forward(self, x):
+        return self.model(x)
+
+# Initialize models
+image_proj_model = ImageProjectionModel()
+ip_adapter_model = IPAdapterModel()
+
+# Load weights into models
+image_proj_model.load_state_dict(image_proj_sd)
+ip_adapter_model.load_state_dict(ip_adapter_sd)
+
+# Set models to evaluation mode
+image_proj_model.eval()
+ip_adapter_model.eval()
+
+#Inference pipeline
+def inference(image_tensor):
+    """
+    Run inference using the loaded models.
+
+    Args:
+        image_tensor: Preprocessed image tensor from CLIPProcessor
+
+    Returns:
+        Final inference results
+    """
+    with torch.no_grad():
+        # Step 1: Project the image features
+        image_proj = image_proj_model(image_tensor)
+
+        # Step 2: Pass the projected features through the IP Adapter
+        result = ip_adapter_model(image_proj)
+
+    return result
+
+# Using CLIP for image preprocessing
+processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+
+#Image file path
+image_path = "path/to/image.jpg"
+
+# Preprocess the image
+inputs = processor(images=image_path, return_tensors="pt")
+image_features = clip_model.get_image_features(inputs["pixel_values"])
+
+# Normalize the image features as per CLIP's recommendations
+image_features = image_features / image_features.norm(dim=-1, keepdim=True)
+
+# Run inference
+output = inference(image_features)
+print("Inference output:", output)
+```
+
+#### Parameters:
+- `ckpt`: Path to the trained model checkpoint file.
+- `map_location="cpu"`: Specifies that the model should be loaded onto the CPU.
+- `image_proj_sd`: Dictionary to store the components related to image projection.
+- `ip_sd`: Dictionary to store the components related to the IP adapter.
+- `"unet"`, `"image_proj_model"`, `"adapter_modules"`: Prefixes indicating components of the model.

examples/research_projects/ip_adapter/requirements.txt

@@ -0,0 +1,4 @@
+accelerate
+torchvision
+transformers>=4.25.1
+ip_adapter

examples/research_projects/ip_adapter/tutorial_train_faceid.py

@@ -0,0 +1,415 @@
+import argparse
+import itertools
+import json
+import os
+import random
+import time
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+from accelerate import Accelerator
+from accelerate.utils import ProjectConfiguration
+from ip_adapter.attention_processor_faceid import LoRAAttnProcessor, LoRAIPAttnProcessor
+from ip_adapter.ip_adapter_faceid import MLPProjModel
+from PIL import Image
+from torchvision import transforms
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
+
+
+# Dataset
+class MyDataset(torch.utils.data.Dataset):
+    def __init__(
+        self, json_file, tokenizer, size=512, t_drop_rate=0.05, i_drop_rate=0.05, ti_drop_rate=0.05, image_root_path=""
+    ):
+        super().__init__()
+
+        self.tokenizer = tokenizer
+        self.size = size
+        self.i_drop_rate = i_drop_rate
+        self.t_drop_rate = t_drop_rate
+        self.ti_drop_rate = ti_drop_rate
+        self.image_root_path = image_root_path
+
+        self.data = json.load(
+            open(json_file)
+        )  # list of dict: [{"image_file": "1.png", "id_embed_file": "faceid.bin"}]
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(self.size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop(self.size),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+    def __getitem__(self, idx):
+        item = self.data[idx]
+        text = item["text"]
+        image_file = item["image_file"]
+
+        # read image
+        raw_image = Image.open(os.path.join(self.image_root_path, image_file))
+        image = self.transform(raw_image.convert("RGB"))
+
+        face_id_embed = torch.load(item["id_embed_file"], map_location="cpu")
+        face_id_embed = torch.from_numpy(face_id_embed)
+
+        # drop
+        drop_image_embed = 0
+        rand_num = random.random()
+        if rand_num < self.i_drop_rate:
+            drop_image_embed = 1
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate):
+            text = ""
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate + self.ti_drop_rate):
+            text = ""
+            drop_image_embed = 1
+        if drop_image_embed:
+            face_id_embed = torch.zeros_like(face_id_embed)
+        # get text and tokenize
+        text_input_ids = self.tokenizer(
+            text,
+            max_length=self.tokenizer.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+
+        return {
+            "image": image,
+            "text_input_ids": text_input_ids,
+            "face_id_embed": face_id_embed,
+            "drop_image_embed": drop_image_embed,
+        }
+
+    def __len__(self):
+        return len(self.data)
+
+
+def collate_fn(data):
+    images = torch.stack([example["image"] for example in data])
+    text_input_ids = torch.cat([example["text_input_ids"] for example in data], dim=0)
+    face_id_embed = torch.stack([example["face_id_embed"] for example in data])
+    drop_image_embeds = [example["drop_image_embed"] for example in data]
+
+    return {
+        "images": images,
+        "text_input_ids": text_input_ids,
+        "face_id_embed": face_id_embed,
+        "drop_image_embeds": drop_image_embeds,
+    }
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj_model = image_proj_model
+        self.adapter_modules = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj_model(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj_model.load_state_dict(state_dict["image_proj"], strict=True)
+        self.adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+        print(f"Successfully loaded weights from checkpoint {ckpt_path}")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_ip_adapter_path",
+        type=str,
+        default=None,
+        help="Path to pretrained ip adapter model. If not specified weights are initialized randomly.",
+    )
+    parser.add_argument(
+        "--data_json_file",
+        type=str,
+        default=None,
+        required=True,
+        help="Training data",
+    )
+    parser.add_argument(
+        "--data_root_path",
+        type=str,
+        default="",
+        required=True,
+        help="Training data root path",
+    )
+    parser.add_argument(
+        "--image_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to CLIP image encoder",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-ip_adapter",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=("The resolution for input images"),
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Learning rate to use.",
+    )
+    parser.add_argument("--weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--train_batch_size", type=int, default=8, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--save_steps",
+        type=int,
+        default=2000,
+        help=("Save a checkpoint of the training state every X updates"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    return args
+
+
+def main():
+    args = parse_args()
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    accelerator = Accelerator(
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+    )
+
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    # image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+    # freeze parameters of models to save more memory
+    unet.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    # image_encoder.requires_grad_(False)
+
+    # ip-adapter
+    image_proj_model = MLPProjModel(
+        cross_attention_dim=unet.config.cross_attention_dim,
+        id_embeddings_dim=512,
+        num_tokens=4,
+    )
+    # init adapter modules
+    lora_rank = 128
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            attn_procs[name] = LoRAAttnProcessor(
+                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=lora_rank
+            )
+        else:
+            layer_name = name.split(".processor")[0]
+            weights = {
+                "to_k_ip.weight": unet_sd[layer_name + ".to_k.weight"],
+                "to_v_ip.weight": unet_sd[layer_name + ".to_v.weight"],
+            }
+            attn_procs[name] = LoRAIPAttnProcessor(
+                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=lora_rank
+            )
+            attn_procs[name].load_state_dict(weights, strict=False)
+    unet.set_attn_processor(attn_procs)
+    adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+
+    ip_adapter = IPAdapter(unet, image_proj_model, adapter_modules, args.pretrained_ip_adapter_path)
+
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+    # unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    # image_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # optimizer
+    params_to_opt = itertools.chain(ip_adapter.image_proj_model.parameters(), ip_adapter.adapter_modules.parameters())
+    optimizer = torch.optim.AdamW(params_to_opt, lr=args.learning_rate, weight_decay=args.weight_decay)
+
+    # dataloader
+    train_dataset = MyDataset(
+        args.data_json_file, tokenizer=tokenizer, size=args.resolution, image_root_path=args.data_root_path
+    )
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Prepare everything with our `accelerator`.
+    ip_adapter, optimizer, train_dataloader = accelerator.prepare(ip_adapter, optimizer, train_dataloader)
+
+    global_step = 0
+    for epoch in range(0, args.num_train_epochs):
+        begin = time.perf_counter()
+        for step, batch in enumerate(train_dataloader):
+            load_data_time = time.perf_counter() - begin
+            with accelerator.accumulate(ip_adapter):
+                # Convert images to latent space
+                with torch.no_grad():
+                    latents = vae.encode(
+                        batch["images"].to(accelerator.device, dtype=weight_dtype)
+                    ).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                image_embeds = batch["face_id_embed"].to(accelerator.device, dtype=weight_dtype)
+
+                with torch.no_grad():
+                    encoder_hidden_states = text_encoder(batch["text_input_ids"].to(accelerator.device))[0]
+
+                noise_pred = ip_adapter(noisy_latents, timesteps, encoder_hidden_states, image_embeds)
+
+                loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean().item()
+
+                # Backpropagate
+                accelerator.backward(loss)
+                optimizer.step()
+                optimizer.zero_grad()
+
+                if accelerator.is_main_process:
+                    print(
+                        "Epoch {}, step {}, data_time: {}, time: {}, step_loss: {}".format(
+                            epoch, step, load_data_time, time.perf_counter() - begin, avg_loss
+                        )
+                    )
+
+            global_step += 1
+
+            if global_step % args.save_steps == 0:
+                save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                accelerator.save_state(save_path)
+
+            begin = time.perf_counter()
+
+
+if __name__ == "__main__":
+    main()

examples/research_projects/ip_adapter/tutorial_train_ip-adapter.py

@@ -0,0 +1,422 @@
+import argparse
+import itertools
+import json
+import os
+import random
+import time
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+from accelerate import Accelerator
+from accelerate.utils import ProjectConfiguration
+from ip_adapter.ip_adapter import ImageProjModel
+from ip_adapter.utils import is_torch2_available
+from PIL import Image
+from torchvision import transforms
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
+
+
+if is_torch2_available():
+    from ip_adapter.attention_processor import AttnProcessor2_0 as AttnProcessor
+    from ip_adapter.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor
+else:
+    from ip_adapter.attention_processor import AttnProcessor, IPAttnProcessor
+
+
+# Dataset
+class MyDataset(torch.utils.data.Dataset):
+    def __init__(
+        self, json_file, tokenizer, size=512, t_drop_rate=0.05, i_drop_rate=0.05, ti_drop_rate=0.05, image_root_path=""
+    ):
+        super().__init__()
+
+        self.tokenizer = tokenizer
+        self.size = size
+        self.i_drop_rate = i_drop_rate
+        self.t_drop_rate = t_drop_rate
+        self.ti_drop_rate = ti_drop_rate
+        self.image_root_path = image_root_path
+
+        self.data = json.load(open(json_file))  # list of dict: [{"image_file": "1.png", "text": "A dog"}]
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(self.size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop(self.size),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+        self.clip_image_processor = CLIPImageProcessor()
+
+    def __getitem__(self, idx):
+        item = self.data[idx]
+        text = item["text"]
+        image_file = item["image_file"]
+
+        # read image
+        raw_image = Image.open(os.path.join(self.image_root_path, image_file))
+        image = self.transform(raw_image.convert("RGB"))
+        clip_image = self.clip_image_processor(images=raw_image, return_tensors="pt").pixel_values
+
+        # drop
+        drop_image_embed = 0
+        rand_num = random.random()
+        if rand_num < self.i_drop_rate:
+            drop_image_embed = 1
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate):
+            text = ""
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate + self.ti_drop_rate):
+            text = ""
+            drop_image_embed = 1
+        # get text and tokenize
+        text_input_ids = self.tokenizer(
+            text,
+            max_length=self.tokenizer.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+
+        return {
+            "image": image,
+            "text_input_ids": text_input_ids,
+            "clip_image": clip_image,
+            "drop_image_embed": drop_image_embed,
+        }
+
+    def __len__(self):
+        return len(self.data)
+
+
+def collate_fn(data):
+    images = torch.stack([example["image"] for example in data])
+    text_input_ids = torch.cat([example["text_input_ids"] for example in data], dim=0)
+    clip_images = torch.cat([example["clip_image"] for example in data], dim=0)
+    drop_image_embeds = [example["drop_image_embed"] for example in data]
+
+    return {
+        "images": images,
+        "text_input_ids": text_input_ids,
+        "clip_images": clip_images,
+        "drop_image_embeds": drop_image_embeds,
+    }
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj_model = image_proj_model
+        self.adapter_modules = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj_model(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj_model.load_state_dict(state_dict["image_proj"], strict=True)
+        self.adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+        print(f"Successfully loaded weights from checkpoint {ckpt_path}")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_ip_adapter_path",
+        type=str,
+        default=None,
+        help="Path to pretrained ip adapter model. If not specified weights are initialized randomly.",
+    )
+    parser.add_argument(
+        "--data_json_file",
+        type=str,
+        default=None,
+        required=True,
+        help="Training data",
+    )
+    parser.add_argument(
+        "--data_root_path",
+        type=str,
+        default="",
+        required=True,
+        help="Training data root path",
+    )
+    parser.add_argument(
+        "--image_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to CLIP image encoder",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-ip_adapter",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=("The resolution for input images"),
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Learning rate to use.",
+    )
+    parser.add_argument("--weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--train_batch_size", type=int, default=8, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--save_steps",
+        type=int,
+        default=2000,
+        help=("Save a checkpoint of the training state every X updates"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    return args
+
+
+def main():
+    args = parse_args()
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    accelerator = Accelerator(
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+    )
+
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+    # freeze parameters of models to save more memory
+    unet.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    image_encoder.requires_grad_(False)
+
+    # ip-adapter
+    image_proj_model = ImageProjModel(
+        cross_attention_dim=unet.config.cross_attention_dim,
+        clip_embeddings_dim=image_encoder.config.projection_dim,
+        clip_extra_context_tokens=4,
+    )
+    # init adapter modules
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            attn_procs[name] = AttnProcessor()
+        else:
+            layer_name = name.split(".processor")[0]
+            weights = {
+                "to_k_ip.weight": unet_sd[layer_name + ".to_k.weight"],
+                "to_v_ip.weight": unet_sd[layer_name + ".to_v.weight"],
+            }
+            attn_procs[name] = IPAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
+            attn_procs[name].load_state_dict(weights)
+    unet.set_attn_processor(attn_procs)
+    adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+
+    ip_adapter = IPAdapter(unet, image_proj_model, adapter_modules, args.pretrained_ip_adapter_path)
+
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+    # unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    image_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # optimizer
+    params_to_opt = itertools.chain(ip_adapter.image_proj_model.parameters(), ip_adapter.adapter_modules.parameters())
+    optimizer = torch.optim.AdamW(params_to_opt, lr=args.learning_rate, weight_decay=args.weight_decay)
+
+    # dataloader
+    train_dataset = MyDataset(
+        args.data_json_file, tokenizer=tokenizer, size=args.resolution, image_root_path=args.data_root_path
+    )
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Prepare everything with our `accelerator`.
+    ip_adapter, optimizer, train_dataloader = accelerator.prepare(ip_adapter, optimizer, train_dataloader)
+
+    global_step = 0
+    for epoch in range(0, args.num_train_epochs):
+        begin = time.perf_counter()
+        for step, batch in enumerate(train_dataloader):
+            load_data_time = time.perf_counter() - begin
+            with accelerator.accumulate(ip_adapter):
+                # Convert images to latent space
+                with torch.no_grad():
+                    latents = vae.encode(
+                        batch["images"].to(accelerator.device, dtype=weight_dtype)
+                    ).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                with torch.no_grad():
+                    image_embeds = image_encoder(
+                        batch["clip_images"].to(accelerator.device, dtype=weight_dtype)
+                    ).image_embeds
+                image_embeds_ = []
+                for image_embed, drop_image_embed in zip(image_embeds, batch["drop_image_embeds"]):
+                    if drop_image_embed == 1:
+                        image_embeds_.append(torch.zeros_like(image_embed))
+                    else:
+                        image_embeds_.append(image_embed)
+                image_embeds = torch.stack(image_embeds_)
+
+                with torch.no_grad():
+                    encoder_hidden_states = text_encoder(batch["text_input_ids"].to(accelerator.device))[0]
+
+                noise_pred = ip_adapter(noisy_latents, timesteps, encoder_hidden_states, image_embeds)
+
+                loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean().item()
+
+                # Backpropagate
+                accelerator.backward(loss)
+                optimizer.step()
+                optimizer.zero_grad()
+
+                if accelerator.is_main_process:
+                    print(
+                        "Epoch {}, step {}, data_time: {}, time: {}, step_loss: {}".format(
+                            epoch, step, load_data_time, time.perf_counter() - begin, avg_loss
+                        )
+                    )
+
+            global_step += 1
+
+            if global_step % args.save_steps == 0:
+                save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                accelerator.save_state(save_path)
+
+            begin = time.perf_counter()
+
+
+if __name__ == "__main__":
+    main()

examples/research_projects/ip_adapter/tutorial_train_plus.py

@@ -0,0 +1,445 @@
+import argparse
+import itertools
+import json
+import os
+import random
+import time
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+from accelerate import Accelerator
+from accelerate.utils import ProjectConfiguration
+from ip_adapter.resampler import Resampler
+from ip_adapter.utils import is_torch2_available
+from PIL import Image
+from torchvision import transforms
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
+
+
+if is_torch2_available():
+    from ip_adapter.attention_processor import AttnProcessor2_0 as AttnProcessor
+    from ip_adapter.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor
+else:
+    from ip_adapter.attention_processor import AttnProcessor, IPAttnProcessor
+
+
+# Dataset
+class MyDataset(torch.utils.data.Dataset):
+    def __init__(
+        self, json_file, tokenizer, size=512, t_drop_rate=0.05, i_drop_rate=0.05, ti_drop_rate=0.05, image_root_path=""
+    ):
+        super().__init__()
+
+        self.tokenizer = tokenizer
+        self.size = size
+        self.i_drop_rate = i_drop_rate
+        self.t_drop_rate = t_drop_rate
+        self.ti_drop_rate = ti_drop_rate
+        self.image_root_path = image_root_path
+
+        self.data = json.load(open(json_file))  # list of dict: [{"image_file": "1.png", "text": "A dog"}]
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(self.size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop(self.size),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+        self.clip_image_processor = CLIPImageProcessor()
+
+    def __getitem__(self, idx):
+        item = self.data[idx]
+        text = item["text"]
+        image_file = item["image_file"]
+
+        # read image
+        raw_image = Image.open(os.path.join(self.image_root_path, image_file))
+        image = self.transform(raw_image.convert("RGB"))
+        clip_image = self.clip_image_processor(images=raw_image, return_tensors="pt").pixel_values
+
+        # drop
+        drop_image_embed = 0
+        rand_num = random.random()
+        if rand_num < self.i_drop_rate:
+            drop_image_embed = 1
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate):
+            text = ""
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate + self.ti_drop_rate):
+            text = ""
+            drop_image_embed = 1
+        # get text and tokenize
+        text_input_ids = self.tokenizer(
+            text,
+            max_length=self.tokenizer.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+
+        return {
+            "image": image,
+            "text_input_ids": text_input_ids,
+            "clip_image": clip_image,
+            "drop_image_embed": drop_image_embed,
+        }
+
+    def __len__(self):
+        return len(self.data)
+
+
+def collate_fn(data):
+    images = torch.stack([example["image"] for example in data])
+    text_input_ids = torch.cat([example["text_input_ids"] for example in data], dim=0)
+    clip_images = torch.cat([example["clip_image"] for example in data], dim=0)
+    drop_image_embeds = [example["drop_image_embed"] for example in data]
+
+    return {
+        "images": images,
+        "text_input_ids": text_input_ids,
+        "clip_images": clip_images,
+        "drop_image_embeds": drop_image_embeds,
+    }
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj_model = image_proj_model
+        self.adapter_modules = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj_model(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        # Check if 'latents' exists in both the saved state_dict and the current model's state_dict
+        strict_load_image_proj_model = True
+        if "latents" in state_dict["image_proj"] and "latents" in self.image_proj_model.state_dict():
+            # Check if the shapes are mismatched
+            if state_dict["image_proj"]["latents"].shape != self.image_proj_model.state_dict()["latents"].shape:
+                print(f"Shapes of 'image_proj.latents' in checkpoint {ckpt_path} and current model do not match.")
+                print("Removing 'latents' from checkpoint and loading the rest of the weights.")
+                del state_dict["image_proj"]["latents"]
+                strict_load_image_proj_model = False
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj_model.load_state_dict(state_dict["image_proj"], strict=strict_load_image_proj_model)
+        self.adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+        print(f"Successfully loaded weights from checkpoint {ckpt_path}")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_ip_adapter_path",
+        type=str,
+        default=None,
+        help="Path to pretrained ip adapter model. If not specified weights are initialized randomly.",
+    )
+    parser.add_argument(
+        "--num_tokens",
+        type=int,
+        default=16,
+        help="Number of tokens to query from the CLIP image encoding.",
+    )
+    parser.add_argument(
+        "--data_json_file",
+        type=str,
+        default=None,
+        required=True,
+        help="Training data",
+    )
+    parser.add_argument(
+        "--data_root_path",
+        type=str,
+        default="",
+        required=True,
+        help="Training data root path",
+    )
+    parser.add_argument(
+        "--image_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to CLIP image encoder",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-ip_adapter",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=("The resolution for input images"),
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Learning rate to use.",
+    )
+    parser.add_argument("--weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--train_batch_size", type=int, default=8, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--save_steps",
+        type=int,
+        default=2000,
+        help=("Save a checkpoint of the training state every X updates"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    return args
+
+
+def main():
+    args = parse_args()
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    accelerator = Accelerator(
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+    )
+
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+    # freeze parameters of models to save more memory
+    unet.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    image_encoder.requires_grad_(False)
+
+    # ip-adapter-plus
+    image_proj_model = Resampler(
+        dim=unet.config.cross_attention_dim,
+        depth=4,
+        dim_head=64,
+        heads=12,
+        num_queries=args.num_tokens,
+        embedding_dim=image_encoder.config.hidden_size,
+        output_dim=unet.config.cross_attention_dim,
+        ff_mult=4,
+    )
+    # init adapter modules
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            attn_procs[name] = AttnProcessor()
+        else:
+            layer_name = name.split(".processor")[0]
+            weights = {
+                "to_k_ip.weight": unet_sd[layer_name + ".to_k.weight"],
+                "to_v_ip.weight": unet_sd[layer_name + ".to_v.weight"],
+            }
+            attn_procs[name] = IPAttnProcessor(
+                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, num_tokens=args.num_tokens
+            )
+            attn_procs[name].load_state_dict(weights)
+    unet.set_attn_processor(attn_procs)
+    adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+
+    ip_adapter = IPAdapter(unet, image_proj_model, adapter_modules, args.pretrained_ip_adapter_path)
+
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+    # unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    image_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # optimizer
+    params_to_opt = itertools.chain(ip_adapter.image_proj_model.parameters(), ip_adapter.adapter_modules.parameters())
+    optimizer = torch.optim.AdamW(params_to_opt, lr=args.learning_rate, weight_decay=args.weight_decay)
+
+    # dataloader
+    train_dataset = MyDataset(
+        args.data_json_file, tokenizer=tokenizer, size=args.resolution, image_root_path=args.data_root_path
+    )
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Prepare everything with our `accelerator`.
+    ip_adapter, optimizer, train_dataloader = accelerator.prepare(ip_adapter, optimizer, train_dataloader)
+
+    global_step = 0
+    for epoch in range(0, args.num_train_epochs):
+        begin = time.perf_counter()
+        for step, batch in enumerate(train_dataloader):
+            load_data_time = time.perf_counter() - begin
+            with accelerator.accumulate(ip_adapter):
+                # Convert images to latent space
+                with torch.no_grad():
+                    latents = vae.encode(
+                        batch["images"].to(accelerator.device, dtype=weight_dtype)
+                    ).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                clip_images = []
+                for clip_image, drop_image_embed in zip(batch["clip_images"], batch["drop_image_embeds"]):
+                    if drop_image_embed == 1:
+                        clip_images.append(torch.zeros_like(clip_image))
+                    else:
+                        clip_images.append(clip_image)
+                clip_images = torch.stack(clip_images, dim=0)
+                with torch.no_grad():
+                    image_embeds = image_encoder(
+                        clip_images.to(accelerator.device, dtype=weight_dtype), output_hidden_states=True
+                    ).hidden_states[-2]
+
+                with torch.no_grad():
+                    encoder_hidden_states = text_encoder(batch["text_input_ids"].to(accelerator.device))[0]
+
+                noise_pred = ip_adapter(noisy_latents, timesteps, encoder_hidden_states, image_embeds)
+
+                loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean().item()
+
+                # Backpropagate
+                accelerator.backward(loss)
+                optimizer.step()
+                optimizer.zero_grad()
+
+                if accelerator.is_main_process:
+                    print(
+                        "Epoch {}, step {}, data_time: {}, time: {}, step_loss: {}".format(
+                            epoch, step, load_data_time, time.perf_counter() - begin, avg_loss
+                        )
+                    )
+
+            global_step += 1
+
+            if global_step % args.save_steps == 0:
+                save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                accelerator.save_state(save_path)
+
+            begin = time.perf_counter()
+
+
+if __name__ == "__main__":
+    main()

examples/research_projects/ip_adapter/tutorial_train_sdxl.py

@@ -0,0 +1,520 @@
+import argparse
+import itertools
+import json
+import os
+import random
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from accelerate import Accelerator
+from accelerate.utils import ProjectConfiguration
+from ip_adapter.ip_adapter import ImageProjModel
+from ip_adapter.utils import is_torch2_available
+from PIL import Image
+from torchvision import transforms
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
+
+
+if is_torch2_available():
+    from ip_adapter.attention_processor import AttnProcessor2_0 as AttnProcessor
+    from ip_adapter.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor
+else:
+    from ip_adapter.attention_processor import AttnProcessor, IPAttnProcessor
+
+
+# Dataset
+class MyDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        json_file,
+        tokenizer,
+        tokenizer_2,
+        size=1024,
+        center_crop=True,
+        t_drop_rate=0.05,
+        i_drop_rate=0.05,
+        ti_drop_rate=0.05,
+        image_root_path="",
+    ):
+        super().__init__()
+
+        self.tokenizer = tokenizer
+        self.tokenizer_2 = tokenizer_2
+        self.size = size
+        self.center_crop = center_crop
+        self.i_drop_rate = i_drop_rate
+        self.t_drop_rate = t_drop_rate
+        self.ti_drop_rate = ti_drop_rate
+        self.image_root_path = image_root_path
+
+        self.data = json.load(open(json_file))  # list of dict: [{"image_file": "1.png", "text": "A dog"}]
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(self.size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+        self.clip_image_processor = CLIPImageProcessor()
+
+    def __getitem__(self, idx):
+        item = self.data[idx]
+        text = item["text"]
+        image_file = item["image_file"]
+
+        # read image
+        raw_image = Image.open(os.path.join(self.image_root_path, image_file))
+
+        # original size
+        original_width, original_height = raw_image.size
+        original_size = torch.tensor([original_height, original_width])
+
+        image_tensor = self.transform(raw_image.convert("RGB"))
+        # random crop
+        delta_h = image_tensor.shape[1] - self.size
+        delta_w = image_tensor.shape[2] - self.size
+        assert not all([delta_h, delta_w])
+
+        if self.center_crop:
+            top = delta_h // 2
+            left = delta_w // 2
+        else:
+            top = np.random.randint(0, delta_h + 1)
+            left = np.random.randint(0, delta_w + 1)
+        image = transforms.functional.crop(image_tensor, top=top, left=left, height=self.size, width=self.size)
+        crop_coords_top_left = torch.tensor([top, left])
+
+        clip_image = self.clip_image_processor(images=raw_image, return_tensors="pt").pixel_values
+
+        # drop
+        drop_image_embed = 0
+        rand_num = random.random()
+        if rand_num < self.i_drop_rate:
+            drop_image_embed = 1
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate):
+            text = ""
+        elif rand_num < (self.i_drop_rate + self.t_drop_rate + self.ti_drop_rate):
+            text = ""
+            drop_image_embed = 1
+
+        # get text and tokenize
+        text_input_ids = self.tokenizer(
+            text,
+            max_length=self.tokenizer.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+
+        text_input_ids_2 = self.tokenizer_2(
+            text,
+            max_length=self.tokenizer_2.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+
+        return {
+            "image": image,
+            "text_input_ids": text_input_ids,
+            "text_input_ids_2": text_input_ids_2,
+            "clip_image": clip_image,
+            "drop_image_embed": drop_image_embed,
+            "original_size": original_size,
+            "crop_coords_top_left": crop_coords_top_left,
+            "target_size": torch.tensor([self.size, self.size]),
+        }
+
+    def __len__(self):
+        return len(self.data)
+
+
+def collate_fn(data):
+    images = torch.stack([example["image"] for example in data])
+    text_input_ids = torch.cat([example["text_input_ids"] for example in data], dim=0)
+    text_input_ids_2 = torch.cat([example["text_input_ids_2"] for example in data], dim=0)
+    clip_images = torch.cat([example["clip_image"] for example in data], dim=0)
+    drop_image_embeds = [example["drop_image_embed"] for example in data]
+    original_size = torch.stack([example["original_size"] for example in data])
+    crop_coords_top_left = torch.stack([example["crop_coords_top_left"] for example in data])
+    target_size = torch.stack([example["target_size"] for example in data])
+
+    return {
+        "images": images,
+        "text_input_ids": text_input_ids,
+        "text_input_ids_2": text_input_ids_2,
+        "clip_images": clip_images,
+        "drop_image_embeds": drop_image_embeds,
+        "original_size": original_size,
+        "crop_coords_top_left": crop_coords_top_left,
+        "target_size": target_size,
+    }
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj_model = image_proj_model
+        self.adapter_modules = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, unet_added_cond_kwargs, image_embeds):
+        ip_tokens = self.image_proj_model(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(
+            noisy_latents, timesteps, encoder_hidden_states, added_cond_kwargs=unet_added_cond_kwargs
+        ).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj_model.load_state_dict(state_dict["image_proj"], strict=True)
+        self.adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj_model.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.adapter_modules.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+        print(f"Successfully loaded weights from checkpoint {ckpt_path}")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_ip_adapter_path",
+        type=str,
+        default=None,
+        help="Path to pretrained ip adapter model. If not specified weights are initialized randomly.",
+    )
+    parser.add_argument(
+        "--data_json_file",
+        type=str,
+        default=None,
+        required=True,
+        help="Training data",
+    )
+    parser.add_argument(
+        "--data_root_path",
+        type=str,
+        default="",
+        required=True,
+        help="Training data root path",
+    )
+    parser.add_argument(
+        "--image_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to CLIP image encoder",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-ip_adapter",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=("The resolution for input images"),
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Learning rate to use.",
+    )
+    parser.add_argument("--weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--train_batch_size", type=int, default=8, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--noise_offset", type=float, default=None, help="noise offset")
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--save_steps",
+        type=int,
+        default=2000,
+        help=("Save a checkpoint of the training state every X updates"),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    return args
+
+
+def main():
+    args = parse_args()
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    accelerator = Accelerator(
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+    )
+
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
+    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
+    tokenizer_2 = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer_2")
+    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder_2"
+    )
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+    # freeze parameters of models to save more memory
+    unet.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    text_encoder_2.requires_grad_(False)
+    image_encoder.requires_grad_(False)
+
+    # ip-adapter
+    num_tokens = 4
+    image_proj_model = ImageProjModel(
+        cross_attention_dim=unet.config.cross_attention_dim,
+        clip_embeddings_dim=image_encoder.config.projection_dim,
+        clip_extra_context_tokens=num_tokens,
+    )
+    # init adapter modules
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            attn_procs[name] = AttnProcessor()
+        else:
+            layer_name = name.split(".processor")[0]
+            weights = {
+                "to_k_ip.weight": unet_sd[layer_name + ".to_k.weight"],
+                "to_v_ip.weight": unet_sd[layer_name + ".to_v.weight"],
+            }
+            attn_procs[name] = IPAttnProcessor(
+                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, num_tokens=num_tokens
+            )
+            attn_procs[name].load_state_dict(weights)
+    unet.set_attn_processor(attn_procs)
+    adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+
+    ip_adapter = IPAdapter(unet, image_proj_model, adapter_modules, args.pretrained_ip_adapter_path)
+
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+    # unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device)  # use fp32
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+    text_encoder_2.to(accelerator.device, dtype=weight_dtype)
+    image_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # optimizer
+    params_to_opt = itertools.chain(ip_adapter.image_proj_model.parameters(), ip_adapter.adapter_modules.parameters())
+    optimizer = torch.optim.AdamW(params_to_opt, lr=args.learning_rate, weight_decay=args.weight_decay)
+
+    # dataloader
+    train_dataset = MyDataset(
+        args.data_json_file,
+        tokenizer=tokenizer,
+        tokenizer_2=tokenizer_2,
+        size=args.resolution,
+        image_root_path=args.data_root_path,
+    )
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Prepare everything with our `accelerator`.
+    ip_adapter, optimizer, train_dataloader = accelerator.prepare(ip_adapter, optimizer, train_dataloader)
+
+    global_step = 0
+    for epoch in range(0, args.num_train_epochs):
+        begin = time.perf_counter()
+        for step, batch in enumerate(train_dataloader):
+            load_data_time = time.perf_counter() - begin
+            with accelerator.accumulate(ip_adapter):
+                # Convert images to latent space
+                with torch.no_grad():
+                    # vae of sdxl should use fp32
+                    latents = vae.encode(batch["images"].to(accelerator.device, dtype=vae.dtype)).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+                    latents = latents.to(accelerator.device, dtype=weight_dtype)
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                if args.noise_offset:
+                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
+                    noise += args.noise_offset * torch.randn((latents.shape[0], latents.shape[1], 1, 1)).to(
+                        accelerator.device, dtype=weight_dtype
+                    )
+
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                with torch.no_grad():
+                    image_embeds = image_encoder(
+                        batch["clip_images"].to(accelerator.device, dtype=weight_dtype)
+                    ).image_embeds
+                image_embeds_ = []
+                for image_embed, drop_image_embed in zip(image_embeds, batch["drop_image_embeds"]):
+                    if drop_image_embed == 1:
+                        image_embeds_.append(torch.zeros_like(image_embed))
+                    else:
+                        image_embeds_.append(image_embed)
+                image_embeds = torch.stack(image_embeds_)
+
+                with torch.no_grad():
+                    encoder_output = text_encoder(
+                        batch["text_input_ids"].to(accelerator.device), output_hidden_states=True
+                    )
+                    text_embeds = encoder_output.hidden_states[-2]
+                    encoder_output_2 = text_encoder_2(
+                        batch["text_input_ids_2"].to(accelerator.device), output_hidden_states=True
+                    )
+                    pooled_text_embeds = encoder_output_2[0]
+                    text_embeds_2 = encoder_output_2.hidden_states[-2]
+                    text_embeds = torch.concat([text_embeds, text_embeds_2], dim=-1)  # concat
+
+                # add cond
+                add_time_ids = [
+                    batch["original_size"].to(accelerator.device),
+                    batch["crop_coords_top_left"].to(accelerator.device),
+                    batch["target_size"].to(accelerator.device),
+                ]
+                add_time_ids = torch.cat(add_time_ids, dim=1).to(accelerator.device, dtype=weight_dtype)
+                unet_added_cond_kwargs = {"text_embeds": pooled_text_embeds, "time_ids": add_time_ids}
+
+                noise_pred = ip_adapter(noisy_latents, timesteps, text_embeds, unet_added_cond_kwargs, image_embeds)
+
+                loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean().item()
+
+                # Backpropagate
+                accelerator.backward(loss)
+                optimizer.step()
+                optimizer.zero_grad()
+
+                if accelerator.is_main_process:
+                    print(
+                        "Epoch {}, step {}, data_time: {}, time: {}, step_loss: {}".format(
+                            epoch, step, load_data_time, time.perf_counter() - begin, avg_loss
+                        )
+                    )
+
+            global_step += 1
+
+            if global_step % args.save_steps == 0:
+                save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                accelerator.save_state(save_path)
+
+            begin = time.perf_counter()
+
+
+if __name__ == "__main__":
+    main()

examples/research_projects/pytorch_xla/README.md

@@ -7,13 +7,14 @@ It has been tested on v4 and v5p TPU versions. Training code has been tested on
 This script implements Distributed Data Parallel using GSPMD feature in XLA compiler
 where we shard the input batches over the TPU devices. 
 
-As of 9-11-2024, these are some expected step times.
+As of 10-31-2024, these are some expected step times.
 
 | accelerator | global batch size | step time (seconds) |
 | ----------- | ----------------- | --------- |
-| v5p-128 | 1024 | 0.245 |
-| v5p-256 | 2048 | 0.234 |
-| v5p-512 | 4096 | 0.2498 |
+| v5p-512 | 16384 | 1.01 |
+| v5p-256 | 8192 | 1.01 |
+| v5p-128 | 4096 | 1.0 |
+| v5p-64 | 2048 | 1.01 |
 
 ## Create TPU
 
@@ -43,8 +44,9 @@ Install PyTorch and PyTorch/XLA nightly versions:
 gcloud compute tpus tpu-vm ssh ${TPU_NAME} \
 --project=${PROJECT_ID} --zone=${ZONE} --worker=all \
 --command='
-pip3 install --pre torch==2.5.0.dev20240905+cpu torchvision==0.20.0.dev20240905+cpu --index-url https://download.pytorch.org/whl/nightly/cpu
-pip3 install "torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.5.0.dev20240905-cp310-cp310-linux_x86_64.whl" -f https://storage.googleapis.com/libtpu-releases/index.html
+pip3 install --pre torch==2.6.0.dev20241031+cpu torchvision --index-url https://download.pytorch.org/whl/nightly/cpu
+pip3 install "torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241031.cxx11-cp310-cp310-linux_x86_64.whl" -f https://storage.googleapis.com/libtpu-releases/index.html
+pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
 '
 ```
 
@@ -88,17 +90,18 @@ are fixed.
 gcloud compute tpus tpu-vm ssh ${TPU_NAME} \
 --project=${PROJECT_ID} --zone=${ZONE} --worker=all \
 --command='
-export XLA_DISABLE_FUNCTIONALIZATION=1 
+export XLA_DISABLE_FUNCTIONALIZATION=0
 export PROFILE_DIR=/tmp/
 export CACHE_DIR=/tmp/
 export DATASET_NAME=lambdalabs/naruto-blip-captions
 export PER_HOST_BATCH_SIZE=32 # This is known to work on TPU v4. Can set this to 64 for TPU v5p
 export TRAIN_STEPS=50
 export OUTPUT_DIR=/tmp/trained-model/
-python diffusers/examples/research_projects/pytorch_xla/train_text_to_image_xla.py --pretrained_model_name_or_path=stabilityai/stable-diffusion-2-base --dataset_name=$DATASET_NAME --resolution=512 --center_crop --random_flip --train_batch_size=$PER_HOST_BATCH_SIZE  --max_train_steps=$TRAIN_STEPS --learning_rate=1e-06 --mixed_precision=bf16 --profile_duration=80000 --output_dir=$OUTPUT_DIR --dataloader_num_workers=4 --loader_prefetch_size=4 --device_prefetch_size=4'
-   
+python diffusers/examples/research_projects/pytorch_xla/train_text_to_image_xla.py --pretrained_model_name_or_path=stabilityai/stable-diffusion-2-base --dataset_name=$DATASET_NAME --resolution=512 --center_crop --random_flip --train_batch_size=$PER_HOST_BATCH_SIZE  --max_train_steps=$TRAIN_STEPS --learning_rate=1e-06 --mixed_precision=bf16 --profile_duration=80000 --output_dir=$OUTPUT_DIR --dataloader_num_workers=8 --loader_prefetch_size=4 --device_prefetch_size=4'
 ```
 
+Pass `--print_loss` if you would like to see the loss printed at every step. Be aware that printing the loss at every step disrupts the optimized flow execution, thus the step time will be longer. 
+
 ### Environment Envs Explained
 
 *   `XLA_DISABLE_FUNCTIONALIZATION`: To optimize the performance for AdamW optimizer.

examples/research_projects/pytorch_xla/train_text_to_image_xla.py

@@ -140,33 +140,43 @@ class TrainSD:
         self.optimizer.step()
 
     def start_training(self):
-        times = []
-        last_time = time.time()
-        step = 0
-        while True:
-            if self.global_step >= self.args.max_train_steps:
-                xm.mark_step()
-                break
-            if step == 4 and PROFILE_DIR is not None:
-                xm.wait_device_ops()
-                xp.trace_detached(f"localhost:{PORT}", PROFILE_DIR, duration_ms=args.profile_duration)
+        dataloader_exception = False
+        measure_start_step = args.measure_start_step
+        assert measure_start_step < self.args.max_train_steps
+        total_time = 0
+        for step in range(0, self.args.max_train_steps):
             try:
                 batch = next(self.dataloader)
             except Exception as e:
+                dataloader_exception = True
                 print(e)
                 break
+            if step == measure_start_step and PROFILE_DIR is not None:
+                xm.wait_device_ops()
+                xp.trace_detached(f"localhost:{PORT}", PROFILE_DIR, duration_ms=args.profile_duration)
+                last_time = time.time()
             loss = self.step_fn(batch["pixel_values"], batch["input_ids"])
-            step_time = time.time() - last_time
-            if step >= 10:
-                times.append(step_time)
-            print(f"step: {step}, step_time: {step_time}")
-            if step % 5 == 0:
-                print(f"step: {step}, loss: {loss}")
-            last_time = time.time()
             self.global_step += 1
-            step += 1
-        # print(f"Average step time: {sum(times)/len(times)}")
-        xm.wait_device_ops()
+
+            def print_loss_closure(step, loss):
+                print(f"Step: {step}, Loss: {loss}")
+
+            if args.print_loss:
+                xm.add_step_closure(
+                    print_loss_closure,
+                    args=(
+                        self.global_step,
+                        loss,
+                    ),
+                )
+        xm.mark_step()
+        if not dataloader_exception:
+            xm.wait_device_ops()
+            total_time = time.time() - last_time
+            print(f"Average step time: {total_time/(self.args.max_train_steps-measure_start_step)}")
+        else:
+            print("dataloader exception happen, skip result")
+            return
 
     def step_fn(
         self,
@@ -180,7 +190,10 @@ class TrainSD:
             noise = torch.randn_like(latents).to(self.device, dtype=self.weight_dtype)
             bsz = latents.shape[0]
             timesteps = torch.randint(
-                0, self.noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
+                0,
+                self.noise_scheduler.config.num_train_timesteps,
+                (bsz,),
+                device=latents.device,
             )
             timesteps = timesteps.long()
 
@@ -224,9 +237,6 @@ class TrainSD:
 
 def parse_args():
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
-    parser.add_argument(
-        "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
-    )
     parser.add_argument("--profile_duration", type=int, default=10000, help="Profile duration in ms")
     parser.add_argument(
         "--pretrained_model_name_or_path",
@@ -258,12 +268,6 @@ def parse_args():
             " or to a folder containing files that 🤗 Datasets can understand."
         ),
     )
-    parser.add_argument(
-        "--dataset_config_name",
-        type=str,
-        default=None,
-        help="The config of the Dataset, leave as None if there's only one config.",
-    )
     parser.add_argument(
         "--train_data_dir",
         type=str,
@@ -283,15 +287,6 @@ def parse_args():
         default="text",
         help="The column of the dataset containing a caption or a list of captions.",
     )
-    parser.add_argument(
-        "--max_train_samples",
-        type=int,
-        default=None,
-        help=(
-            "For debugging purposes or quicker training, truncate the number of training examples to this "
-            "value if set."
-        ),
-    )
     parser.add_argument(
         "--output_dir",
         type=str,
@@ -304,7 +299,6 @@ def parse_args():
         default=None,
         help="The directory where the downloaded models and datasets will be stored.",
     )
-    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
     parser.add_argument(
         "--resolution",
         type=int,
@@ -374,12 +368,19 @@ def parse_args():
         default=1,
         help=("Number of subprocesses to use for data loading to cpu."),
     )
+    parser.add_argument(
+        "--loader_prefetch_factor",
+        type=int,
+        default=2,
+        help=("Number of batches loaded in advance by each worker."),
+    )
     parser.add_argument(
         "--device_prefetch_size",
         type=int,
         default=1,
         help=("Number of subprocesses to use for data loading to tpu from cpu. "),
     )
+    parser.add_argument("--measure_start_step", type=int, default=10, help="Step to start profiling.")
     parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
     parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
     parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
@@ -394,12 +395,8 @@ def parse_args():
         "--mixed_precision",
         type=str,
         default=None,
-        choices=["no", "fp16", "bf16"],
-        help=(
-            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
-            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
-            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
-        ),
+        choices=["no", "bf16"],
+        help=("Whether to use mixed precision. Bf16 requires PyTorch >= 1.10"),
     )
     parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
     parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
@@ -409,6 +406,12 @@ def parse_args():
         default=None,
         help="The name of the repository to keep in sync with the local `output_dir`.",
     )
+    parser.add_argument(
+        "--print_loss",
+        default=False,
+        action="store_true",
+        help=("Print loss at every step."),
+    )
 
     args = parser.parse_args()
 
@@ -436,7 +439,6 @@ def load_dataset(args):
         # Downloading and loading a dataset from the hub.
         dataset = datasets.load_dataset(
             args.dataset_name,
-            args.dataset_config_name,
             cache_dir=args.cache_dir,
             data_dir=args.train_data_dir,
         )
@@ -481,9 +483,7 @@ def main(args):
     _ = xp.start_server(PORT)
 
     num_devices = xr.global_runtime_device_count()
-    device_ids = np.arange(num_devices)
-    mesh_shape = (num_devices, 1)
-    mesh = xs.Mesh(device_ids, mesh_shape, ("x", "y"))
+    mesh = xs.get_1d_mesh("data")
     xs.set_global_mesh(mesh)
 
     text_encoder = CLIPTextModel.from_pretrained(
@@ -520,6 +520,7 @@ def main(args):
     from torch_xla.distributed.fsdp.utils import apply_xla_patch_to_nn_linear
 
     unet = apply_xla_patch_to_nn_linear(unet, xs.xla_patched_nn_linear_forward)
+    unet.enable_xla_flash_attention(partition_spec=("data", None, None, None))
 
     vae.requires_grad_(False)
     text_encoder.requires_grad_(False)
@@ -530,15 +531,12 @@ def main(args):
     # as these weights are only used for inference, keeping weights in full
     # precision is not required.
     weight_dtype = torch.float32
-    if args.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif args.mixed_precision == "bf16":
+    if args.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
     device = xm.xla_device()
-    print("device: ", device)
-    print("weight_dtype: ", weight_dtype)
 
+    # Move text_encode and vae to device and cast to weight_dtype
     text_encoder = text_encoder.to(device, dtype=weight_dtype)
     vae = vae.to(device, dtype=weight_dtype)
     unet = unet.to(device, dtype=weight_dtype)
@@ -606,24 +604,27 @@ def main(args):
         collate_fn=collate_fn,
         num_workers=args.dataloader_num_workers,
         batch_size=args.train_batch_size,
+        prefetch_factor=args.loader_prefetch_factor,
     )
 
     train_dataloader = pl.MpDeviceLoader(
         train_dataloader,
         device,
         input_sharding={
-            "pixel_values": xs.ShardingSpec(mesh, ("x", None, None, None), minibatch=True),
-            "input_ids": xs.ShardingSpec(mesh, ("x", None), minibatch=True),
+            "pixel_values": xs.ShardingSpec(mesh, ("data", None, None, None), minibatch=True),
+            "input_ids": xs.ShardingSpec(mesh, ("data", None), minibatch=True),
         },
         loader_prefetch_size=args.loader_prefetch_size,
         device_prefetch_size=args.device_prefetch_size,
     )
 
+    num_hosts = xr.process_count()
+    num_devices_per_host = num_devices // num_hosts
     if xm.is_master_ordinal():
         print("***** Running training *****")
-        print(f"Instantaneous batch size per device = {args.train_batch_size}")
+        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host }")
         print(
-            f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_devices}"
+            f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_hosts}"
         )
         print(f"  Total optimization steps = {args.max_train_steps}")
 

examples/server/README.md

@@ -0,0 +1,61 @@
+
+# Create a server
+
+Diffusers' pipelines can be used as an inference engine for a server. It supports concurrent and multithreaded requests to generate images that may be requested by multiple users at the same time.
+
+This guide will show you how to use the [`StableDiffusion3Pipeline`] in a server, but feel free to use any pipeline you want.
+
+
+Start by navigating to the `examples/server` folder and installing all of the dependencies.
+
+```py
+pip install .
+pip install -f requirements.txt
+```
+
+Launch the server with the following command.
+
+```py
+python server.py
+```
+
+The server is accessed at http://localhost:8000. You can curl this model with the following command.
+```
+curl -X POST -H "Content-Type: application/json" --data '{"model": "something", "prompt": "a kitten in front of a fireplace"}' http://localhost:8000/v1/images/generations
+```
+
+If you need to upgrade some dependencies, you can use either [pip-tools](https://github.com/jazzband/pip-tools) or [uv](https://github.com/astral-sh/uv). For example, upgrade the dependencies with `uv` using the following command.
+
+```
+uv pip compile requirements.in -o requirements.txt
+```
+
+
+The server is built with [FastAPI](https://fastapi.tiangolo.com/async/). The endpoint for `v1/images/generations` is shown below.
+```py
+@app.post("/v1/images/generations")
+async def generate_image(image_input: TextToImageInput):
+    try:
+        loop = asyncio.get_event_loop()
+        scheduler = shared_pipeline.pipeline.scheduler.from_config(shared_pipeline.pipeline.scheduler.config)
+        pipeline = StableDiffusion3Pipeline.from_pipe(shared_pipeline.pipeline, scheduler=scheduler)
+        generator = torch.Generator(device="cuda")
+        generator.manual_seed(random.randint(0, 10000000))
+        output = await loop.run_in_executor(None, lambda: pipeline(image_input.prompt, generator = generator))
+        logger.info(f"output: {output}")
+        image_url = save_image(output.images[0])
+        return {"data": [{"url": image_url}]}
+    except Exception as e:
+        if isinstance(e, HTTPException):
+            raise e
+        elif hasattr(e, 'message'):
+            raise HTTPException(status_code=500, detail=e.message + traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e) + traceback.format_exc())
+```
+The `generate_image` function is defined as asynchronous with the [async](https://fastapi.tiangolo.com/async/) keyword so that FastAPI knows that whatever is happening in this function won't necessarily return a result right away. Once it hits some point in the function that it needs to await some other [Task](https://docs.python.org/3/library/asyncio-task.html#asyncio.Task), the main thread goes back to answering other HTTP requests. This is shown in the code below with the [await](https://fastapi.tiangolo.com/async/#async-and-await) keyword.
+```py
+output = await loop.run_in_executor(None, lambda: pipeline(image_input.prompt, generator = generator))
+```
+At this point, the execution of the pipeline function is placed onto a [new thread](https://docs.python.org/3/library/asyncio-eventloop.html#asyncio.loop.run_in_executor), and the main thread performs other things until a result is returned from the `pipeline`.
+
+Another important aspect of this implementation is creating a `pipeline` from `shared_pipeline`. The goal behind this is to avoid loading the underlying model more than once onto the GPU while still allowing for each new request that is running on a separate thread to have its own generator and scheduler. The scheduler, in particular, is not thread-safe, and it will cause errors like: `IndexError: index 21 is out of bounds for dimension 0 with size 21` if you try to use the same scheduler across multiple threads.

examples/server/requirements.in

@@ -0,0 +1,9 @@
+torch~=2.4.0
+transformers==4.46.1
+sentencepiece
+aiohttp
+py-consul
+prometheus_client >= 0.18.0
+prometheus-fastapi-instrumentator >= 7.0.0
+fastapi
+uvicorn

examples/server/requirements.txt

@@ -0,0 +1,124 @@
+# This file was autogenerated by uv via the following command:
+#    uv pip compile requirements.in -o requirements.txt
+aiohappyeyeballs==2.4.3
+    # via aiohttp
+aiohttp==3.10.10
+    # via -r requirements.in
+aiosignal==1.3.1
+    # via aiohttp
+annotated-types==0.7.0
+    # via pydantic
+anyio==4.6.2.post1
+    # via starlette
+attrs==24.2.0
+    # via aiohttp
+certifi==2024.8.30
+    # via requests
+charset-normalizer==3.4.0
+    # via requests
+click==8.1.7
+    # via uvicorn
+fastapi==0.115.3
+    # via -r requirements.in
+filelock==3.16.1
+    # via
+    #   huggingface-hub
+    #   torch
+    #   transformers
+frozenlist==1.5.0
+    # via
+    #   aiohttp
+    #   aiosignal
+fsspec==2024.10.0
+    # via
+    #   huggingface-hub
+    #   torch
+h11==0.14.0
+    # via uvicorn
+huggingface-hub==0.26.1
+    # via
+    #   tokenizers
+    #   transformers
+idna==3.10
+    # via
+    #   anyio
+    #   requests
+    #   yarl
+jinja2==3.1.4
+    # via torch
+markupsafe==3.0.2
+    # via jinja2
+mpmath==1.3.0
+    # via sympy
+multidict==6.1.0
+    # via
+    #   aiohttp
+    #   yarl
+networkx==3.4.2
+    # via torch
+numpy==2.1.2
+    # via transformers
+packaging==24.1
+    # via
+    #   huggingface-hub
+    #   transformers
+prometheus-client==0.21.0
+    # via
+    #   -r requirements.in
+    #   prometheus-fastapi-instrumentator
+prometheus-fastapi-instrumentator==7.0.0
+    # via -r requirements.in
+propcache==0.2.0
+    # via yarl
+py-consul==1.5.3
+    # via -r requirements.in
+pydantic==2.9.2
+    # via fastapi
+pydantic-core==2.23.4
+    # via pydantic
+pyyaml==6.0.2
+    # via
+    #   huggingface-hub
+    #   transformers
+regex==2024.9.11
+    # via transformers
+requests==2.32.3
+    # via
+    #   huggingface-hub
+    #   py-consul
+    #   transformers
+safetensors==0.4.5
+    # via transformers
+sentencepiece==0.2.0
+    # via -r requirements.in
+sniffio==1.3.1
+    # via anyio
+starlette==0.41.0
+    # via
+    #   fastapi
+    #   prometheus-fastapi-instrumentator
+sympy==1.13.3
+    # via torch
+tokenizers==0.20.1
+    # via transformers
+torch==2.4.1
+    # via -r requirements.in
+tqdm==4.66.5
+    # via
+    #   huggingface-hub
+    #   transformers
+transformers==4.46.1
+    # via -r requirements.in
+typing-extensions==4.12.2
+    # via
+    #   fastapi
+    #   huggingface-hub
+    #   pydantic
+    #   pydantic-core
+    #   torch
+urllib3==2.2.3
+    # via requests
+uvicorn==0.32.0
+    # via -r requirements.in
+yarl==1.16.0
+    # via aiohttp

examples/server/server.py

@@ -0,0 +1,133 @@
+import asyncio
+import logging
+import os
+import random
+import tempfile
+import traceback
+import uuid
+
+import aiohttp
+import torch
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.staticfiles import StaticFiles
+from pydantic import BaseModel
+
+from diffusers.pipelines.stable_diffusion_3 import StableDiffusion3Pipeline
+
+
+logger = logging.getLogger(__name__)
+
+
+class TextToImageInput(BaseModel):
+    model: str
+    prompt: str
+    size: str | None = None
+    n: int | None = None
+
+
+class HttpClient:
+    session: aiohttp.ClientSession = None
+
+    def start(self):
+        self.session = aiohttp.ClientSession()
+
+    async def stop(self):
+        await self.session.close()
+        self.session = None
+
+    def __call__(self) -> aiohttp.ClientSession:
+        assert self.session is not None
+        return self.session
+
+
+class TextToImagePipeline:
+    pipeline: StableDiffusion3Pipeline = None
+    device: str = None
+
+    def start(self):
+        if torch.cuda.is_available():
+            model_path = os.getenv("MODEL_PATH", "stabilityai/stable-diffusion-3.5-large")
+            logger.info("Loading CUDA")
+            self.device = "cuda"
+            self.pipeline = StableDiffusion3Pipeline.from_pretrained(
+                model_path,
+                torch_dtype=torch.bfloat16,
+            ).to(device=self.device)
+        elif torch.backends.mps.is_available():
+            model_path = os.getenv("MODEL_PATH", "stabilityai/stable-diffusion-3.5-medium")
+            logger.info("Loading MPS for Mac M Series")
+            self.device = "mps"
+            self.pipeline = StableDiffusion3Pipeline.from_pretrained(
+                model_path,
+                torch_dtype=torch.bfloat16,
+            ).to(device=self.device)
+        else:
+            raise Exception("No CUDA or MPS device available")
+
+
+app = FastAPI()
+service_url = os.getenv("SERVICE_URL", "http://localhost:8000")
+image_dir = os.path.join(tempfile.gettempdir(), "images")
+if not os.path.exists(image_dir):
+    os.makedirs(image_dir)
+app.mount("/images", StaticFiles(directory=image_dir), name="images")
+http_client = HttpClient()
+shared_pipeline = TextToImagePipeline()
+
+# Configure CORS settings
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],  # Allows all origins
+    allow_credentials=True,
+    allow_methods=["*"],  # Allows all methods, e.g., GET, POST, OPTIONS, etc.
+    allow_headers=["*"],  # Allows all headers
+)
+
+
+@app.on_event("startup")
+def startup():
+    http_client.start()
+    shared_pipeline.start()
+
+
+def save_image(image):
+    filename = "draw" + str(uuid.uuid4()).split("-")[0] + ".png"
+    image_path = os.path.join(image_dir, filename)
+    # write image to disk at image_path
+    logger.info(f"Saving image to {image_path}")
+    image.save(image_path)
+    return os.path.join(service_url, "images", filename)
+
+
+@app.get("/")
+@app.post("/")
+@app.options("/")
+async def base():
+    return "Welcome to Diffusers! Where you can use diffusion models to generate images"
+
+
+@app.post("/v1/images/generations")
+async def generate_image(image_input: TextToImageInput):
+    try:
+        loop = asyncio.get_event_loop()
+        scheduler = shared_pipeline.pipeline.scheduler.from_config(shared_pipeline.pipeline.scheduler.config)
+        pipeline = StableDiffusion3Pipeline.from_pipe(shared_pipeline.pipeline, scheduler=scheduler)
+        generator = torch.Generator(device=shared_pipeline.device)
+        generator.manual_seed(random.randint(0, 10000000))
+        output = await loop.run_in_executor(None, lambda: pipeline(image_input.prompt, generator=generator))
+        logger.info(f"output: {output}")
+        image_url = save_image(output.images[0])
+        return {"data": [{"url": image_url}]}
+    except Exception as e:
+        if isinstance(e, HTTPException):
+            raise e
+        elif hasattr(e, "message"):
+            raise HTTPException(status_code=500, detail=e.message + traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e) + traceback.format_exc())
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    uvicorn.run(app, host="0.0.0.0", port=8000)

examples/text_to_image/train_text_to_image_sdxl.py

@@ -483,7 +483,6 @@ def parse_args(input_args=None):
     # Sanity checks
     if args.dataset_name is None and args.train_data_dir is None:
         raise ValueError("Need either a dataset name or a training folder.")
-
     if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
         raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
 
@@ -824,9 +823,7 @@ def main(args):
     if args.dataset_name is not None:
         # Downloading and loading a dataset from the hub.
         dataset = load_dataset(
-            args.dataset_name,
-            args.dataset_config_name,
-            cache_dir=args.cache_dir,
+            args.dataset_name, args.dataset_config_name, cache_dir=args.cache_dir, data_dir=args.train_data_dir
         )
     else:
         data_files = {}

scripts/convert_cogvideox_to_diffusers.py

@@ -80,6 +80,8 @@ TRANSFORMER_KEYS_RENAME_DICT = {
     "post_attn1_layernorm": "norm2.norm",
     "time_embed.0": "time_embedding.linear_1",
     "time_embed.2": "time_embedding.linear_2",
+    "ofs_embed.0": "ofs_embedding.linear_1",
+    "ofs_embed.2": "ofs_embedding.linear_2",
     "mixins.patch_embed": "patch_embed",
     "mixins.final_layer.norm_final": "norm_out.norm",
     "mixins.final_layer.linear": "proj_out",
@@ -140,6 +142,7 @@ def convert_transformer(
     use_rotary_positional_embeddings: bool,
     i2v: bool,
     dtype: torch.dtype,
+    init_kwargs: Dict[str, Any],
 ):
     PREFIX_KEY = "model.diffusion_model."
 
@@ -149,7 +152,9 @@ def convert_transformer(
         num_layers=num_layers,
         num_attention_heads=num_attention_heads,
         use_rotary_positional_embeddings=use_rotary_positional_embeddings,
-        use_learned_positional_embeddings=i2v,
+        ofs_embed_dim=512 if (i2v and init_kwargs["patch_size_t"] is not None) else None,  # CogVideoX1.5-5B-I2V
+        use_learned_positional_embeddings=i2v and init_kwargs["patch_size_t"] is None,  # CogVideoX-5B-I2V
+        **init_kwargs,
     ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
@@ -163,13 +168,18 @@ def convert_transformer(
             if special_key not in key:
                 continue
             handler_fn_inplace(key, original_state_dict)
+
     transformer.load_state_dict(original_state_dict, strict=True)
     return transformer
 
 
-def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
+def convert_vae(ckpt_path: str, scaling_factor: float, version: str, dtype: torch.dtype):
+    init_kwargs = {"scaling_factor": scaling_factor}
+    if version == "1.5":
+        init_kwargs.update({"invert_scale_latents": True})
+
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
+    vae = AutoencoderKLCogVideoX(**init_kwargs).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -187,6 +197,34 @@ def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
     return vae
 
 
+def get_transformer_init_kwargs(version: str):
+    if version == "1.0":
+        vae_scale_factor_spatial = 8
+        init_kwargs = {
+            "patch_size": 2,
+            "patch_size_t": None,
+            "patch_bias": True,
+            "sample_height": 480 // vae_scale_factor_spatial,
+            "sample_width": 720 // vae_scale_factor_spatial,
+            "sample_frames": 49,
+        }
+
+    elif version == "1.5":
+        vae_scale_factor_spatial = 8
+        init_kwargs = {
+            "patch_size": 2,
+            "patch_size_t": 2,
+            "patch_bias": False,
+            "sample_height": 300,
+            "sample_width": 300,
+            "sample_frames": 81,
+        }
+    else:
+        raise ValueError("Unsupported version of CogVideoX.")
+
+    return init_kwargs
+
+
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -202,6 +240,12 @@ def get_args():
     parser.add_argument(
         "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
     )
+    parser.add_argument(
+        "--typecast_text_encoder",
+        action="store_true",
+        default=False,
+        help="Whether or not to apply fp16/bf16 precision to text_encoder",
+    )
     # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
     parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
     # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
@@ -214,7 +258,18 @@ def get_args():
     parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
     # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
     parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
-    parser.add_argument("--i2v", action="store_true", default=False, help="Whether to save the model weights in fp16")
+    parser.add_argument(
+        "--i2v",
+        action="store_true",
+        default=False,
+        help="Whether the model to be converted is the Image-to-Video version of CogVideoX.",
+    )
+    parser.add_argument(
+        "--version",
+        choices=["1.0", "1.5"],
+        default="1.0",
+        help="Which version of CogVideoX to use for initializing default modeling parameters.",
+    )
     return parser.parse_args()
 
 
@@ -230,6 +285,7 @@ if __name__ == "__main__":
     dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
 
     if args.transformer_ckpt_path is not None:
+        init_kwargs = get_transformer_init_kwargs(args.version)
         transformer = convert_transformer(
             args.transformer_ckpt_path,
             args.num_layers,
@@ -237,14 +293,19 @@ if __name__ == "__main__":
             args.use_rotary_positional_embeddings,
             args.i2v,
             dtype,
+            init_kwargs,
         )
     if args.vae_ckpt_path is not None:
-        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype)
+        # Keep VAE in float32 for better quality
+        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, args.version, torch.float32)
 
     text_encoder_id = "google/t5-v1_1-xxl"
     tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
     text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
 
+    if args.typecast_text_encoder:
+        text_encoder = text_encoder.to(dtype=dtype)
+
     # Apparently, the conversion does not work anymore without this :shrug:
     for param in text_encoder.parameters():
         param.data = param.data.contiguous()
@@ -276,11 +337,6 @@ if __name__ == "__main__":
         scheduler=scheduler,
     )
 
-    if args.fp16:
-        pipe = pipe.to(dtype=torch.float16)
-    if args.bf16:
-        pipe = pipe.to(dtype=torch.bfloat16)
-
     # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
     # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
     # is either fp16/bf16 here).

scripts/convert_cogview3_to_diffusers.py

@@ -36,7 +36,7 @@ from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
 from diffusers.utils.import_utils import is_accelerate_available
 
 
-CTX = init_empty_weights if is_accelerate_available else nullcontext
+CTX = init_empty_weights if is_accelerate_available() else nullcontext
 
 TOKENIZER_MAX_LENGTH = 224
 

scripts/convert_dcae_to_diffusers.py

@@ -0,0 +1,323 @@
+import argparse
+from typing import Any, Dict
+
+import torch
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+
+from diffusers import AutoencoderDC
+
+
+def remap_qkv_(key: str, state_dict: Dict[str, Any]):
+    qkv = state_dict.pop(key)
+    q, k, v = torch.chunk(qkv, 3, dim=0)
+    parent_module, _, _ = key.rpartition(".qkv.conv.weight")
+    state_dict[f"{parent_module}.to_q.weight"] = q.squeeze()
+    state_dict[f"{parent_module}.to_k.weight"] = k.squeeze()
+    state_dict[f"{parent_module}.to_v.weight"] = v.squeeze()
+
+
+def remap_proj_conv_(key: str, state_dict: Dict[str, Any]):
+    parent_module, _, _ = key.rpartition(".proj.conv.weight")
+    state_dict[f"{parent_module}.to_out.weight"] = state_dict.pop(key).squeeze()
+
+
+AE_KEYS_RENAME_DICT = {
+    # common
+    "main.": "",
+    "op_list.": "",
+    "context_module": "attn",
+    "local_module": "conv_out",
+    # NOTE: The below two lines work because scales in the available configs only have a tuple length of 1
+    # If there were more scales, there would be more layers, so a loop would be better to handle this
+    "aggreg.0.0": "to_qkv_multiscale.0.proj_in",
+    "aggreg.0.1": "to_qkv_multiscale.0.proj_out",
+    "depth_conv.conv": "conv_depth",
+    "inverted_conv.conv": "conv_inverted",
+    "point_conv.conv": "conv_point",
+    "point_conv.norm": "norm",
+    "conv.conv.": "conv.",
+    "conv1.conv": "conv1",
+    "conv2.conv": "conv2",
+    "conv2.norm": "norm",
+    "proj.norm": "norm_out",
+    # encoder
+    "encoder.project_in.conv": "encoder.conv_in",
+    "encoder.project_out.0.conv": "encoder.conv_out",
+    "encoder.stages": "encoder.down_blocks",
+    # decoder
+    "decoder.project_in.conv": "decoder.conv_in",
+    "decoder.project_out.0": "decoder.norm_out",
+    "decoder.project_out.2.conv": "decoder.conv_out",
+    "decoder.stages": "decoder.up_blocks",
+}
+
+AE_F32C32_KEYS = {
+    # encoder
+    "encoder.project_in.conv": "encoder.conv_in.conv",
+    # decoder
+    "decoder.project_out.2.conv": "decoder.conv_out.conv",
+}
+
+AE_F64C128_KEYS = {
+    # encoder
+    "encoder.project_in.conv": "encoder.conv_in.conv",
+    # decoder
+    "decoder.project_out.2.conv": "decoder.conv_out.conv",
+}
+
+AE_F128C512_KEYS = {
+    # encoder
+    "encoder.project_in.conv": "encoder.conv_in.conv",
+    # decoder
+    "decoder.project_out.2.conv": "decoder.conv_out.conv",
+}
+
+AE_SPECIAL_KEYS_REMAP = {
+    "qkv.conv.weight": remap_qkv_,
+    "proj.conv.weight": remap_proj_conv_,
+}
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def convert_ae(config_name: str, dtype: torch.dtype):
+    config = get_ae_config(config_name)
+    hub_id = f"mit-han-lab/{config_name}"
+    ckpt_path = hf_hub_download(hub_id, "model.safetensors")
+    original_state_dict = get_state_dict(load_file(ckpt_path))
+
+    ae = AutoencoderDC(**config).to(dtype=dtype)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in AE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in AE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    ae.load_state_dict(original_state_dict, strict=True)
+    return ae
+
+
+def get_ae_config(name: str):
+    if name in ["dc-ae-f32c32-sana-1.0"]:
+        config = {
+            "latent_channels": 32,
+            "encoder_block_types": (
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ),
+            "decoder_block_types": (
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ),
+            "encoder_block_out_channels": (128, 256, 512, 512, 1024, 1024),
+            "decoder_block_out_channels": (128, 256, 512, 512, 1024, 1024),
+            "encoder_qkv_multiscales": ((), (), (), (5,), (5,), (5,)),
+            "decoder_qkv_multiscales": ((), (), (), (5,), (5,), (5,)),
+            "encoder_layers_per_block": (2, 2, 2, 3, 3, 3),
+            "decoder_layers_per_block": [3, 3, 3, 3, 3, 3],
+            "downsample_block_type": "conv",
+            "upsample_block_type": "interpolate",
+            "decoder_norm_types": "rms_norm",
+            "decoder_act_fns": "silu",
+            "scaling_factor": 0.41407,
+        }
+    elif name in ["dc-ae-f32c32-in-1.0", "dc-ae-f32c32-mix-1.0"]:
+        AE_KEYS_RENAME_DICT.update(AE_F32C32_KEYS)
+        config = {
+            "latent_channels": 32,
+            "encoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "decoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024],
+            "decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024],
+            "encoder_layers_per_block": [0, 4, 8, 2, 2, 2],
+            "decoder_layers_per_block": [0, 5, 10, 2, 2, 2],
+            "encoder_qkv_multiscales": ((), (), (), (), (), ()),
+            "decoder_qkv_multiscales": ((), (), (), (), (), ()),
+            "decoder_norm_types": ["batch_norm", "batch_norm", "batch_norm", "rms_norm", "rms_norm", "rms_norm"],
+            "decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu"],
+        }
+        if name == "dc-ae-f32c32-in-1.0":
+            config["scaling_factor"] = 0.3189
+        elif name == "dc-ae-f32c32-mix-1.0":
+            config["scaling_factor"] = 0.4552
+    elif name in ["dc-ae-f64c128-in-1.0", "dc-ae-f64c128-mix-1.0"]:
+        AE_KEYS_RENAME_DICT.update(AE_F64C128_KEYS)
+        config = {
+            "latent_channels": 128,
+            "encoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "decoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048],
+            "decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048],
+            "encoder_layers_per_block": [0, 4, 8, 2, 2, 2, 2],
+            "decoder_layers_per_block": [0, 5, 10, 2, 2, 2, 2],
+            "encoder_qkv_multiscales": ((), (), (), (), (), (), ()),
+            "decoder_qkv_multiscales": ((), (), (), (), (), (), ()),
+            "decoder_norm_types": [
+                "batch_norm",
+                "batch_norm",
+                "batch_norm",
+                "rms_norm",
+                "rms_norm",
+                "rms_norm",
+                "rms_norm",
+            ],
+            "decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu", "silu"],
+        }
+        if name == "dc-ae-f64c128-in-1.0":
+            config["scaling_factor"] = 0.2889
+        elif name == "dc-ae-f64c128-mix-1.0":
+            config["scaling_factor"] = 0.4538
+    elif name in ["dc-ae-f128c512-in-1.0", "dc-ae-f128c512-mix-1.0"]:
+        AE_KEYS_RENAME_DICT.update(AE_F128C512_KEYS)
+        config = {
+            "latent_channels": 512,
+            "encoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "decoder_block_types": [
+                "ResBlock",
+                "ResBlock",
+                "ResBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+                "EfficientViTBlock",
+            ],
+            "encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048, 2048],
+            "decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048, 2048],
+            "encoder_layers_per_block": [0, 4, 8, 2, 2, 2, 2, 2],
+            "decoder_layers_per_block": [0, 5, 10, 2, 2, 2, 2, 2],
+            "encoder_qkv_multiscales": ((), (), (), (), (), (), (), ()),
+            "decoder_qkv_multiscales": ((), (), (), (), (), (), (), ()),
+            "decoder_norm_types": [
+                "batch_norm",
+                "batch_norm",
+                "batch_norm",
+                "rms_norm",
+                "rms_norm",
+                "rms_norm",
+                "rms_norm",
+                "rms_norm",
+            ],
+            "decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu", "silu", "silu"],
+        }
+        if name == "dc-ae-f128c512-in-1.0":
+            config["scaling_factor"] = 0.4883
+        elif name == "dc-ae-f128c512-mix-1.0":
+            config["scaling_factor"] = 0.3620
+    else:
+        raise ValueError("Invalid config name provided.")
+
+    return config
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--config_name",
+        type=str,
+        default="dc-ae-f32c32-sana-1.0",
+        choices=[
+            "dc-ae-f32c32-sana-1.0",
+            "dc-ae-f32c32-in-1.0",
+            "dc-ae-f32c32-mix-1.0",
+            "dc-ae-f64c128-in-1.0",
+            "dc-ae-f64c128-mix-1.0",
+            "dc-ae-f128c512-in-1.0",
+            "dc-ae-f128c512-mix-1.0",
+        ],
+        help="The DCAE checkpoint to convert",
+    )
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    ae = convert_ae(args.config_name, dtype)
+    ae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant)

scripts/convert_flux_to_diffusers.py

@@ -31,12 +31,14 @@ python scripts/convert_flux_to_diffusers.py  \
 --vae
 """
 
-CTX = init_empty_weights if is_accelerate_available else nullcontext
+CTX = init_empty_weights if is_accelerate_available() else nullcontext
 
 parser = argparse.ArgumentParser()
 parser.add_argument("--original_state_dict_repo_id", default=None, type=str)
 parser.add_argument("--filename", default="flux.safetensors", type=str)
 parser.add_argument("--checkpoint_path", default=None, type=str)
+parser.add_argument("--in_channels", type=int, default=64)
+parser.add_argument("--out_channels", type=int, default=None)
 parser.add_argument("--vae", action="store_true")
 parser.add_argument("--transformer", action="store_true")
 parser.add_argument("--output_path", type=str)
@@ -279,10 +281,13 @@ def main(args):
         num_single_layers = 38
         inner_dim = 3072
         mlp_ratio = 4.0
+
         converted_transformer_state_dict = convert_flux_transformer_checkpoint_to_diffusers(
             original_ckpt, num_layers, num_single_layers, inner_dim, mlp_ratio=mlp_ratio
         )
-        transformer = FluxTransformer2DModel(guidance_embeds=has_guidance)
+        transformer = FluxTransformer2DModel(
+            in_channels=args.in_channels, out_channels=args.out_channels, guidance_embeds=has_guidance
+        )
         transformer.load_state_dict(converted_transformer_state_dict, strict=True)
 
         print(

scripts/convert_hunyuan_video_to_diffusers.py

@@ -0,0 +1,257 @@
+import argparse
+from typing import Any, Dict
+
+import torch
+from accelerate import init_empty_weights
+from transformers import AutoModel, AutoTokenizer, CLIPTextModel, CLIPTokenizer
+
+from diffusers import (
+    AutoencoderKLHunyuanVideo,
+    FlowMatchEulerDiscreteScheduler,
+    HunyuanVideoPipeline,
+    HunyuanVideoTransformer3DModel,
+)
+
+
+def remap_norm_scale_shift_(key, state_dict):
+    weight = state_dict.pop(key)
+    shift, scale = weight.chunk(2, dim=0)
+    new_weight = torch.cat([scale, shift], dim=0)
+    state_dict[key.replace("final_layer.adaLN_modulation.1", "norm_out.linear")] = new_weight
+
+
+def remap_txt_in_(key, state_dict):
+    def rename_key(key):
+        new_key = key.replace("individual_token_refiner.blocks", "token_refiner.refiner_blocks")
+        new_key = new_key.replace("adaLN_modulation.1", "norm_out.linear")
+        new_key = new_key.replace("txt_in", "context_embedder")
+        new_key = new_key.replace("t_embedder.mlp.0", "time_text_embed.timestep_embedder.linear_1")
+        new_key = new_key.replace("t_embedder.mlp.2", "time_text_embed.timestep_embedder.linear_2")
+        new_key = new_key.replace("c_embedder", "time_text_embed.text_embedder")
+        new_key = new_key.replace("mlp", "ff")
+        return new_key
+
+    if "self_attn_qkv" in key:
+        weight = state_dict.pop(key)
+        to_q, to_k, to_v = weight.chunk(3, dim=0)
+        state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_q"))] = to_q
+        state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_k"))] = to_k
+        state_dict[rename_key(key.replace("self_attn_qkv", "attn.to_v"))] = to_v
+    else:
+        state_dict[rename_key(key)] = state_dict.pop(key)
+
+
+def remap_img_attn_qkv_(key, state_dict):
+    weight = state_dict.pop(key)
+    to_q, to_k, to_v = weight.chunk(3, dim=0)
+    state_dict[key.replace("img_attn_qkv", "attn.to_q")] = to_q
+    state_dict[key.replace("img_attn_qkv", "attn.to_k")] = to_k
+    state_dict[key.replace("img_attn_qkv", "attn.to_v")] = to_v
+
+
+def remap_txt_attn_qkv_(key, state_dict):
+    weight = state_dict.pop(key)
+    to_q, to_k, to_v = weight.chunk(3, dim=0)
+    state_dict[key.replace("txt_attn_qkv", "attn.add_q_proj")] = to_q
+    state_dict[key.replace("txt_attn_qkv", "attn.add_k_proj")] = to_k
+    state_dict[key.replace("txt_attn_qkv", "attn.add_v_proj")] = to_v
+
+
+def remap_single_transformer_blocks_(key, state_dict):
+    hidden_size = 3072
+
+    if "linear1.weight" in key:
+        linear1_weight = state_dict.pop(key)
+        split_size = (hidden_size, hidden_size, hidden_size, linear1_weight.size(0) - 3 * hidden_size)
+        q, k, v, mlp = torch.split(linear1_weight, split_size, dim=0)
+        new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(".linear1.weight")
+        state_dict[f"{new_key}.attn.to_q.weight"] = q
+        state_dict[f"{new_key}.attn.to_k.weight"] = k
+        state_dict[f"{new_key}.attn.to_v.weight"] = v
+        state_dict[f"{new_key}.proj_mlp.weight"] = mlp
+
+    elif "linear1.bias" in key:
+        linear1_bias = state_dict.pop(key)
+        split_size = (hidden_size, hidden_size, hidden_size, linear1_bias.size(0) - 3 * hidden_size)
+        q_bias, k_bias, v_bias, mlp_bias = torch.split(linear1_bias, split_size, dim=0)
+        new_key = key.replace("single_blocks", "single_transformer_blocks").removesuffix(".linear1.bias")
+        state_dict[f"{new_key}.attn.to_q.bias"] = q_bias
+        state_dict[f"{new_key}.attn.to_k.bias"] = k_bias
+        state_dict[f"{new_key}.attn.to_v.bias"] = v_bias
+        state_dict[f"{new_key}.proj_mlp.bias"] = mlp_bias
+
+    else:
+        new_key = key.replace("single_blocks", "single_transformer_blocks")
+        new_key = new_key.replace("linear2", "proj_out")
+        new_key = new_key.replace("q_norm", "attn.norm_q")
+        new_key = new_key.replace("k_norm", "attn.norm_k")
+        state_dict[new_key] = state_dict.pop(key)
+
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "img_in": "x_embedder",
+    "time_in.mlp.0": "time_text_embed.timestep_embedder.linear_1",
+    "time_in.mlp.2": "time_text_embed.timestep_embedder.linear_2",
+    "guidance_in.mlp.0": "time_text_embed.guidance_embedder.linear_1",
+    "guidance_in.mlp.2": "time_text_embed.guidance_embedder.linear_2",
+    "vector_in.in_layer": "time_text_embed.text_embedder.linear_1",
+    "vector_in.out_layer": "time_text_embed.text_embedder.linear_2",
+    "double_blocks": "transformer_blocks",
+    "img_attn_q_norm": "attn.norm_q",
+    "img_attn_k_norm": "attn.norm_k",
+    "img_attn_proj": "attn.to_out.0",
+    "txt_attn_q_norm": "attn.norm_added_q",
+    "txt_attn_k_norm": "attn.norm_added_k",
+    "txt_attn_proj": "attn.to_add_out",
+    "img_mod.linear": "norm1.linear",
+    "img_norm1": "norm1.norm",
+    "img_norm2": "norm2",
+    "img_mlp": "ff",
+    "txt_mod.linear": "norm1_context.linear",
+    "txt_norm1": "norm1.norm",
+    "txt_norm2": "norm2_context",
+    "txt_mlp": "ff_context",
+    "self_attn_proj": "attn.to_out.0",
+    "modulation.linear": "norm.linear",
+    "pre_norm": "norm.norm",
+    "final_layer.norm_final": "norm_out.norm",
+    "final_layer.linear": "proj_out",
+    "fc1": "net.0.proj",
+    "fc2": "net.2",
+    "input_embedder": "proj_in",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {
+    "txt_in": remap_txt_in_,
+    "img_attn_qkv": remap_img_attn_qkv_,
+    "txt_attn_qkv": remap_txt_attn_qkv_,
+    "single_blocks": remap_single_transformer_blocks_,
+    "final_layer.adaLN_modulation.1": remap_norm_scale_shift_,
+}
+
+VAE_KEYS_RENAME_DICT = {}
+
+VAE_SPECIAL_KEYS_REMAP = {}
+
+
+def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def convert_transformer(ckpt_path: str):
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", weights_only=True))
+
+    with init_empty_weights():
+        transformer = HunyuanVideoTransformer3DModel()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True, assign=True)
+    return transformer
+
+
+def convert_vae(ckpt_path: str):
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", weights_only=True))
+
+    with init_empty_weights():
+        vae = AutoencoderKLHunyuanVideo()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    vae.load_state_dict(original_state_dict, strict=True, assign=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
+    )
+    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original VAE checkpoint")
+    parser.add_argument("--text_encoder_path", type=str, default=None, help="Path to original llama checkpoint")
+    parser.add_argument("--tokenizer_path", type=str, default=None, help="Path to original llama tokenizer")
+    parser.add_argument("--text_encoder_2_path", type=str, default=None, help="Path to original clip checkpoint")
+    parser.add_argument("--save_pipeline", action="store_true")
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--dtype", default="bf16", help="Torch dtype to save the transformer in.")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    dtype = DTYPE_MAPPING[args.dtype]
+
+    if args.save_pipeline:
+        assert args.transformer_ckpt_path is not None and args.vae_ckpt_path is not None
+        assert args.text_encoder_path is not None
+        assert args.tokenizer_path is not None
+        assert args.text_encoder_2_path is not None
+
+    if args.transformer_ckpt_path is not None:
+        transformer = convert_transformer(args.transformer_ckpt_path)
+        transformer = transformer.to(dtype=dtype)
+        if not args.save_pipeline:
+            transformer.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+
+    if args.vae_ckpt_path is not None:
+        vae = convert_vae(args.vae_ckpt_path)
+        if not args.save_pipeline:
+            vae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+
+    if args.save_pipeline:
+        text_encoder = AutoModel.from_pretrained(args.text_encoder_path, torch_dtype=torch.float16)
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path, padding_side="right")
+        text_encoder_2 = CLIPTextModel.from_pretrained(args.text_encoder_2_path, torch_dtype=torch.float16)
+        tokenizer_2 = CLIPTokenizer.from_pretrained(args.text_encoder_2_path)
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+
+        pipe = HunyuanVideoPipeline(
+            transformer=transformer,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            text_encoder_2=text_encoder_2,
+            tokenizer_2=tokenizer_2,
+            scheduler=scheduler,
+        )
+        pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")

scripts/convert_ltx_to_diffusers.py

@@ -0,0 +1,209 @@
+import argparse
+from typing import Any, Dict
+
+import torch
+from safetensors.torch import load_file
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
+
+
+def remove_keys_(key: str, state_dict: Dict[str, Any]):
+    state_dict.pop(key)
+
+
+TOKENIZER_MAX_LENGTH = 128
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "patchify_proj": "proj_in",
+    "adaln_single": "time_embed",
+    "q_norm": "norm_q",
+    "k_norm": "norm_k",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+
+VAE_KEYS_RENAME_DICT = {
+    # decoder
+    "up_blocks.0": "mid_block",
+    "up_blocks.1": "up_blocks.0",
+    "up_blocks.2": "up_blocks.1.upsamplers.0",
+    "up_blocks.3": "up_blocks.1",
+    "up_blocks.4": "up_blocks.2.conv_in",
+    "up_blocks.5": "up_blocks.2.upsamplers.0",
+    "up_blocks.6": "up_blocks.2",
+    "up_blocks.7": "up_blocks.3.conv_in",
+    "up_blocks.8": "up_blocks.3.upsamplers.0",
+    "up_blocks.9": "up_blocks.3",
+    # encoder
+    "down_blocks.0": "down_blocks.0",
+    "down_blocks.1": "down_blocks.0.downsamplers.0",
+    "down_blocks.2": "down_blocks.0.conv_out",
+    "down_blocks.3": "down_blocks.1",
+    "down_blocks.4": "down_blocks.1.downsamplers.0",
+    "down_blocks.5": "down_blocks.1.conv_out",
+    "down_blocks.6": "down_blocks.2",
+    "down_blocks.7": "down_blocks.2.downsamplers.0",
+    "down_blocks.8": "down_blocks.3",
+    "down_blocks.9": "mid_block",
+    # common
+    "conv_shortcut": "conv_shortcut.conv",
+    "res_blocks": "resnets",
+    "norm3.norm": "norm3",
+    "per_channel_statistics.mean-of-means": "latents_mean",
+    "per_channel_statistics.std-of-means": "latents_std",
+}
+
+VAE_SPECIAL_KEYS_REMAP = {
+    "per_channel_statistics.channel": remove_keys_,
+    "per_channel_statistics.mean-of-means": remove_keys_,
+    "per_channel_statistics.mean-of-stds": remove_keys_,
+}
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def convert_transformer(
+    ckpt_path: str,
+    dtype: torch.dtype,
+):
+    PREFIX_KEY = ""
+
+    original_state_dict = get_state_dict(load_file(ckpt_path))
+    transformer = LTXVideoTransformer3DModel().to(dtype=dtype)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[len(PREFIX_KEY) :]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True)
+    return transformer
+
+
+def convert_vae(ckpt_path: str, dtype: torch.dtype):
+    original_state_dict = get_state_dict(load_file(ckpt_path))
+    vae = AutoencoderKLLTXVideo().to(dtype=dtype)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    vae.load_state_dict(original_state_dict, strict=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
+    )
+    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
+    parser.add_argument(
+        "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
+    )
+    parser.add_argument(
+        "--typecast_text_encoder",
+        action="store_true",
+        default=False,
+        help="Whether or not to apply fp16/bf16 precision to text_encoder",
+    )
+    parser.add_argument("--save_pipeline", action="store_true")
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    if args.save_pipeline:
+        assert args.transformer_ckpt_path is not None and args.vae_ckpt_path is not None
+
+    if args.transformer_ckpt_path is not None:
+        transformer: LTXVideoTransformer3DModel = convert_transformer(args.transformer_ckpt_path, dtype)
+        if not args.save_pipeline:
+            transformer.save_pretrained(
+                args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant
+            )
+
+    if args.vae_ckpt_path is not None:
+        vae: AutoencoderKLLTXVideo = convert_vae(args.vae_ckpt_path, dtype)
+        if not args.save_pipeline:
+            vae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant)
+
+    if args.save_pipeline:
+        text_encoder_id = "google/t5-v1_1-xxl"
+        tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
+        text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
+
+        if args.typecast_text_encoder:
+            text_encoder = text_encoder.to(dtype=dtype)
+
+        # Apparently, the conversion does not work anymore without this :shrug:
+        for param in text_encoder.parameters():
+            param.data = param.data.contiguous()
+
+        scheduler = FlowMatchEulerDiscreteScheduler(
+            use_dynamic_shifting=True,
+            base_shift=0.95,
+            max_shift=2.05,
+            base_image_seq_len=1024,
+            max_image_seq_len=4096,
+            shift_terminal=0.1,
+        )
+
+        pipe = LTXPipeline(
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+        )
+
+        pipe.save_pretrained(args.output_path, safe_serialization=True, variant=variant, max_shard_size="5GB")

scripts/convert_mochi_to_diffusers.py

@@ -10,7 +10,7 @@ from diffusers import AutoencoderKLMochi, FlowMatchEulerDiscreteScheduler, Mochi
 from diffusers.utils.import_utils import is_accelerate_available
 
 
-CTX = init_empty_weights if is_accelerate_available else nullcontext
+CTX = init_empty_weights if is_accelerate_available() else nullcontext
 
 TOKENIZER_MAX_LENGTH = 256