Merge branch 'main' into fix-test

update

.github/workflows/nightly_tests.yml

@@ -12,110 +12,96 @@ env:
   PYTEST_TIMEOUT: 600
   RUN_SLOW: yes
   RUN_NIGHTLY: yes
+  PIPELINE_USAGE_CUTOFF: 5000
   SLACK_API_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
 
 jobs:
-  run_nightly_tests:
-    strategy:
-      fail-fast: false
-      matrix:
-        config:
-          - name: Nightly PyTorch CUDA tests on Ubuntu
-            framework: pytorch
-            runner: docker-gpu
-            image: diffusers/diffusers-pytorch-cuda
-            report: torch_cuda
-          - name: Nightly Flax TPU tests on Ubuntu
-            framework: flax
-            runner: docker-tpu
-            image: diffusers/diffusers-flax-tpu
-            report: flax_tpu
-          - name: Nightly ONNXRuntime CUDA tests on Ubuntu
-            framework: onnxruntime
-            runner: docker-gpu
-            image: diffusers/diffusers-onnxruntime-cuda
-            report: onnx_cuda
-
-    name: ${{ matrix.config.name }}
-
-    runs-on: ${{ matrix.config.runner }}
-
-    container:
-      image: ${{ matrix.config.image }}
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ ${{ matrix.config.runner == 'docker-tpu' && '--privileged' || '--gpus 0'}}
-
-    defaults:
-      run:
-        shell: bash
-
+  setup_torch_cuda_pipeline_matrix:
+    name: Setup Torch Pipelines Matrix
+    runs-on: ubuntu-latest
+    outputs:
+      pipeline_test_matrix: ${{ steps.fetch_pipeline_matrix.outputs.pipeline_test_matrix }}
     steps:
       - name: Checkout diffusers
         uses: actions/checkout@v3
         with:
           fetch-depth: 2
-
-      - name: NVIDIA-SMI
-        if: ${{ matrix.config.runner == 'docker-gpu' }}
-        run: |
-          nvidia-smi
-
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
       - name: Install dependencies
         run: |
+          pip install -e .
+          pip install huggingface_hub
+      - name: Fetch Pipeline Matrix
+        id: fetch_pipeline_matrix
+        run: |
+          matrix=$(python utils/fetch_torch_cuda_pipeline_test_matrix.py)
+          echo $matrix
+          echo "pipeline_test_matrix=$matrix" >> $GITHUB_OUTPUT
+
+      - name: Pipeline Tests Artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: test-pipelines.json
+          path: reports
+
+  run_nightly_tests_for_torch_pipelines:
+    name: Torch Pipelines CUDA Nightly Tests
+    needs: setup_torch_cuda_pipeline_matrix
+    strategy:
+      fail-fast: false
+      matrix:
+        module: ${{ fromJson(needs.setup_torch_cuda_pipeline_matrix.outputs.pipeline_test_matrix) }}
+    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: NVIDIA-SMI
+        run: nvidia-smi
+      
+      - name: Install dependencies
+        run: |
+          apt-get update && apt-get install libsndfile1-dev libgl1 -y
           python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           python -m uv pip install -e [quality,test]
-          python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers
-          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
+          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
           python -m uv pip install pytest-reportlog
-
+      
       - name: Environment
         run: |
           python utils/print_env.py
-
-      - name: Run nightly PyTorch CUDA tests
-        if: ${{ matrix.config.framework == 'pytorch' }}
+      
+      - name: Nightly PyTorch CUDA checkpoint (pipelines) tests 
         env:
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+          # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+          CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
             -s -v -k "not Flax and not Onnx" \
-            --make-reports=tests_${{ matrix.config.report }} \
-            --report-log=${{ matrix.config.report }}.log \
-            tests/ 
-
-      - name: Run nightly Flax TPU tests
-        if: ${{ matrix.config.framework == 'flax' }}
-        env:
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
-        run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m pytest -n 0 \
-            -s -v -k "Flax" \
-            --make-reports=tests_${{ matrix.config.report }} \
-            --report-log=${{ matrix.config.report }}.log \
-            tests/
-
-      - name: Run nightly ONNXRuntime CUDA tests
-        if: ${{ matrix.config.framework == 'onnxruntime' }}
-        env:
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
-        run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "Onnx" \
-            --make-reports=tests_${{ matrix.config.report }} \
-            --report-log=${{ matrix.config.report }}.log \ 
-            tests/
-
+            --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+            --report-log=tests_pipeline_${{ matrix.module }}_cuda.log \ 
+            tests/pipelines/${{ matrix.module }}
+      
       - name: Failure short reports
         if: ${{ failure() }}
-        run: cat reports/tests_${{ matrix.config.report }}_failures_short.txt
+        run: |
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_stats.txt
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_failures_short.txt
 
       - name: Test suite reports artifacts
         if: ${{ always() }}
         uses: actions/upload-artifact@v2
         with:
-          name: ${{ matrix.config.report }}_test_reports
+          name: pipeline_${{ matrix.module }}_test_reports
           path: reports
       
       - name: Generate Report and Notify Channel
@@ -124,6 +110,248 @@ jobs:
           pip install slack_sdk tabulate
           python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
 
+  run_nightly_tests_for_other_torch_modules:
+    name: Torch Non-Pipelines CUDA Nightly Tests
+    runs-on: docker-gpu
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+    defaults:
+      run:
+        shell: bash
+    strategy:
+      matrix:
+        module: [models, schedulers, others, examples]
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        apt-get update && apt-get install libsndfile1-dev libgl1 -y
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install pytest-reportlog
+
+    - name: Environment
+      run: python utils/print_env.py
+
+    - name: Run nightly PyTorch CUDA tests for non-pipeline modules
+      if: ${{ matrix.module != 'examples'}} 
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+        CUBLAS_WORKSPACE_CONFIG: :16:8
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "not Flax and not Onnx" \
+          --make-reports=tests_torch_${{ matrix.module }}_cuda \
+          --report-log=tests_torch_${{ matrix.module }}_cuda.log \ 
+          tests/${{ matrix.module }}
+
+    - name: Run nightly example tests with Torch
+      if: ${{ matrix.module == 'examples' }}
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+        CUBLAS_WORKSPACE_CONFIG: :16:8
+      run: |
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v --make-reports=examples_torch_cuda \
+          --report-log=examples_torch_cuda.log \ 
+          examples/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_torch_${{ matrix.module }}_cuda_stats.txt 
+        cat reports/tests_torch_${{ matrix.module }}_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: torch_${{ matrix.module }}_cuda_test_reports
+        path: reports
+
+    - name: Generate Report and Notify Channel
+      if: always()
+      run: |
+        pip install slack_sdk tabulate
+        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_lora_nightly_tests:
+    name: Nightly LoRA Tests with PEFT and TORCH
+    runs-on: docker-gpu
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --gpus 0
+    defaults:
+      run:
+        shell: bash
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        apt-get update && apt-get install libsndfile1-dev libgl1 -y
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
+        python -m uv pip install pytest-reportlog
+
+    - name: Environment
+      run: python utils/print_env.py
+
+    - name: Run nightly LoRA tests with PEFT and Torch
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+        CUBLAS_WORKSPACE_CONFIG: :16:8
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "not Flax and not Onnx" \
+          --make-reports=tests_torch_lora_cuda \
+          --report-log=tests_torch_lora_cuda.log \ 
+          tests/lora
+    
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_torch_lora_cuda_stats.txt 
+        cat reports/tests_torch_lora_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: torch_lora_cuda_test_reports
+        path: reports
+
+    - name: Generate Report and Notify Channel
+      if: always()
+      run: |
+        pip install slack_sdk tabulate
+        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+  
+  run_flax_tpu_tests:
+    name: Nightly Flax TPU Tests
+    runs-on: docker-tpu
+    container:
+      image: diffusers/diffusers-flax-tpu
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --privileged
+    defaults:
+      run:
+        shell: bash
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        apt-get update && apt-get install libsndfile1-dev libgl1 -y
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install pytest-reportlog
+
+    - name: Environment
+      run: python utils/print_env.py
+
+    - name: Run nightly Flax TPU tests
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      run: |
+        python -m pytest -n 0 \
+          -s -v -k "Flax" \
+          --make-reports=tests_flax_tpu \
+          --report-log=tests_flax_tpu.log \ 
+          tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_flax_tpu_stats.txt
+        cat reports/tests_flax_tpu_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: flax_tpu_test_reports
+        path: reports
+
+    - name: Generate Report and Notify Channel
+      if: always()
+      run: |
+        pip install slack_sdk tabulate
+        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_nightly_onnx_tests:
+    name: Nightly ONNXRuntime CUDA tests on Ubuntu
+    runs-on: docker-gpu
+    container:
+      image: diffusers/diffusers-onnxruntime-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
+    
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: nvidia-smi
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install pytest-reportlog
+
+    - name: Environment
+      run: python utils/print_env.py
+    
+    - name: Run nightly ONNXRuntime CUDA tests
+      env:
+        HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "Onnx" \
+          --make-reports=tests_onnx_cuda \
+          --report-log=tests_onnx_cuda.log \ 
+          tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_onnx_cuda_stats.txt
+        cat reports/tests_onnx_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: ${{ matrix.config.report }}_test_reports
+        path: reports
+    
+    - name: Generate Report and Notify Channel
+      if: always()
+      run: |
+        pip install slack_sdk tabulate
+        python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+
   run_nightly_tests_apple_m1:
     name: Nightly PyTorch MPS tests on MacOS
     runs-on: [ self-hosted, apple-m1 ]

.github/workflows/pr_test_peft_backend.yml

@@ -35,6 +35,10 @@ jobs:
         run: |
           ruff check examples tests src utils scripts
           ruff format examples tests src utils scripts --check
+      - 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
@@ -54,6 +58,10 @@ jobs:
           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]
@@ -105,4 +113,4 @@ jobs:
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
           -s -v \
           --make-reports=tests_${{ matrix.config.report }} \
-          tests/lora/test_lora_layers_peft.py
+          tests/lora/

.github/workflows/pr_tests.yml

@@ -43,6 +43,10 @@ jobs:
         run: |
           ruff check examples tests src utils scripts
           ruff format examples tests src utils scripts --check
+      - 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
@@ -62,6 +66,10 @@ jobs:
           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]

.github/workflows/push_tests.yml

@@ -21,10 +21,7 @@ env:
 jobs:
   setup_torch_cuda_pipeline_matrix:
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
-    container:
-      image: diffusers/diffusers-pytorch-cpu # this is a CPU image, but we need it to fetch the matrix
-      options: --shm-size "16gb" --ipc host
+    runs-on: ubuntu-latest
     outputs:
       pipeline_test_matrix: ${{ steps.fetch_pipeline_matrix.outputs.pipeline_test_matrix }}
     steps:
@@ -32,24 +29,20 @@ jobs:
         uses: actions/checkout@v3
         with:
           fetch-depth: 2
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
       - name: Install dependencies
         run: |
-          apt-get update && apt-get install libsndfile1-dev libgl1 -y
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m uv pip install -e [quality,test]
-          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
-
-      - name: Environment
-        run: |
-          python utils/print_env.py
-
+          pip install -e .
+          pip install huggingface_hub
       - name: Fetch Pipeline Matrix
         id: fetch_pipeline_matrix
         run: |
           matrix=$(python utils/fetch_torch_cuda_pipeline_test_matrix.py)
           echo $matrix
           echo "pipeline_test_matrix=$matrix" >> $GITHUB_OUTPUT
-
       - name: Pipeline Tests Artifacts
         if: ${{ always() }}
         uses: actions/upload-artifact@v2

CITATION.cff

@@ -19,6 +19,16 @@ authors:
     family-names: Rasul
   - given-names: Mishig
     family-names: Davaadorj
+  - given-names: Dhruv
+    family-names: Nair
+  - given-names: Sayak
+    family-names: Paul
+  - given-names: Steven
+    family-names: Liu
+  - given-names: William
+    family-names: Berman
+  - given-names: Yiyi
+    family-names: Xu
   - given-names: Thomas
     family-names: Wolf
 repository-code: 'https://github.com/huggingface/diffusers'

README.md

@@ -238,7 +238,7 @@ We also want to thank @heejkoo for the very helpful overview of papers, code and
 
 ```bibtex
 @misc{von-platen-etal-2022-diffusers,
-  author = {Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Thomas Wolf},
+  author = {Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Dhruv Nair and Sayak Paul and William Berman and Yiyi Xu and Steven Liu and Thomas Wolf},
   title = {Diffusers: State-of-the-art diffusion models},
   year = {2022},
   publisher = {GitHub},

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

@@ -408,6 +408,29 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers)
 
 </Tip>
 
+<table>
+    <tr>
+      <th align=center>Without FreeInit enabled</th>
+      <th align=center>With FreeInit enabled</th>
+    </tr>
+    <tr>
+        <td align=center>
+          panda playing a guitar
+          <br />
+          <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-no-freeinit.gif"
+              alt="panda playing a guitar"
+              style="width: 300px;" />
+        </td>
+        <td align=center>
+          panda playing a guitar
+          <br/>
+          <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-freeinit.gif"
+              alt="panda playing a guitar"
+              style="width: 300px;" />
+        </td>
+    </tr>
+</table>
+
 ## Using AnimateLCM
 
 [AnimateLCM](https://animatelcm.github.io/) is a motion module checkpoint and an [LCM LoRA](https://huggingface.co/docs/diffusers/using-diffusers/inference_with_lcm_lora) that have been created using a consistency learning strategy that decouples the distillation of the image generation priors and the motion generation priors.

docs/source/en/training/controlnet.md

@@ -88,7 +88,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/custom_diffusion.md

@@ -54,7 +54,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -84,7 +84,7 @@ Many of the basic parameters are described in the [DreamBooth](dreambooth#script
 - `--freeze_model`: freezes the key and value parameters in the cross-attention layer; the default is `crossattn_kv`, but you can set it to `crossattn` to train all the parameters in the cross-attention layer
 - `--concepts_list`: to learn multiple concepts, provide a path to a JSON file containing the concepts
 - `--modifier_token`: a special word used to represent the learned concept
-- `--initializer_token`:
+- `--initializer_token`: a special word used to initialize the embeddings of the `modifier_token`
 
 ### Prior preservation loss
 

docs/source/en/training/dreambooth.md

@@ -67,7 +67,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -180,7 +180,7 @@ elif args.pretrained_model_name_or_path:
         revision=args.revision,
         use_fast=False,
     )
-    
+
 # Load scheduler and models
 noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
 text_encoder = text_encoder_cls.from_pretrained(

docs/source/en/training/instructpix2pix.md

@@ -51,7 +51,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -89,7 +89,7 @@ The dataset preprocessing code and training loop are found in the [`main()`](htt
 
 As with the script parameters, a walkthrough of the training script is provided in the [Text-to-image](text2image#training-script) training guide. Instead, this guide takes a look at the InstructPix2Pix relevant parts of the script.
 
-The script begins by modifing the [number of input channels](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L445) in the first convolutional layer of the UNet to account for InstructPix2Pix's additional conditioning image:
+The script begins by modifying the [number of input channels](https://github.com/huggingface/diffusers/blob/64603389da01082055a901f2883c4810d1144edb/examples/instruct_pix2pix/train_instruct_pix2pix.py#L445) in the first convolutional layer of the UNet to account for InstructPix2Pix's additional conditioning image:
 
 ```py
 in_channels = 8

docs/source/en/training/kandinsky.md

@@ -59,7 +59,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -235,7 +235,7 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image_prior.py \
   --validation_prompts="A robot pokemon, 4k photo" \
   --report_to="wandb" \
   --push_to_hub \
-  --output_dir="kandi2-prior-pokemon-model" 
+  --output_dir="kandi2-prior-pokemon-model"
 ```
 
 </hfoption>
@@ -259,7 +259,7 @@ accelerate launch --mixed_precision="fp16"  train_text_to_image_decoder.py \
   --validation_prompts="A robot pokemon, 4k photo" \
   --report_to="wandb" \
   --push_to_hub \
-  --output_dir="kandi2-decoder-pokemon-model" 
+  --output_dir="kandi2-decoder-pokemon-model"
 ```
 
 </hfoption>

docs/source/en/training/lcm_distill.md

@@ -53,7 +53,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -252,4 +252,4 @@ The SDXL training script is discussed in more detail in the [SDXL training](sdxl
 Congratulations on distilling a LCM model! To learn more about LCM, the following may be helpful:
 
 - Learn how to use [LCMs for inference](../using-diffusers/lcm) for text-to-image, image-to-image, and with LoRA checkpoints.
-- Read the [SDXL in 4 steps with Latent Consistency LoRAs](https://huggingface.co/blog/lcm_lora) blog post to learn more about SDXL LCM-LoRA's for super fast inference, quality comparisons, benchmarks, and more.
+- Read the [SDXL in 4 steps with Latent Consistency LoRAs](https://huggingface.co/blog/lcm_lora) blog post to learn more about SDXL LCM-LoRA's for super fast inference, quality comparisons, benchmarks, and more.

docs/source/en/training/sdxl.md

@@ -59,7 +59,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/t2i_adapters.md

@@ -53,7 +53,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/text2image.md

@@ -69,7 +69,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/text_inversion.md

@@ -67,7 +67,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/unconditional_training.md

@@ -51,7 +51,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()

docs/source/en/training/wuerstchen.md

@@ -53,7 +53,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```bash
+```py
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -173,7 +173,7 @@ pipeline = AutoPipelineForText2Image.from_pretrained("path/to/saved/model", torc
 
 caption = "A cute bird pokemon holding a shield"
 images = pipeline(
-    caption, 
+    caption,
     width=1024,
     height=1536,
     prior_timesteps=DEFAULT_STAGE_C_TIMESTEPS,

docs/source/en/tutorials/using_peft_for_inference.md

@@ -45,7 +45,7 @@ Make sure to include the token `toy_face` in the prompt and then you can perform
 ```python
 prompt = "toy_face of a hacker with a hoodie"
 
-lora_scale= 0.9
+lora_scale = 0.9
 image = pipe(
     prompt, num_inference_steps=30, cross_attention_kwargs={"scale": lora_scale}, generator=torch.manual_seed(0)
 ).images[0]
@@ -114,7 +114,7 @@ To return to only using one adapter, use the [`~diffusers.loaders.UNet2DConditio
 pipe.set_adapters("toy")
 
 prompt = "toy_face of a hacker with a hoodie"
-lora_scale= 0.9
+lora_scale = 0.9
 image = pipe(
     prompt, num_inference_steps=30, cross_attention_kwargs={"scale": lora_scale}, generator=torch.manual_seed(0)
 ).images[0]
@@ -127,11 +127,68 @@ Or to disable all adapters entirely, use the [`~diffusers.loaders.UNet2DConditio
 pipe.disable_lora()
 
 prompt = "toy_face of a hacker with a hoodie"
-lora_scale= 0.9
 image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
 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 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
+pipe.enable_lora()  # enable lora again, after we disabled it above
+prompt = "toy_face of a hacker with a hoodie, pixel art"
+adapter_weight_scales = { "unet": { "down": 1, "mid": 0, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-down](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_down.png)
+
+Let's see how turning off the `down` part and turning on the `mid` and `up` part respectively changes the image.
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 1, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-mid](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mid.png)
+
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 0, "up": 1} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-up](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_up.png)
+
+Looks cool!
+
+This is a really powerful feature. You can use it to control the adapter strengths down to per-transformer level. And you can even use it for multiple adapters.
+```python
+adapter_weight_scales_toy = 0.5
+adapter_weight_scales_pixel = {
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # because, in this example, "mid" is not given, all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters(["toy", "pixel"], [adapter_weight_scales_toy, adapter_weight_scales_pixel])
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+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
 
 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.LoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:

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

@@ -239,5 +239,7 @@ pipeline.to("cuda")
 prompt = "柴犬、カラフルアート"
 
 image = pipeline(prompt=prompt).images[0]
+```
 
-```
+> [!TIP]
+> When using `trust_remote_code=True`, it is also strongly encouraged to pass a commit hash as a `revision` to make sure the author of the models did not update the code with some malicious new lines (unless you fully trust the authors of the models).

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

@@ -60,6 +60,23 @@ repo_id = "runwayml/stable-diffusion-v1-5"
 pipe = StableDiffusionImg2ImgPipeline.from_pretrained(repo_id)
 ```
 
+You can use the Space below to gauge the memory requirements of a pipeline you want to load beforehand without downloading the pipeline checkpoints:
+
+<div class="block dark:hidden">
+	<iframe 
+        src="https://diffusers-compute-pipeline-size.hf.space?__theme=light"
+        width="850"
+        height="1600"
+    ></iframe>
+</div>
+<div class="hidden dark:block">
+    <iframe 
+        src="https://diffusers-compute-pipeline-size.hf.space?__theme=dark"
+        width="850"
+        height="1600"
+    ></iframe>
+</div>
+
 ### Local pipeline
 
 To load a diffusion pipeline locally, use [`git-lfs`](https://git-lfs.github.com/) to manually download the checkpoint (in this case, [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5)) to your local disk. This creates a local folder, `./stable-diffusion-v1-5`, on your disk:

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

@@ -153,18 +153,43 @@ image
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
 </div>
 
-<Tip>
-
-For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
-
-</Tip>
-
 To unload the LoRA weights, use the [`~loaders.LoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
 
 ```py
 pipeline.unload_lora_weights()
 ```
 
+### Adjust LoRA weight scale
+
+For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
+
+For more granular control on the amount of LoRA weights used per layer, you can use [`~loaders.LoraLoaderMixin.set_adapters`] and pass a dictionary specifying by how much to scale the weights in each layer by.
+```python
+pipe = ... # create pipeline
+pipe.load_lora_weights(..., adapter_name="my_adapter") 
+scales = {
+    "text_encoder": 0.5,
+    "text_encoder_2": 0.5,  # only usable if pipe has a 2nd text encoder
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # in this example "mid" is not given, therefore all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters("my_adapter", scales)
+```
+
+This also works with multiple adapters - see [this guide](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference#customize-adapters-strength) for how to do it.
+
+<Tip warning={true}>
+
+Currently, [`~loaders.LoraLoaderMixin.set_adapters`] only supports scaling attention weights. If a LoRA has other parts (e.g., resnets or down-/upsamplers), they will keep a scale of 1.0.
+
+</Tip>
+
 ### Kohya and TheLastBen
 
 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.

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

@@ -21,7 +21,7 @@ This guide will show you how to use SVD to generate short videos from images.
 Before you begin, make sure you have the following libraries installed:
 
 ```py
-!pip install -q -U diffusers transformers accelerate 
+!pip install -q -U diffusers transformers accelerate
 ```
 
 The are two variants of this model, [SVD](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid) and [SVD-XT](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt). The SVD checkpoint is trained to generate 14 frames and the SVD-XT checkpoint is further finetuned to generate 25 frames.
@@ -86,7 +86,7 @@ Video generation is very memory intensive because you're essentially generating
 + frames = pipe(image, decode_chunk_size=2, generator=generator, num_frames=25).frames[0]
 ```
 
-Using all these tricks togethere should lower the memory requirement to less than 8GB VRAM.
+Using all these tricks together should lower the memory requirement to less than 8GB VRAM.
 
 ## Micro-conditioning
 

docs/source/ko/optimization/fp16.md

@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # 메모리와 속도
 
-메모리 또는 속도에 대해 🤗 Diffusers *추론*을 최적화하기 위한 몇 가지 기술과 아이디어를 제시합니다. 
+메모리 또는 속도에 대해 🤗 Diffusers *추론*을 최적화하기 위한 몇 가지 기술과 아이디어를 제시합니다.
 일반적으로, memory-efficient attention을 위해 [xFormers](https://github.com/facebookresearch/xformers) 사용을 추천하기 때문에, 추천하는 [설치 방법](xformers)을 보고 설치해 보세요.
 
 다음 설정이 성능과 메모리에 미치는 영향에 대해 설명합니다.
@@ -27,7 +27,7 @@ specific language governing permissions and limitations under the License.
 | memory-efficient attention | 2.63s  | x3.61   |
 
 <em>
-   NVIDIA TITAN RTX에서 50 DDIM 스텝의 "a photo of an astronaut riding a horse on mars" 프롬프트로 512x512 크기의 단일 이미지를 생성하였습니다. 
+   NVIDIA TITAN RTX에서 50 DDIM 스텝의 "a photo of an astronaut riding a horse on mars" 프롬프트로 512x512 크기의 단일 이미지를 생성하였습니다.
 </em>
 
 ## cuDNN auto-tuner 활성화하기
@@ -44,11 +44,11 @@ torch.backends.cudnn.benchmark = True
 
 ### fp32 대신 tf32 사용하기  (Ampere 및 이후 CUDA 장치들에서)
 
-Ampere 및 이후 CUDA 장치에서 행렬곱 및 컨볼루션은 TensorFloat32(TF32) 모드를 사용하여 더 빠르지만 약간 덜 정확할 수 있습니다. 
-기본적으로 PyTorch는 컨볼루션에 대해 TF32 모드를 활성화하지만 행렬 곱셈은 활성화하지 않습니다. 
-네트워크에 완전한 float32 정밀도가 필요한 경우가 아니면 행렬 곱셈에 대해서도 이 설정을 활성화하는 것이 좋습니다. 
-이는 일반적으로 무시할 수 있는 수치의 정확도 손실이 있지만, 계산 속도를 크게 높일 수 있습니다. 
-그것에 대해 [여기](https://huggingface.co/docs/transformers/v4.18.0/en/performance#tf32)서 더 읽을 수 있습니다. 
+Ampere 및 이후 CUDA 장치에서 행렬곱 및 컨볼루션은 TensorFloat32(TF32) 모드를 사용하여 더 빠르지만 약간 덜 정확할 수 있습니다.
+기본적으로 PyTorch는 컨볼루션에 대해 TF32 모드를 활성화하지만 행렬 곱셈은 활성화하지 않습니다.
+네트워크에 완전한 float32 정밀도가 필요한 경우가 아니면 행렬 곱셈에 대해서도 이 설정을 활성화하는 것이 좋습니다.
+이는 일반적으로 무시할 수 있는 수치의 정확도 손실이 있지만, 계산 속도를 크게 높일 수 있습니다.
+그것에 대해 [여기](https://huggingface.co/docs/transformers/v4.18.0/en/performance#tf32)서 더 읽을 수 있습니다.
 추론하기 전에 다음을 추가하기만 하면 됩니다:
 
 ```python
@@ -59,13 +59,13 @@ torch.backends.cuda.matmul.allow_tf32 = True
 
 ## 반정밀도 가중치
 
-더 많은 GPU 메모리를 절약하고 더 빠른 속도를 얻기 위해 모델 가중치를 반정밀도(half precision)로 직접 불러오고 실행할 수 있습니다. 
+더 많은 GPU 메모리를 절약하고 더 빠른 속도를 얻기 위해 모델 가중치를 반정밀도(half precision)로 직접 불러오고 실행할 수 있습니다.
 여기에는 `fp16`이라는 브랜치에 저장된 float16 버전의 가중치를 불러오고, 그 때 `float16` 유형을 사용하도록 PyTorch에 지시하는 작업이 포함됩니다.
 
 ```Python
 pipe = StableDiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-v1-5",
-    
+
     torch_dtype=torch.float16,
 )
 pipe = pipe.to("cuda")
@@ -75,7 +75,7 @@ image = pipe(prompt).images[0]
 ```
 
 <Tip warning={true}>
-  어떤 파이프라인에서도 [`torch.autocast`](https://pytorch.org/docs/stable/amp.html#torch.autocast) 를 사용하는 것은 검은색 이미지를 생성할 수 있고, 순수한 float16 정밀도를 사용하는 것보다 항상 느리기 때문에 사용하지 않는 것이 좋습니다. 
+  어떤 파이프라인에서도 [`torch.autocast`](https://pytorch.org/docs/stable/amp.html#torch.autocast) 를 사용하는 것은 검은색 이미지를 생성할 수 있고, 순수한 float16 정밀도를 사용하는 것보다 항상 느리기 때문에 사용하지 않는 것이 좋습니다.
 </Tip>
 
 ## 추가 메모리 절약을 위한 슬라이스 어텐션
@@ -95,7 +95,7 @@ from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-v1-5",
-    
+
     torch_dtype=torch.float16,
 )
 pipe = pipe.to("cuda")
@@ -122,7 +122,7 @@ from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-v1-5",
-    
+
     torch_dtype=torch.float16,
 )
 pipe = pipe.to("cuda")
@@ -148,7 +148,7 @@ from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-v1-5",
-    
+
     torch_dtype=torch.float16,
 )
 
@@ -165,7 +165,7 @@ image = pipe(prompt).images[0]
 또 다른 최적화 방법인 <a href="#model_offloading">모델 오프로딩</a>을 사용하는 것을 고려하십시오. 이는 훨씬 빠르지만 메모리 절약이 크지는 않습니다.
 </Tip>
 
-또한 ttention slicing과 연결해서 최소 메모리(< 2GB)로도 동작할 수 있습니다. 
+또한 ttention slicing과 연결해서 최소 메모리(< 2GB)로도 동작할 수 있습니다.
 
 
 ```Python
@@ -174,7 +174,7 @@ from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
     "runwayml/stable-diffusion-v1-5",
-    
+
     torch_dtype=torch.float16,
 )
 
@@ -204,7 +204,7 @@ import torch
 from diffusers import StableDiffusionPipeline
 
 pipe = StableDiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5",  
+    "runwayml/stable-diffusion-v1-5",
     torch_dtype=torch.float16,
 )
 
@@ -355,7 +355,7 @@ unet_traced = torch.jit.load("unet_traced.pt")
 class TracedUNet(torch.nn.Module):
     def __init__(self):
         super().__init__()
-        self.in_channels = pipe.unet.in_channels
+        self.in_channels = pipe.unet.config.in_channels
         self.device = pipe.unet.device
 
     def forward(self, latent_model_input, t, encoder_hidden_states):
@@ -387,7 +387,7 @@ with torch.inference_mode():
 | A100-SXM4-40GB    	| 18.6it/s            	| 29.it/s                        	|
 | A100-SXM-80GB    	| 18.7it/s            	| 29.5it/s                        	|
 
-이를 활용하려면 다음을 만족해야 합니다: 
+이를 활용하려면 다음을 만족해야 합니다:
  - PyTorch > 1.12
  - Cuda 사용 가능
  - [xformers 라이브러리를 설치함](xformers)

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

@@ -14,7 +14,7 @@ specific language governing permissions and limitations under the License.
 
 [[open-in-colab]]
 
-🧨 Diffusers는 사용자 친화적이며 유연한 도구 상자로, 사용사례에 맞게 diffusion 시스템을 구축 할 수 있도록 설계되었습니다. 이 도구 상자의 핵심은 모델과 스케줄러입니다. [`DiffusionPipeline`]은 편의를 위해 이러한 구성 요소를 번들로 제공하지만, 파이프라인을 분리하고 모델과 스케줄러를 개별적으로 사용해 새로운 diffusion 시스템을 만들 수도 있습니다. 
+🧨 Diffusers는 사용자 친화적이며 유연한 도구 상자로, 사용사례에 맞게 diffusion 시스템을 구축 할 수 있도록 설계되었습니다. 이 도구 상자의 핵심은 모델과 스케줄러입니다. [`DiffusionPipeline`]은 편의를 위해 이러한 구성 요소를 번들로 제공하지만, 파이프라인을 분리하고 모델과 스케줄러를 개별적으로 사용해 새로운 diffusion 시스템을 만들 수도 있습니다.
 
 이 튜토리얼에서는 기본 파이프라인부터 시작해 Stable Diffusion 파이프라인까지 진행하며 모델과 스케줄러를 사용해 추론을 위한 diffusion 시스템을 조립하는 방법을 배웁니다.
 
@@ -36,7 +36,7 @@ specific language governing permissions and limitations under the License.
 
 정말 쉽습니다. 그런데 파이프라인은 어떻게 이렇게 할 수 있었을까요? 파이프라인을 세분화하여 내부에서 어떤 일이 일어나고 있는지 살펴보겠습니다.
 
-위 예시에서 파이프라인에는 [`UNet2DModel`] 모델과 [`DDPMScheduler`]가 포함되어 있습니다. 파이프라인은 원하는 출력 크기의 랜덤 노이즈를 받아 모델을 여러번 통과시켜 이미지의 노이즈를 제거합니다. 각 timestep에서 모델은 *noise residual*을 예측하고 스케줄러는 이를 사용하여 노이즈가 적은 이미지를 예측합니다. 파이프라인은 지정된 추론 스텝수에 도달할 때까지 이 과정을 반복합니다. 
+위 예시에서 파이프라인에는 [`UNet2DModel`] 모델과 [`DDPMScheduler`]가 포함되어 있습니다. 파이프라인은 원하는 출력 크기의 랜덤 노이즈를 받아 모델을 여러번 통과시켜 이미지의 노이즈를 제거합니다. 각 timestep에서 모델은 *noise residual*을 예측하고 스케줄러는 이를 사용하여 노이즈가 적은 이미지를 예측합니다. 파이프라인은 지정된 추론 스텝수에 도달할 때까지 이 과정을 반복합니다.
 
 모델과 스케줄러를 별도로 사용하여 파이프라인을 다시 생성하기 위해 자체적인 노이즈 제거 프로세스를 작성해 보겠습니다.
 
@@ -210,7 +210,7 @@ Stable Diffusion 은 text-to-image *latent diffusion* 모델입니다. latent di
 
 ```py
 >>> latents = torch.randn(
-...     (batch_size, unet.in_channels, height // 8, width // 8),
+...     (batch_size, unet.config.in_channels, height // 8, width // 8),
 ...     generator=generator,
 ...     device=torch_device,
 ... )

examples/README.md

@@ -42,7 +42,7 @@ Training examples show how to pretrain or fine-tune diffusion models for a varie
 | [**Dreambooth**](./dreambooth) | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_dreambooth_training.ipynb)
 | [**ControlNet**](./controlnet) | ✅ | ✅ | -
 | [**InstructPix2Pix**](./instruct_pix2pix) | ✅ | ✅ | -
-| [**Reinforcement Learning for Control**](https://github.com/huggingface/diffusers/blob/main/examples/reinforcement_learning/run_diffusers_locomotion.py)                    | - | - | coming soon.
+| [**Reinforcement Learning for Control**](./reinforcement_learning)                    | - | - | coming soon.
 
 ## Community
 

examples/advanced_diffusion_training/README.md

@@ -308,6 +308,6 @@ accelerate launch train_dreambooth_lora_sdxl_advanced.py \
 Check out [these recommended practices](https://huggingface.co/blog/sdxl_lora_advanced_script#additional-good-practices)
 
 ## Running on Colab Notebook
-Check out [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_DreamBooth_LoRA_advanced_example.ipynb). 
+Check out [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_Dreambooth_LoRA_advanced_example.ipynb).
 to train using the advanced features (including pivotal tuning), and [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_DreamBooth_LoRA_.ipynb) to train on a free colab, using some of the advanced features (excluding pivotal tuning)
 

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -656,7 +656,6 @@ def parse_args(input_args=None):
     )
     parser.add_argument(
         "--use_dora",
-        type=bool,
         action="store_true",
         default=False,
         help=(

examples/community/README.md

@@ -1,10 +1,12 @@
-# Community Examples
+# Community Pipeline Examples
 
 > **For more information about community pipelines, please have a look at [this issue](https://github.com/huggingface/diffusers/issues/841).**
 
-**Community** examples consist of both inference and training examples that have been added by the community.
-Please have a look at the following table to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste ready code example that you can try out.
-If a community doesn't work as expected, please open an issue and ping the author on it.
+**Community pipeline** examples consist pipelines that have been added by the community.
+Please have a look at the following tables to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste ready code example that you can try out.
+If a community pipeline doesn't work as expected, please open an issue and ping the author on it.
+
+Please also check out our [Community Scripts](https://github.com/huggingface/diffusers/blob/main/examples/community/README_community_scripts.md) examples for tips and tricks that you can use with diffusers without having to run a community pipeline.
 
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
@@ -83,14 +85,25 @@ This depth estimation pipeline processes a single input image through multiple d
 
 ```python
 import numpy as np
+import torch
 from PIL import Image
 from diffusers import DiffusionPipeline
 from diffusers.utils import load_image
 
+# Original DDIM version (higher quality)
 pipe = DiffusionPipeline.from_pretrained(
-    "Bingxin/Marigold",
+    "prs-eth/marigold-v1-0",
     custom_pipeline="marigold_depth_estimation"
     # torch_dtype=torch.float16,                # (optional) Run with half-precision (16-bit float).
+    # variant="fp16",                           # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
+)
+
+# (New) LCM version (faster speed)
+pipe = DiffusionPipeline.from_pretrained(
+    "prs-eth/marigold-lcm-v1-0",
+    custom_pipeline="marigold_depth_estimation"
+    # torch_dtype=torch.float16,                # (optional) Run with half-precision (16-bit float).
+    # variant="fp16",                           # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
 )
 
 pipe.to("cuda")
@@ -99,12 +112,21 @@ img_path_or_url = "https://share.phys.ethz.ch/~pf/bingkedata/marigold/pipeline_e
 image: Image.Image = load_image(img_path_or_url)
 
 pipeline_output = pipe(
-    image,                  # Input image.
+    image,                    # Input image.
+    # ----- recommended setting for DDIM version -----
     # denoising_steps=10,     # (optional) Number of denoising steps of each inference pass. Default: 10.
     # ensemble_size=10,       # (optional) Number of inference passes in the ensemble. Default: 10.
+    # ------------------------------------------------
+    
+    # ----- recommended setting for LCM version ------
+    # denoising_steps=4,
+    # ensemble_size=5,
+    # -------------------------------------------------
+    
     # processing_res=768,     # (optional) Maximum resolution of processing. If set to 0: will not resize at all. Defaults to 768.
     # match_input_res=True,   # (optional) Resize depth prediction to match input resolution.
     # batch_size=0,           # (optional) Inference batch size, no bigger than `num_ensemble`. If set to 0, the script will automatically decide the proper batch size. Defaults to 0.
+    # seed=2024,              # (optional) Random seed can be set to ensure additional reproducibility. Default: None (unseeded). Note: forcing --batch_size 1 helps to increase reproducibility. To ensure full reproducibility, deterministic mode needs to be used.
     # color_map="Spectral",   # (optional) Colormap used to colorize the depth map. Defaults to "Spectral". Set to `None` to skip colormap generation.
     # show_progress_bar=True, # (optional) If true, will show progress bars of the inference progress.
 )
@@ -933,7 +955,7 @@ image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
 ### Checkpoint Merger Pipeline
 Based on the AUTOMATIC1111/webui for checkpoint merging. This is a custom pipeline that merges upto 3 pretrained model checkpoints as long as they are in the HuggingFace model_index.json format.
 
-The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect atleast 13GB RAM Usage on Kaggle GPU kernels and
+The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect at least 13GB RAM Usage on Kaggle GPU kernels and
 on colab you might run out of the 12GB memory even while merging two checkpoints.
 
 Usage:-
@@ -1887,7 +1909,7 @@ In the above code, the `prompt2` is appended to the `prompt`, which is more than
 
 For more results, checkout [PR #6114](https://github.com/huggingface/diffusers/pull/6114).
 
-## Example Images Mixing (with CoCa)
+### Example Images Mixing (with CoCa)
 ```python
 import requests
 from io import BytesIO
@@ -2934,7 +2956,7 @@ pipe(prompt =prompt, rp_args = rp_args)
 
 The Pipeline supports `compel` syntax. Input prompts using the `compel` structure will be automatically applied and processed.
 
-## Diffusion Posterior Sampling Pipeline
+### Diffusion Posterior Sampling Pipeline
 * Reference paper
     ```
     @article{chung2022diffusion,
@@ -3741,3 +3763,80 @@ onestep_image = pipe(prompt, num_inference_steps=1).images[0]
 # Multistep sampling
 multistep_image = pipe(prompt, num_inference_steps=4).images[0]
 ```
+
+# Perturbed-Attention Guidance
+
+[Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)
+
+This implementation is based on [Diffusers](https://huggingface.co/docs/diffusers/index). StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).
+
+## Example Usage
+
+```
+import os
+import torch
+
+from accelerate.utils import set_seed
+
+from diffusers import StableDiffusionPipeline
+from diffusers.utils import load_image, make_image_grid
+from diffusers.utils.torch_utils import randn_tensor
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="hyoungwoncho/sd_perturbed_attention_guidance",
+    torch_dtype=torch.float16
+)
+
+device="cuda"
+pipe = pipe.to(device)
+
+pag_scale = 5.0
+pag_applied_layers_index = ['m0']
+
+batch_size = 4
+seed=10
+
+base_dir = "./results/"
+grid_dir = base_dir + "/pag" + str(pag_scale) + "/"
+
+if not os.path.exists(grid_dir):
+    os.makedirs(grid_dir)
+
+set_seed(seed)
+
+latent_input = randn_tensor(shape=(batch_size,4,64,64),generator=None, device=device, dtype=torch.float16)
+
+output_baseline = pipe(
+    "",
+    width=512,
+    height=512,
+    num_inference_steps=50,
+    guidance_scale=0.0,
+    pag_scale=0.0,
+    pag_applied_layers_index=pag_applied_layers_index,
+    num_images_per_prompt=batch_size,
+    latents=latent_input
+).images
+
+output_pag = pipe(
+    "",
+    width=512,
+    height=512,
+    num_inference_steps=50,
+    guidance_scale=0.0,
+    pag_scale=5.0,
+    pag_applied_layers_index=pag_applied_layers_index,
+    num_images_per_prompt=batch_size,
+    latents=latent_input
+).images
+
+grid_image = make_image_grid(output_baseline + output_pag, rows=2, cols=batch_size)
+grid_image.save(grid_dir + "sample.png")
+```
+
+## PAG Parameters
+
+pag_scale : gudiance scale of PAG (ex: 5.0)
+
+pag_applied_layers_index : index of the layer to apply perturbation (ex: ['m0'])

examples/community/README_community_scripts.md

@@ -0,0 +1,232 @@
+# Community Scripts
+
+**Community scripts** consist of inference examples using Diffusers pipelines that have been added by the community. 
+Please have a look at the following table to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste code example that you can try out.
+If a community script doesn't work as expected, please open an issue and ping the author on it.
+
+| 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)                                   | | [Álvaro Somoza](https://github.com/asomoza)|
+| asymmetric tiling                                                                                                  |configure seamless image tiling independently for the X and Y axes                                                                                                                                                                                                      | [Asymmetric Tiling](#asymmetric-tiling )                                   | | [alexisrolland](https://github.com/alexisrolland)|
+
+
+## Example usages
+
+### IP Adapter Negative Noise
+
+Diffusers pipelines are fully integrated with IP-Adapter, which allows you to prompt the diffusion model with an image. However, it does not support negative image prompts (there is no `negative_ip_adapter_image` argument) the same way it supports negative text prompts. When you pass an `ip_adapter_image,` it will create a zero-filled tensor as a negative image. This script shows you how to create a negative noise from `ip_adapter_image` and use it to significantly improve the generation quality while preserving the composition of images.
+
+[cubiq](https://github.com/cubiq) initially developed this feature in his [repository](https://github.com/cubiq/ComfyUI_IPAdapter_plus). The community script was contributed by [asomoza](https://github.com/Somoza). You can find more details about this experimentation [this discussion](https://github.com/huggingface/diffusers/discussions/7167)
+
+IP-Adapter without negative noise
+|source|result|
+|---|---|
+|![20240229150812](https://github.com/huggingface/diffusers/assets/5442875/901d8bd8-7a59-4fe7-bda1-a0e0d6c7dffd)|![20240229163923_normal](https://github.com/huggingface/diffusers/assets/5442875/3432e25a-ece6-45f4-a3f4-fca354f40b5b)|
+
+IP-Adapter with negative noise
+|source|result|
+|---|---|
+|![20240229150812](https://github.com/huggingface/diffusers/assets/5442875/901d8bd8-7a59-4fe7-bda1-a0e0d6c7dffd)|![20240229163923](https://github.com/huggingface/diffusers/assets/5442875/736fd15a-36ba-40c0-a7d8-6ec1ac26f788)|
+
+```python
+import torch
+
+from diffusers import AutoencoderKL, DPMSolverMultistepScheduler, StableDiffusionXLPipeline
+from diffusers.models import ImageProjection
+from diffusers.utils import load_image
+
+
+def encode_image(
+    image_encoder,
+    feature_extractor,
+    image,
+    device,
+    num_images_per_prompt,
+    output_hidden_states=None,
+    negative_image=None,
+):
+    dtype = next(image_encoder.parameters()).dtype
+
+    if not isinstance(image, torch.Tensor):
+        image = feature_extractor(image, return_tensors="pt").pixel_values
+
+    image = image.to(device=device, dtype=dtype)
+    if output_hidden_states:
+        image_enc_hidden_states = image_encoder(image, output_hidden_states=True).hidden_states[-2]
+        image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+
+        if negative_image is None:
+            uncond_image_enc_hidden_states = image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+        else:
+            if not isinstance(negative_image, torch.Tensor):
+                negative_image = feature_extractor(negative_image, return_tensors="pt").pixel_values
+            negative_image = negative_image.to(device=device, dtype=dtype)
+            uncond_image_enc_hidden_states = image_encoder(negative_image, output_hidden_states=True).hidden_states[-2]
+
+        uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+        return image_enc_hidden_states, uncond_image_enc_hidden_states
+    else:
+        image_embeds = image_encoder(image).image_embeds
+        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        uncond_image_embeds = torch.zeros_like(image_embeds)
+
+        return image_embeds, uncond_image_embeds
+
+
+@torch.no_grad()
+def prepare_ip_adapter_image_embeds(
+    unet,
+    image_encoder,
+    feature_extractor,
+    ip_adapter_image,
+    do_classifier_free_guidance,
+    device,
+    num_images_per_prompt,
+    ip_adapter_negative_image=None,
+):
+    if not isinstance(ip_adapter_image, list):
+        ip_adapter_image = [ip_adapter_image]
+
+    if len(ip_adapter_image) != len(unet.encoder_hid_proj.image_projection_layers):
+        raise ValueError(
+            f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+        )
+
+    image_embeds = []
+    for single_ip_adapter_image, image_proj_layer in zip(
+        ip_adapter_image, unet.encoder_hid_proj.image_projection_layers
+    ):
+        output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+        single_image_embeds, single_negative_image_embeds = encode_image(
+            image_encoder,
+            feature_extractor,
+            single_ip_adapter_image,
+            device,
+            1,
+            output_hidden_state,
+            negative_image=ip_adapter_negative_image,
+        )
+        single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+        single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0)
+
+        if do_classifier_free_guidance:
+            single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+            single_image_embeds = single_image_embeds.to(device)
+
+        image_embeds.append(single_image_embeds)
+
+    return image_embeds
+
+
+vae = AutoencoderKL.from_pretrained(
+    "madebyollin/sdxl-vae-fp16-fix",
+    torch_dtype=torch.float16,
+).to("cuda")
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "RunDiffusion/Juggernaut-XL-v9",
+    torch_dtype=torch.float16,
+    vae=vae,
+    variant="fp16",
+).to("cuda")
+
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
+pipeline.scheduler.config.use_karras_sigmas = True
+
+pipeline.load_ip_adapter(
+    "h94/IP-Adapter",
+    subfolder="sdxl_models",
+    weight_name="ip-adapter-plus_sdxl_vit-h.safetensors",
+    image_encoder_folder="models/image_encoder",
+)
+pipeline.set_ip_adapter_scale(0.7)
+
+ip_image = load_image("source.png")
+negative_ip_image = load_image("noise.png")
+
+image_embeds = prepare_ip_adapter_image_embeds(
+    unet=pipeline.unet,
+    image_encoder=pipeline.image_encoder,
+    feature_extractor=pipeline.feature_extractor,
+    ip_adapter_image=[[ip_image]],
+    do_classifier_free_guidance=True,
+    device="cuda",
+    num_images_per_prompt=1,
+    ip_adapter_negative_image=negative_ip_image,
+)
+
+
+prompt = "cinematic photo of a cyborg in the city, 4k, high quality, intricate, highly detailed"
+negative_prompt = "blurry, smooth, plastic"
+
+image = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    ip_adapter_image_embeds=image_embeds,
+    guidance_scale=6.0,
+    num_inference_steps=25,
+    generator=torch.Generator(device="cpu").manual_seed(1556265306),
+).images[0]
+
+image.save("result.png")
+```
+
+### Asymmetric Tiling
+Stable Diffusion is not trained to generate seamless textures. However, you can use this simple script to add tiling to your generation. This script is contributed by [alexisrolland](https://github.com/alexisrolland). See more details in the [this issue](https://github.com/huggingface/diffusers/issues/556)
+
+
+|Generated|Tiled|
+|---|---|
+|![20240313003235_573631814](https://github.com/huggingface/diffusers/assets/5442875/eca174fb-06a4-464e-a3a7-00dbb024543e)|![wall](https://github.com/huggingface/diffusers/assets/5442875/b4aa774b-2a6a-4316-a8eb-8f30b5f4d024)|
+
+
+```py
+import torch
+from typing import Optional
+from diffusers import StableDiffusionPipeline
+from diffusers.models.lora import LoRACompatibleConv
+
+def seamless_tiling(pipeline, x_axis, y_axis):
+    def asymmetric_conv2d_convforward(self, input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None):
+        self.paddingX = (self._reversed_padding_repeated_twice[0], self._reversed_padding_repeated_twice[1], 0, 0)
+        self.paddingY = (0, 0, self._reversed_padding_repeated_twice[2], self._reversed_padding_repeated_twice[3])
+        working = torch.nn.functional.pad(input, self.paddingX, mode=x_mode)
+        working = torch.nn.functional.pad(working, self.paddingY, mode=y_mode)
+        return torch.nn.functional.conv2d(working, weight, bias, self.stride, torch.nn.modules.utils._pair(0), self.dilation, self.groups)
+    x_mode = 'circular' if x_axis else 'constant'
+    y_mode = 'circular' if y_axis else 'constant'
+    targets = [pipeline.vae, pipeline.text_encoder, pipeline.unet]
+    convolution_layers = []
+    for target in targets:
+        for module in target.modules():
+            if isinstance(module, torch.nn.Conv2d):
+                convolution_layers.append(module)
+    for layer in convolution_layers:
+        if isinstance(layer, LoRACompatibleConv) and layer.lora_layer is None:
+            layer.lora_layer = lambda * x: 0
+        layer._conv_forward = asymmetric_conv2d_convforward.__get__(layer, torch.nn.Conv2d)
+    return pipeline
+
+pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True)
+pipeline.enable_model_cpu_offload()
+prompt = ["texture of a red brick wall"]
+seed = 123456
+generator = torch.Generator(device='cuda').manual_seed(seed)
+
+pipeline = seamless_tiling(pipeline=pipeline, x_axis=True, y_axis=True)
+image = pipeline(
+    prompt=prompt,
+    width=512,
+    height=512,
+    num_inference_steps=20,
+    guidance_scale=7,
+    num_images_per_prompt=1,
+    generator=generator
+).images[0]
+seamless_tiling(pipeline=pipeline, x_axis=False, y_axis=False)
+
+torch.cuda.empty_cache()
+image.save('image.png')
+```

examples/community/checkpoint_merger.py

@@ -103,7 +103,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
         print(f"Combining with alpha={alpha}, interpolation mode={interp}")
 
         checkpoint_count = len(pretrained_model_name_or_path_list)
-        # Ignore result from model_index_json comparision of the two checkpoints
+        # Ignore result from model_index_json comparison of the two checkpoints
         force = kwargs.pop("force", False)
 
         # If less than 2 checkpoints, nothing to merge. If more than 3, not supported for now.
@@ -217,7 +217,7 @@ class CheckpointMergerPipeline(DiffusionPipeline):
                         ]
                         checkpoint_path_2 = files[0] if len(files) > 0 else None
                 # For an attr if both checkpoint_path_1 and 2 are None, ignore.
-                # If atleast one is present, deal with it according to interp method, of course only if the state_dict keys match.
+                # If at least one is present, deal with it according to interp method, of course only if the state_dict keys match.
                 if checkpoint_path_1 is None and checkpoint_path_2 is None:
                     print(f"Skipping {attr}: not present in 2nd or 3d model")
                     continue

examples/community/imagic_stable_diffusion.py

@@ -1,7 +1,8 @@
 """
-    modeled after the textual_inversion.py / train_dreambooth.py and the work
-    of justinpinkney here: https://github.com/justinpinkney/stable-diffusion/blob/main/notebooks/imagic.ipynb
+modeled after the textual_inversion.py / train_dreambooth.py and the work
+of justinpinkney here: https://github.com/justinpinkney/stable-diffusion/blob/main/notebooks/imagic.ipynb
 """
+
 import inspect
 import warnings
 from typing import List, Optional, Union

examples/community/latent_consistency_img2img.py

@@ -440,7 +440,7 @@ def betas_for_alpha_bar(
             return math.exp(t * -12.0)
 
     else:
-        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
 
     betas = []
     for i in range(num_diffusion_timesteps):

examples/community/latent_consistency_interpolate.py

@@ -726,7 +726,7 @@ class LatentConsistencyModelWalkPipeline(
             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 pipeine class.
+                `._callback_tensor_inputs` attribute of your pipeline class.
             embedding_interpolation_type (`str`, *optional*, defaults to `"lerp"`):
                 The type of interpolation to use for interpolating between text embeddings. Choose between `"lerp"` and `"slerp"`.
             latent_interpolation_type (`str`, *optional*, defaults to `"slerp"`):
@@ -779,7 +779,7 @@ class LatentConsistencyModelWalkPipeline(
         else:
             batch_size = prompt_embeds.shape[0]
         if batch_size < 2:
-            raise ValueError(f"`prompt` must have length of atleast 2 but found {batch_size}")
+            raise ValueError(f"`prompt` must have length of at least 2 but found {batch_size}")
         if num_images_per_prompt != 1:
             raise ValueError("`num_images_per_prompt` must be `1` as no other value is supported yet")
         if prompt_embeds is not None:
@@ -883,7 +883,7 @@ class LatentConsistencyModelWalkPipeline(
                 ) as batch_progress_bar:
                     for batch_index in range(0, bs, process_batch_size):
                         batch_inference_latents = inference_latents[batch_index : batch_index + process_batch_size]
-                        batch_inference_embedddings = inference_embeddings[
+                        batch_inference_embeddings = inference_embeddings[
                             batch_index : batch_index + process_batch_size
                         ]
 
@@ -892,7 +892,7 @@ class LatentConsistencyModelWalkPipeline(
                         )
                         timesteps = self.scheduler.timesteps
 
-                        current_bs = batch_inference_embedddings.shape[0]
+                        current_bs = batch_inference_embeddings.shape[0]
                         w = torch.tensor(self.guidance_scale - 1).repeat(current_bs)
                         w_embedding = self.get_guidance_scale_embedding(
                             w, embedding_dim=self.unet.config.time_cond_proj_dim
@@ -901,14 +901,14 @@ class LatentConsistencyModelWalkPipeline(
                         # 10. Perform inference for current batch
                         with self.progress_bar(total=num_inference_steps) as progress_bar:
                             for index, t in enumerate(timesteps):
-                                batch_inference_latents = batch_inference_latents.to(batch_inference_embedddings.dtype)
+                                batch_inference_latents = batch_inference_latents.to(batch_inference_embeddings.dtype)
 
                                 # model prediction (v-prediction, eps, x)
                                 model_pred = self.unet(
                                     batch_inference_latents,
                                     t,
                                     timestep_cond=w_embedding,
-                                    encoder_hidden_states=batch_inference_embedddings,
+                                    encoder_hidden_states=batch_inference_embeddings,
                                     cross_attention_kwargs=self.cross_attention_kwargs,
                                     return_dict=False,
                                 )[0]
@@ -924,8 +924,8 @@ class LatentConsistencyModelWalkPipeline(
                                     callback_outputs = callback_on_step_end(self, index, t, callback_kwargs)
 
                                     batch_inference_latents = callback_outputs.pop("latents", batch_inference_latents)
-                                    batch_inference_embedddings = callback_outputs.pop(
-                                        "prompt_embeds", batch_inference_embedddings
+                                    batch_inference_embeddings = callback_outputs.pop(
+                                        "prompt_embeds", batch_inference_embeddings
                                     )
                                     w_embedding = callback_outputs.pop("w_embedding", w_embedding)
                                     denoised = callback_outputs.pop("denoised", denoised)
@@ -939,7 +939,7 @@ class LatentConsistencyModelWalkPipeline(
                                         step_idx = index // getattr(self.scheduler, "order", 1)
                                         callback(step_idx, t, batch_inference_latents)
 
-                        denoised = denoised.to(batch_inference_embedddings.dtype)
+                        denoised = denoised.to(batch_inference_embeddings.dtype)
 
                         # Note: This is not supported because you would get black images in your latent walk if
                         #       NSFW concept is detected

examples/community/latent_consistency_txt2img.py

@@ -348,7 +348,7 @@ def betas_for_alpha_bar(
             return math.exp(t * -12.0)
 
     else:
-        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
 
     betas = []
     for i in range(num_diffusion_timesteps):

examples/community/llm_grounded_diffusion.py

@@ -530,7 +530,7 @@ class LLMGroundedDiffusionPipeline(
                 )
 
         if len(phrases) != len(boxes):
-            ValueError(
+            raise ValueError(
                 "length of `phrases` and `boxes` has to be same, but"
                 f" got: `phrases` {len(phrases)} != `boxes` {len(boxes)}"
             )

examples/community/lpw_stable_diffusion_xl.py

@@ -164,7 +164,7 @@ def get_prompts_tokens_with_weights(clip_tokenizer: CLIPTokenizer, prompt: str):
         text_tokens (list)
             A list contains token ids
         text_weight (list)
-            A list contains the correspodent weight of token ids
+            A list contains the correspondent weight of token ids
 
     Example:
         import torch
@@ -1028,7 +1028,7 @@ class SDXLLongPromptWeightingPipeline(
                 # because `num_inference_steps` might be even given that every timestep
                 # (except the highest one) is duplicated. If `num_inference_steps` is even it would
                 # mean that we cut the timesteps in the middle of the denoising step
-                # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+                # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1
                 # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
                 num_inference_steps = num_inference_steps + 1
 
@@ -1531,7 +1531,7 @@ class SDXLLongPromptWeightingPipeline(
             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 pipeine class.
+                `._callback_tensor_inputs` attribute of your pipeline class.
 
         Examples:
 
@@ -2131,7 +2131,7 @@ class SDXLLongPromptWeightingPipeline(
             **kwargs,
         )
 
-    # Overrride to properly handle the loading and unloading of the additional text encoder.
+    # Override to properly handle the loading and unloading of the additional text encoder.
     def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL

examples/community/marigold_depth_estimation.py

@@ -18,6 +18,7 @@
 # --------------------------------------------------------------------------
 
 
+import logging
 import math
 from typing import Dict, Union
 
@@ -25,6 +26,7 @@ import matplotlib
 import numpy as np
 import torch
 from PIL import Image
+from PIL.Image import Resampling
 from scipy.optimize import minimize
 from torch.utils.data import DataLoader, TensorDataset
 from tqdm.auto import tqdm
@@ -34,13 +36,14 @@ from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
     DiffusionPipeline,
+    LCMScheduler,
     UNet2DConditionModel,
 )
 from diffusers.utils import BaseOutput, check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.25.0")
 
 
 class MarigoldDepthOutput(BaseOutput):
@@ -61,6 +64,19 @@ class MarigoldDepthOutput(BaseOutput):
     uncertainty: Union[None, np.ndarray]
 
 
+def get_pil_resample_method(method_str: str) -> Resampling:
+    resample_method_dic = {
+        "bilinear": Resampling.BILINEAR,
+        "bicubic": Resampling.BICUBIC,
+        "nearest": Resampling.NEAREST,
+    }
+    resample_method = resample_method_dic.get(method_str, None)
+    if resample_method is None:
+        raise ValueError(f"Unknown resampling method: {resample_method}")
+    else:
+        return resample_method
+
+
 class MarigoldPipeline(DiffusionPipeline):
     """
     Pipeline for monocular depth estimation using Marigold: https://marigoldmonodepth.github.io.
@@ -113,7 +129,9 @@ class MarigoldPipeline(DiffusionPipeline):
         ensemble_size: int = 10,
         processing_res: int = 768,
         match_input_res: bool = True,
+        resample_method: str = "bilinear",
         batch_size: int = 0,
+        seed: Union[int, None] = None,
         color_map: str = "Spectral",
         show_progress_bar: bool = True,
         ensemble_kwargs: Dict = None,
@@ -129,7 +147,9 @@ class MarigoldPipeline(DiffusionPipeline):
                 If set to 0: will not resize at all.
             match_input_res (`bool`, *optional*, defaults to `True`):
                 Resize depth prediction to match input resolution.
-                Only valid if `limit_input_res` is not None.
+                Only valid if `processing_res` > 0.
+            resample_method: (`str`, *optional*, defaults to `bilinear`):
+                Resampling method used to resize images and depth predictions. This can be one of `bilinear`, `bicubic` or `nearest`, defaults to: `bilinear`.
             denoising_steps (`int`, *optional*, defaults to `10`):
                 Number of diffusion denoising steps (DDIM) during inference.
             ensemble_size (`int`, *optional*, defaults to `10`):
@@ -137,6 +157,8 @@ class MarigoldPipeline(DiffusionPipeline):
             batch_size (`int`, *optional*, defaults to `0`):
                 Inference batch size, no bigger than `num_ensemble`.
                 If set to 0, the script will automatically decide the proper batch size.
+            seed (`int`, *optional*, defaults to `None`)
+                Reproducibility seed.
             show_progress_bar (`bool`, *optional*, defaults to `True`):
                 Display a progress bar of diffusion denoising.
             color_map (`str`, *optional*, defaults to `"Spectral"`, pass `None` to skip colorized depth map generation):
@@ -146,8 +168,7 @@ class MarigoldPipeline(DiffusionPipeline):
         Returns:
             `MarigoldDepthOutput`: Output class for Marigold monocular depth prediction pipeline, including:
             - **depth_np** (`np.ndarray`) Predicted depth map, with depth values in the range of [0, 1]
-            - **depth_colored** (`None` or `PIL.Image.Image`) Colorized depth map, with the shape of [3, H, W] and
-                    values in [0, 1]. None if `color_map` is `None`
+            - **depth_colored** (`PIL.Image.Image`) Colorized depth map, with the shape of [3, H, W] and values in [0, 1], None if `color_map` is `None`
             - **uncertainty** (`None` or `np.ndarray`) Uncalibrated uncertainty(MAD, median absolute deviation)
                     coming from ensembling. None if `ensemble_size = 1`
         """
@@ -158,13 +179,21 @@ class MarigoldPipeline(DiffusionPipeline):
         if not match_input_res:
             assert processing_res is not None, "Value error: `resize_output_back` is only valid with "
         assert processing_res >= 0
-        assert denoising_steps >= 1
         assert ensemble_size >= 1
 
+        # Check if denoising step is reasonable
+        self._check_inference_step(denoising_steps)
+
+        resample_method: Resampling = get_pil_resample_method(resample_method)
+
         # ----------------- Image Preprocess -----------------
         # Resize image
         if processing_res > 0:
-            input_image = self.resize_max_res(input_image, max_edge_resolution=processing_res)
+            input_image = self.resize_max_res(
+                input_image,
+                max_edge_resolution=processing_res,
+                resample_method=resample_method,
+            )
         # Convert the image to RGB, to 1.remove the alpha channel 2.convert B&W to 3-channel
         input_image = input_image.convert("RGB")
         image = np.asarray(input_image)
@@ -203,9 +232,10 @@ class MarigoldPipeline(DiffusionPipeline):
                 rgb_in=batched_img,
                 num_inference_steps=denoising_steps,
                 show_pbar=show_progress_bar,
+                seed=seed,
             )
-            depth_pred_ls.append(depth_pred_raw.detach().clone())
-        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze()
+            depth_pred_ls.append(depth_pred_raw.detach())
+        depth_preds = torch.concat(depth_pred_ls, dim=0).squeeze()
         torch.cuda.empty_cache()  # clear vram cache for ensembling
 
         # ----------------- Test-time ensembling -----------------
@@ -227,7 +257,7 @@ class MarigoldPipeline(DiffusionPipeline):
         # Resize back to original resolution
         if match_input_res:
             pred_img = Image.fromarray(depth_pred)
-            pred_img = pred_img.resize(input_size)
+            pred_img = pred_img.resize(input_size, resample=resample_method)
             depth_pred = np.asarray(pred_img)
 
         # Clip output range
@@ -243,12 +273,32 @@ class MarigoldPipeline(DiffusionPipeline):
             depth_colored_img = Image.fromarray(depth_colored_hwc)
         else:
             depth_colored_img = None
+
         return MarigoldDepthOutput(
             depth_np=depth_pred,
             depth_colored=depth_colored_img,
             uncertainty=pred_uncert,
         )
 
+    def _check_inference_step(self, n_step: int):
+        """
+        Check if denoising step is reasonable
+        Args:
+            n_step (`int`): denoising steps
+        """
+        assert n_step >= 1
+
+        if isinstance(self.scheduler, DDIMScheduler):
+            if n_step < 10:
+                logging.warning(
+                    f"Too few denoising steps: {n_step}. Recommended to use the LCM checkpoint for few-step inference."
+                )
+        elif isinstance(self.scheduler, LCMScheduler):
+            if not 1 <= n_step <= 4:
+                logging.warning(f"Non-optimal setting of denoising steps: {n_step}. Recommended setting is 1-4 steps.")
+        else:
+            raise RuntimeError(f"Unsupported scheduler type: {type(self.scheduler)}")
+
     def _encode_empty_text(self):
         """
         Encode text embedding for empty prompt.
@@ -265,7 +315,13 @@ class MarigoldPipeline(DiffusionPipeline):
         self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype)
 
     @torch.no_grad()
-    def single_infer(self, rgb_in: torch.Tensor, num_inference_steps: int, show_pbar: bool) -> torch.Tensor:
+    def single_infer(
+        self,
+        rgb_in: torch.Tensor,
+        num_inference_steps: int,
+        seed: Union[int, None],
+        show_pbar: bool,
+    ) -> torch.Tensor:
         """
         Perform an individual depth prediction without ensembling.
 
@@ -286,10 +342,20 @@ class MarigoldPipeline(DiffusionPipeline):
         timesteps = self.scheduler.timesteps  # [T]
 
         # Encode image
-        rgb_latent = self._encode_rgb(rgb_in)
+        rgb_latent = self.encode_rgb(rgb_in)
 
         # Initial depth map (noise)
-        depth_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype)  # [B, 4, h, w]
+        if seed is None:
+            rand_num_generator = None
+        else:
+            rand_num_generator = torch.Generator(device=device)
+            rand_num_generator.manual_seed(seed)
+        depth_latent = torch.randn(
+            rgb_latent.shape,
+            device=device,
+            dtype=self.dtype,
+            generator=rand_num_generator,
+        )  # [B, 4, h, w]
 
         # Batched empty text embedding
         if self.empty_text_embed is None:
@@ -314,9 +380,9 @@ class MarigoldPipeline(DiffusionPipeline):
             noise_pred = self.unet(unet_input, t, encoder_hidden_states=batch_empty_text_embed).sample  # [B, 4, h, w]
 
             # compute the previous noisy sample x_t -> x_t-1
-            depth_latent = self.scheduler.step(noise_pred, t, depth_latent).prev_sample
-        torch.cuda.empty_cache()
-        depth = self._decode_depth(depth_latent)
+            depth_latent = self.scheduler.step(noise_pred, t, depth_latent, generator=rand_num_generator).prev_sample
+
+        depth = self.decode_depth(depth_latent)
 
         # clip prediction
         depth = torch.clip(depth, -1.0, 1.0)
@@ -325,7 +391,7 @@ class MarigoldPipeline(DiffusionPipeline):
 
         return depth
 
-    def _encode_rgb(self, rgb_in: torch.Tensor) -> torch.Tensor:
+    def encode_rgb(self, rgb_in: torch.Tensor) -> torch.Tensor:
         """
         Encode RGB image into latent.
 
@@ -344,7 +410,7 @@ class MarigoldPipeline(DiffusionPipeline):
         rgb_latent = mean * self.rgb_latent_scale_factor
         return rgb_latent
 
-    def _decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
         """
         Decode depth latent into depth map.
 
@@ -365,7 +431,7 @@ class MarigoldPipeline(DiffusionPipeline):
         return depth_mean
 
     @staticmethod
-    def resize_max_res(img: Image.Image, max_edge_resolution: int) -> Image.Image:
+    def resize_max_res(img: Image.Image, max_edge_resolution: int, resample_method=Resampling.BILINEAR) -> Image.Image:
         """
         Resize image to limit maximum edge length while keeping aspect ratio.
 
@@ -374,6 +440,8 @@ class MarigoldPipeline(DiffusionPipeline):
                 Image to be resized.
             max_edge_resolution (`int`):
                 Maximum edge length (pixel).
+            resample_method (`PIL.Image.Resampling`):
+                Resampling method used to resize images.
 
         Returns:
             `Image.Image`: Resized image.
@@ -384,7 +452,7 @@ class MarigoldPipeline(DiffusionPipeline):
         new_width = int(original_width * downscale_factor)
         new_height = int(original_height * downscale_factor)
 
-        resized_img = img.resize((new_width, new_height))
+        resized_img = img.resize((new_width, new_height), resample=resample_method)
         return resized_img
 
     @staticmethod

examples/community/mixture_tiling.py

@@ -196,7 +196,7 @@ class StableDiffusionTilingPipeline(DiffusionPipeline, StableDiffusionExtrasMixi
             guidance_scale_tiles: specific weights for classifier-free guidance in each tile.
             guidance_scale_tiles: specific weights for classifier-free guidance in each tile. If None, the value provided in guidance_scale will be used.
             seed_tiles: specific seeds for the initialization latents in each tile. These will override the latents generated for the whole canvas using the standard seed parameter.
-            seed_tiles_mode: either "full" "exclusive". If "full", all the latents affected by the tile be overriden. If "exclusive", only the latents that are affected exclusively by this tile (and no other tiles) will be overrriden.
+            seed_tiles_mode: either "full" "exclusive". If "full", all the latents affected by the tile be overriden. If "exclusive", only the latents that are affected exclusively by this tile (and no other tiles) will be overriden.
             seed_reroll_regions: a list of tuples in the form (start row, end row, start column, end column, seed) defining regions in pixel space for which the latents will be overriden using the given seed. Takes priority over seed_tiles.
             cpu_vae: the decoder from latent space to pixel space can require too mucho GPU RAM for large images. If you find out of memory errors at the end of the generation process, try setting this parameter to True to run the decoder in CPU. Slower, but should run without memory issues.
 
@@ -325,7 +325,7 @@ class StableDiffusionTilingPipeline(DiffusionPipeline, StableDiffusionExtrasMixi
         if accepts_eta:
             extra_step_kwargs["eta"] = eta
 
-        # Mask for tile weights strenght
+        # Mask for tile weights strength
         tile_weights = self._gaussian_weights(tile_width, tile_height, batch_size)
 
         # Diffusion timesteps

examples/community/pipeline_animatediff_controlnet.py

@@ -832,7 +832,7 @@ class AnimateDiffControlNetPipeline(
             clip_skip (`int`, *optional*):
                 Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                 the output of the pre-final layer will be used for computing the prompt embeddings.
-            allback_on_step_end (`Callable`, *optional*):
+            callback_on_step_end (`Callable`, *optional*):
                 A function that calls at the end of each denoising steps during the inference. The function is called
                 with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                 callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
@@ -840,7 +840,7 @@ class AnimateDiffControlNetPipeline(
             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 pipeine class.
+                `._callback_tensor_inputs` attribute of your pipeline class.
 
         Examples:
 

examples/community/pipeline_demofusion_sdxl.py

@@ -1280,7 +1280,7 @@ class DemoFusionSDXLPipeline(
 
         return output_images
 
-    # Overrride to properly handle the loading and unloading of the additional text encoder.
+    # Override to properly handle the loading and unloading of the additional text encoder.
     def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
         # We could have accessed the unet config from `lora_state_dict()` too. We pass
         # it here explicitly to be able to tell that it's coming from an SDXL

examples/community/pipeline_sdxl_style_aligned.py

@@ -887,7 +887,7 @@ class StyleAlignedSDXLPipeline(
                 # because `num_inference_steps` might be even given that every timestep
                 # (except the highest one) is duplicated. If `num_inference_steps` is even it would
                 # mean that we cut the timesteps in the middle of the denoising step
-                # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+                # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1
                 # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
                 num_inference_steps = num_inference_steps + 1
 

examples/community/pipeline_stable_diffusion_upscale_ldm3d.py

@@ -26,7 +26,7 @@ from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInver
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
-from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_ldm3d import LDM3DPipelineOutput
+from diffusers.pipelines.stable_diffusion_ldm3d.pipeline_stable_diffusion_ldm3d import LDM3DPipelineOutput
 from diffusers.schedulers import DDPMScheduler, KarrasDiffusionSchedulers
 from diffusers.utils import (
     USE_PEFT_BACKEND,

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -206,7 +206,7 @@ def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool
             dimensions: ``batch x channels x height x width``.
     """
 
-    # checkpoint. TOD(Yiyi) - need to clean this up later
+    # checkpoint. #TODO(Yiyi) - need to clean this up later
     if image is None:
         raise ValueError("`image` input cannot be undefined.")
 
@@ -277,7 +277,7 @@ def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool
         # images are in latent space and thus can't
         # be masked set masked_image to None
         # we assume that the checkpoint is not an inpainting
-        # checkpoint. TOD(Yiyi) - need to clean this up later
+        # checkpoint. #TODO(Yiyi) - need to clean this up later
         masked_image = None
     else:
         masked_image = image * (mask < 0.5)
@@ -1073,7 +1073,7 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
                 # because `num_inference_steps` might be even given that every timestep
                 # (except the highest one) is duplicated. If `num_inference_steps` is even it would
                 # mean that we cut the timesteps in the middle of the denoising step
-                # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+                # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1
                 # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
                 num_inference_steps = num_inference_steps + 1
 

examples/community/pipeline_stable_diffusion_xl_instantid.py

@@ -46,6 +46,11 @@ except Exception:
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+logger.warning(
+    "To use instant id pipelines, please make sure you have the `insightface` library installed: `pip install insightface`."
+    "Please refer to: https://huggingface.co/InstantX/InstantID for further instructions regarding inference"
+)
+
 
 def FeedForward(dim, mult=4):
     inner_dim = int(dim * mult)
@@ -701,7 +706,7 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
             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 pipeine class.
+                `._callback_tensor_inputs` attribute of your pipeline class.
 
         Examples:
 

examples/community/scheduling_ufogen.py

@@ -81,7 +81,7 @@ def betas_for_alpha_bar(
             return math.exp(t * -12.0)
 
     else:
-        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
 
     betas = []
     for i in range(num_diffusion_timesteps):

examples/community/seed_resize_stable_diffusion.py

@@ -1,6 +1,7 @@
 """
-    modified based on diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+modified based on diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
 """
+
 import inspect
 from typing import Callable, List, Optional, Union
 

examples/community/stable_diffusion_ipex.py

@@ -224,7 +224,7 @@ class StableDiffusionIPEXPipeline(
         # 5. Prepare latent variables
         latents = self.prepare_latents(
             batch_size * num_images_per_prompt,
-            self.unet.in_channels,
+            self.unet.config.in_channels,
             height,
             width,
             prompt_embeds.dtype,
@@ -679,7 +679,7 @@ class StableDiffusionIPEXPipeline(
         timesteps = self.scheduler.timesteps
 
         # 5. Prepare latent variables
-        num_channels_latents = self.unet.in_channels
+        num_channels_latents = self.unet.config.in_channels
         latents = self.prepare_latents(
             batch_size * num_images_per_prompt,
             num_channels_latents,

examples/community/stable_diffusion_tensorrt_txt2img.py

@@ -917,7 +917,7 @@ class TensorRTStableDiffusionPipeline(StableDiffusionPipeline):
             text_embeddings = self.__encode_prompt(prompt, negative_prompt)
 
             # Pre-initialize latents
-            num_channels_latents = self.unet.in_channels
+            num_channels_latents = self.unet.config.in_channels
             latents = self.prepare_latents(
                 batch_size,
                 num_channels_latents,

examples/community/unclip_text_interpolation.py

@@ -35,7 +35,6 @@ def slerp(val, low, high):
 
 
 class UnCLIPTextInterpolationPipeline(DiffusionPipeline):
-
     """
     Pipeline for prompt-to-prompt interpolation on CLIP text embeddings and using the UnCLIP / Dall-E to decode them to images.
 
@@ -49,7 +48,7 @@ class UnCLIPTextInterpolationPipeline(DiffusionPipeline):
             Tokenizer of class
             [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
         prior ([`PriorTransformer`]):
-            The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+            The canonical unCLIP prior to approximate the image embedding from the text embedding.
         text_proj ([`UnCLIPTextProjModel`]):
             Utility class to prepare and combine the embeddings before they are passed to the decoder.
         decoder ([`UNet2DConditionModel`]):

examples/controlnet/train_controlnet_sdxl.py

@@ -125,7 +125,11 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
         )
 
     image_logs = []
-    inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast("cuda")
+    inference_ctx = (
+        contextlib.nullcontext()
+        if (is_final_validation or torch.backends.mps.is_available())
+        else torch.autocast("cuda")
+    )
 
     for validation_prompt, validation_image in zip(validation_prompts, validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
@@ -792,6 +796,12 @@ def main(args):
 
     logging_dir = Path(args.output_dir, args.logging_dir)
 
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
 
     accelerator = Accelerator(

examples/dreambooth/README_sdxl.md

@@ -259,13 +259,17 @@ The authors found that by using DoRA, both the learning capacity and training st
 > This is also aligned with some of the quantitative analysis shown in the paper. 
 
 **Usage**
-1. To use DoRA you need to install `peft` from main: 
+1. To use DoRA you need to upgrade the installation of `peft`: 
 ```bash
-pip install git+https://github.com/huggingface/peft.git
+pip install-U peft
 ```
 2. Enable DoRA training by adding this flag
 ```bash
 --use_dora
 ```
 **Inference** 
-The inference is the same as if you train a regular LoRA 🤗
+The inference is the same as if you train a regular LoRA 🤗
+
+## Format compatibility
+
+You can pass `--output_kohya_format` to additionally generate a state dictionary which should be compatible with other platforms and tools such as Automatic 1111, Comfy, Kohya, etc. The `output_dir` will contain a file named "pytorch_lora_weights_kohya.safetensors".

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
 import gc
 import itertools
 import json
@@ -41,6 +40,7 @@ from peft import LoraConfig, set_peft_model_state_dict
 from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
+from safetensors.torch import load_file, save_file
 from torch.utils.data import Dataset
 from torchvision import transforms
 from torchvision.transforms.functional import crop
@@ -62,7 +62,9 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
 from diffusers.utils import (
     check_min_version,
+    convert_all_state_dict_to_peft,
     convert_state_dict_to_diffusers,
+    convert_state_dict_to_kohya,
     convert_unet_state_dict_to_peft,
     is_wandb_available,
 )
@@ -205,11 +207,18 @@ def log_validation(
     generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
     # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
     # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
-    inference_ctx = (
-        contextlib.nullcontext() if "playground" in args.pretrained_model_name_or_path else torch.cuda.amp.autocast()
-    )
+    enable_autocast = True
+    if torch.backends.mps.is_available() or (
+        accelerator.mixed_precision == "fp16" or accelerator.mixed_precision == "bf16"
+    ):
+        enable_autocast = False
+    if "playground" in args.pretrained_model_name_or_path:
+        enable_autocast = False
 
-    with inference_ctx:
+    with torch.autocast(
+        accelerator.device.type,
+        enabled=enable_autocast,
+    ):
         images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
 
     for tracker in accelerator.trackers:
@@ -227,7 +236,8 @@ def log_validation(
             )
 
     del pipeline
-    torch.cuda.empty_cache()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
 
     return images
 
@@ -396,6 +406,11 @@ def parse_args(input_args=None):
         default="lora-dreambooth-model",
         help="The output directory where the model predictions and checkpoints will be written.",
     )
+    parser.add_argument(
+        "--output_kohya_format",
+        action="store_true",
+        help="Flag to additionally generate final state dict in the Kohya format so that it becomes compatible with A111, Comfy, Kohya, etc.",
+    )
     parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
     parser.add_argument(
         "--resolution",
@@ -959,6 +974,12 @@ def main(args):
     if args.do_edm_style_training and args.snr_gamma is not None:
         raise ValueError("Min-SNR formulation is not supported when conducting EDM-style training.")
 
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     logging_dir = Path(args.output_dir, args.logging_dir)
 
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
@@ -1001,7 +1022,8 @@ def main(args):
         cur_class_images = len(list(class_images_dir.iterdir()))
 
         if cur_class_images < args.num_class_images:
-            torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
+            has_supported_fp16_accelerator = torch.cuda.is_available() or torch.backends.mps.is_available()
+            torch_dtype = torch.float16 if has_supported_fp16_accelerator else torch.float32
             if args.prior_generation_precision == "fp32":
                 torch_dtype = torch.float32
             elif args.prior_generation_precision == "fp16":
@@ -1126,6 +1148,12 @@ def main(args):
     elif accelerator.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
+    if torch.backends.mps.is_available() and weight_dtype == torch.bfloat16:
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     # Move unet, vae and text_encoder to device and cast to weight_dtype
     unet.to(accelerator.device, dtype=weight_dtype)
 
@@ -1270,7 +1298,7 @@ def main(args):
 
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
-    if args.allow_tf32:
+    if args.allow_tf32 and torch.cuda.is_available():
         torch.backends.cuda.matmul.allow_tf32 = True
 
     if args.scale_lr:
@@ -1447,7 +1475,8 @@ def main(args):
     if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
         del tokenizers, text_encoders
         gc.collect()
-        torch.cuda.empty_cache()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
 
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
@@ -1890,6 +1919,11 @@ def main(args):
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )
+        if args.output_kohya_format:
+            lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
+            peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
+            kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
+            save_file(kohya_state_dict, f"{args.output_dir}/pytorch_lora_weights_kohya.safetensors")
 
         # Final inference
         # Load previous pipeline

examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py

@@ -71,12 +71,7 @@ TORCH_DTYPE_MAPPING = {"fp32": torch.float32, "fp16": torch.float16, "bf16": tor
 
 
 def log_validation(
-    pipeline,
-    args,
-    accelerator,
-    generator,
-    global_step,
-    is_final_validation=False,
+    pipeline, args, accelerator, generator, global_step, is_final_validation=False, enable_autocast=True
 ):
     logger.info(
         f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
@@ -96,7 +91,7 @@ def log_validation(
         else Image.open(image_url_or_path).convert("RGB")
     )(args.val_image_url_or_path)
 
-    with torch.autocast(str(accelerator.device).replace(":0", ""), enabled=accelerator.mixed_precision == "fp16"):
+    with torch.autocast(accelerator.device.type, enabled=enable_autocast):
         edited_images = []
         # Run inference
         for val_img_idx in range(args.num_validation_images):
@@ -497,6 +492,13 @@ def main():
             ),
         )
     logging_dir = os.path.join(args.output_dir, args.logging_dir)
+
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
@@ -981,6 +983,13 @@ def main():
     if accelerator.is_main_process:
         accelerator.init_trackers("instruct-pix2pix-xl", config=vars(args))
 
+    # Some configurations require autocast to be disabled.
+    enable_autocast = True
+    if torch.backends.mps.is_available() or (
+        accelerator.mixed_precision == "fp16" or accelerator.mixed_precision == "bf16"
+    ):
+        enable_autocast = False
+
     # Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
@@ -1193,6 +1202,7 @@ def main():
                         generator,
                         global_step,
                         is_final_validation=False,
+                        enable_autocast=enable_autocast,
                     )
 
                     if args.use_ema:
@@ -1242,6 +1252,7 @@ def main():
                 generator,
                 global_step,
                 is_final_validation=True,
+                enable_autocast=enable_autocast,
             )
 
     accelerator.end_training()

examples/research_projects/diffusion_orpo/README.md

@@ -0,0 +1,121 @@
+This project is an attempt to check if it's possible to apply to [ORPO](https://arxiv.org/abs/2403.07691) on a text-conditioned diffusion model to align it on preference data WITHOUT a reference model. The implementation is based on https://github.com/huggingface/trl/pull/1435/. 
+
+> [!WARNING] 
+> We assume that MSE in the diffusion formulation approximates the log-probs as required by ORPO (hat-tip to [@kashif](https://github.com/kashif) for the idea). So, please consider this to be extremely experimental.
+
+## Training
+
+Here's training command you can use on a 40GB A100 to validate things on a [small preference
+dataset](https://hf.co/datasets/kashif/pickascore): 
+
+```bash
+accelerate launch train_diffusion_orpo_sdxl_lora.py \
+  --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0  \
+  --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
+  --output_dir="diffusion-sdxl-orpo" \
+  --mixed_precision="fp16" \
+  --dataset_name=kashif/pickascore \
+  --train_batch_size=8 \
+  --gradient_accumulation_steps=2 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --rank=8 \
+  --learning_rate=1e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=2000 \
+  --checkpointing_steps=500 \
+  --run_validation --validation_steps=50 \
+  --seed="0" \
+  --report_to="wandb" \
+  --push_to_hub
+```
+
+We also provide a simple script to scale up the training on the [yuvalkirstain/pickapic_v2](https://huggingface.co/datasets/yuvalkirstain/pickapic_v2) dataset:
+
+```bash
+accelerate launch --multi_gpu train_diffusion_orpo_sdxl_lora_wds.py \
+  --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0  \
+  --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
+  --dataset_path="pipe:aws s3 cp s3://diffusion-preference-opt/{00000..00644}.tar -" \
+  --output_dir="diffusion-sdxl-orpo-wds" \
+  --mixed_precision="fp16" \
+  --gradient_accumulation_steps=1 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --rank=8 \
+  --dataloader_num_workers=8 \
+  --learning_rate=3e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=50000 \
+  --checkpointing_steps=2000 \
+  --run_validation --validation_steps=500 \
+  --seed="0" \
+  --report_to="wandb" \
+  --push_to_hub
+```
+
+We tested the above on a node of 8 H100s but it should also work on A100s. It requires the `webdataset` library for faster dataloading. Note that we kept the dataset shards on an S3 bucket but it should be also possible to have them stored locally. 
+
+You can use the code below to convert the original dataset into `webdataset` shards:
+
+```python
+import os
+import io
+import ray
+import webdataset as wds
+from datasets import Dataset
+from PIL import Image
+
+ray.init(num_cpus=8)
+
+
+def convert_to_image(im_bytes):
+    return Image.open(io.BytesIO(im_bytes)).convert("RGB")
+
+def main():
+    dataset_path = "/pickapic_v2/data"
+    wds_shards_path = "/pickapic_v2_webdataset"
+    # get all .parquet files in the dataset path
+    dataset_files = [
+        os.path.join(dataset_path, f)
+        for f in os.listdir(dataset_path)
+        if f.endswith(".parquet")
+    ]
+
+    @ray.remote
+    def create_shard(path):
+        # get basename of the file
+        basename = os.path.basename(path)
+        # get the shard number data-00123-of-01034.parquet -> 00123
+        shard_num = basename.split("-")[1]
+        dataset = Dataset.from_parquet(path)
+        # create a webdataset shard
+        shard = wds.TarWriter(os.path.join(wds_shards_path, f"{shard_num}.tar"))
+        
+        for i, example in enumerate(dataset):
+            wds_example = {
+                "__key__": str(i),
+                "original_prompt.txt": example["caption"],
+                "jpg_0.jpg": convert_to_image(example["jpg_0"]),
+                "jpg_1.jpg": convert_to_image(example["jpg_1"]),
+                "label_0.txt": str(example["label_0"]),
+                "label_1.txt": str(example["label_1"])
+            }
+            shard.write(wds_example)
+        shard.close()
+
+    futures = [create_shard.remote(path) for path in dataset_files]
+    ray.get(futures)
+
+
+if __name__ == "__main__":
+    main()
+```
+
+## Inference
+
+Refer to [sayakpaul/diffusion-sdxl-orpo](https://huggingface.co/sayakpaul/diffusion-sdxl-orpo) for an experimental checkpoint.

examples/research_projects/diffusion_orpo/requirements.txt

@@ -0,0 +1,7 @@
+datasets
+accelerate
+transformers
+torchvision
+wandb
+peft
+webdataset

examples/research_projects/multi_token_textual_inversion/multi_token_clip.py

@@ -23,6 +23,7 @@ TODO:
 6. Integrate to training x
 7. Test
 """
+
 import copy
 import random
 

examples/research_projects/promptdiffusion/convert_original_promptdiffusion_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for stable diffusion checkpoints which _only_ contain a controlnet. """
+"""Conversion script for stable diffusion checkpoints which _only_ contain a controlnet."""
 
 import argparse
 import re

examples/research_projects/promptdiffusion/pipeline_prompt_diffusion.py

@@ -1005,7 +1005,7 @@ class PromptDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lo
             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 pipeine class.
+                `._callback_tensor_inputs` attribute of your pipeline class.
 
         Examples:
 

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -501,6 +501,12 @@ def main(args):
 
     logging_dir = Path(args.output_dir, args.logging_dir)
 
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
     kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
     accelerator = Accelerator(
@@ -973,6 +979,13 @@ def main(args):
     if accelerator.is_main_process:
         accelerator.init_trackers("text2image-fine-tune", config=vars(args))
 
+    # Some configurations require autocast to be disabled.
+    enable_autocast = True
+    if torch.backends.mps.is_available() or (
+        accelerator.mixed_precision == "fp16" or accelerator.mixed_precision == "bf16"
+    ):
+        enable_autocast = False
+
     # Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
@@ -1199,7 +1212,10 @@ def main(args):
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
 
-                with torch.cuda.amp.autocast():
+                with torch.autocast(
+                    accelerator.device.type,
+                    enabled=enable_autocast,
+                ):
                     images = [
                         pipeline(**pipeline_args, generator=generator).images[0]
                         for _ in range(args.num_validation_images)

examples/text_to_image/train_text_to_image_sdxl.py

@@ -590,6 +590,12 @@ def main(args):
 
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
 
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
         mixed_precision=args.mixed_precision,
@@ -911,6 +917,7 @@ def main(args):
         )
         precomputed_dataset = precomputed_dataset.with_transform(preprocess_train)
 
+    del compute_vae_encodings_fn, compute_embeddings_fn, text_encoder_one, text_encoder_two
     del text_encoders, tokenizers, vae
     gc.collect()
     torch.cuda.empty_cache()
@@ -979,6 +986,13 @@ def main(args):
         model = model._orig_mod if is_compiled_module(model) else model
         return model
 
+    # Some configurations require autocast to be disabled.
+    enable_autocast = True
+    if torch.backends.mps.is_available() or (
+        accelerator.mixed_precision == "fp16" or accelerator.mixed_precision == "bf16"
+    ):
+        enable_autocast = False
+
     # Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
 
@@ -1212,7 +1226,10 @@ def main(args):
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
 
-                with torch.cuda.amp.autocast():
+                with torch.autocast(
+                    accelerator.device.type,
+                    enabled=enable_autocast,
+                ):
                     images = [
                         pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0]
                         for _ in range(args.num_validation_images)
@@ -1267,7 +1284,7 @@ def main(args):
         if args.validation_prompt and args.num_validation_images > 0:
             pipeline = pipeline.to(accelerator.device)
             generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-            with torch.cuda.amp.autocast():
+            with torch.autocast(accelerator.device.type, enabled=enable_autocast):
                 images = [
                     pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
                     for _ in range(args.num_validation_images)

scripts/change_naming_configs_and_checkpoints.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the LDM checkpoints. """
+"""Conversion script for the LDM checkpoints."""
 
 import argparse
 import json

scripts/convert_i2vgen_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the LDM checkpoints. """
+"""Conversion script for the LDM checkpoints."""
 
 import argparse
 

scripts/convert_if.py

@@ -1195,9 +1195,9 @@ def superres_check_against_original(dump_path, unet_checkpoint_path):
         if_II_model = IFStageIII(device="cuda", dir_or_name=orig_path, model_kwargs={"precision": "fp32"}).model
 
     batch_size = 1
-    channels = model.in_channels // 2
-    height = model.sample_size
-    width = model.sample_size
+    channels = model.config.in_channels // 2
+    height = model.config.sample_size
+    width = model.config.sample_size
     height = 1024
     width = 1024
 

scripts/convert_ldm_original_checkpoint_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the LDM checkpoints. """
+"""Conversion script for the LDM checkpoints."""
 
 import argparse
 import json

scripts/convert_lora_safetensor_to_diffusers.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-""" Conversion script for the LoRA's safetensors checkpoints. """
+"""Conversion script for the LoRA's safetensors checkpoints."""
 
 import argparse
 

scripts/convert_ms_text_to_video_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the LDM checkpoints. """
+"""Conversion script for the LDM checkpoints."""
 
 import argparse
 

scripts/convert_ncsnpp_original_checkpoint_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the NCSNPP checkpoints. """
+"""Conversion script for the NCSNPP checkpoints."""
 
 import argparse
 import json

scripts/convert_original_audioldm2_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the AudioLDM2 checkpoints."""
+"""Conversion script for the AudioLDM2 checkpoints."""
 
 import argparse
 import re

scripts/convert_original_audioldm_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the AudioLDM checkpoints."""
+"""Conversion script for the AudioLDM checkpoints."""
 
 import argparse
 import re

scripts/convert_original_controlnet_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for stable diffusion checkpoints which _only_ contain a controlnet. """
+"""Conversion script for stable diffusion checkpoints which _only_ contain a controlnet."""
 
 import argparse
 

scripts/convert_original_musicldm_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the MusicLDM checkpoints."""
+"""Conversion script for the MusicLDM checkpoints."""
 
 import argparse
 import re

scripts/convert_original_stable_diffusion_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the LDM checkpoints. """
+"""Conversion script for the LDM checkpoints."""
 
 import argparse
 import importlib

scripts/convert_versatile_diffusion_to_diffusers.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the Versatile Stable Diffusion checkpoints. """
+"""Conversion script for the Versatile Stable Diffusion checkpoints."""
 
 import argparse
 from argparse import Namespace

scripts/convert_zero123_to_diffusers.py

@@ -11,6 +11,7 @@ $ python convert_zero123_to_diffusers.py \
    --original_config_file /path/zero123/configs/sd-objaverse-finetune-c_concat-256.yaml
 ```
 """
+
 import argparse
 
 import torch

setup.py

@@ -72,7 +72,11 @@ To create the package for PyPI.
 9. Upload the final version to the actual PyPI:
    twine upload dist/* -r pypi
 
-10. Prepare the release notes and publish them on GitHub once everything is looking hunky-dory.
+10. Prepare the release notes and publish them on GitHub once everything is looking hunky-dory. You can use the following
+    Space to fetch all the commits applicable for the release: https://huggingface.co/spaces/lysandre/github-release. Repo should
+    be `huggingface/diffusers`. `tag` should be the previous release tag (v0.26.1, for example), and `branch` should be
+    the latest release branch (v0.27.0-release, for example). It denotes all commits that have happened on branch
+    v0.27.0-release after the tag v0.26.1 was created.
 
 11. Run `make post-release` (or, for a patch release, `make post-patch`). If you were on a branch for the release,
     you need to go back to main before executing this.
@@ -81,9 +85,8 @@ To create the package for PyPI.
 import os
 import re
 import sys
-from distutils.core import Command
 
-from setuptools import find_packages, setup
+from setuptools import Command, find_packages, setup
 
 
 # IMPORTANT:
@@ -163,7 +166,7 @@ def deps_list(*pkgs):
 
 class DepsTableUpdateCommand(Command):
     """
-    A custom distutils command that updates the dependency table.
+    A custom command that updates the dependency table.
     usage: python setup.py deps_table_update
     """
 

src/diffusers/configuration_utils.py

@@ -13,7 +13,8 @@
 # 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.
-""" ConfigMixin base class and utilities."""
+"""ConfigMixin base class and utilities."""
+
 import dataclasses
 import functools
 import importlib

src/diffusers/loaders/ip_adapter.py

@@ -19,7 +19,7 @@ import torch
 from huggingface_hub.utils import validate_hf_hub_args
 from safetensors import safe_open
 
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict
 from ..utils import (
     _get_model_file,
     is_accelerate_available,
@@ -182,7 +182,7 @@ class IPAdapterMixin:
                             elif key.startswith("ip_adapter."):
                                 state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
                 else:
-                    state_dict = torch.load(model_file, map_location="cpu")
+                    state_dict = load_state_dict(model_file)
             else:
                 state_dict = pretrained_model_name_or_path_or_dict
 

src/diffusers/loaders/lora.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import copy
 import inspect
 import os
 from pathlib import Path
@@ -25,7 +26,7 @@ from packaging import version
 from torch import nn
 
 from .. import __version__
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
@@ -36,6 +37,7 @@ from ..utils import (
     get_adapter_name,
     get_peft_kwargs,
     is_accelerate_available,
+    is_peft_version,
     is_transformers_available,
     logging,
     recurse_remove_peft_layers,
@@ -113,7 +115,7 @@ class LoraLoaderMixin:
         # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
         state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
 
-        is_correct_format = all("lora" in key for key in state_dict.keys())
+        is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
 
@@ -281,7 +283,7 @@ class LoraLoaderMixin:
                     subfolder=subfolder,
                     user_agent=user_agent,
                 )
-                state_dict = torch.load(model_file, map_location="cpu")
+                state_dict = load_state_dict(model_file)
         else:
             state_dict = pretrained_model_name_or_path_or_dict
 
@@ -451,6 +453,15 @@ class LoraLoaderMixin:
                     rank[key] = val.shape[1]
 
             lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
+            if "use_dora" in lora_config_kwargs:
+                if lora_config_kwargs["use_dora"]:
+                    if is_peft_version("<", "0.9.0"):
+                        raise ValueError(
+                            "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                        )
+                else:
+                    if is_peft_version("<", "0.9.0"):
+                        lora_config_kwargs.pop("use_dora")
             lora_config = LoraConfig(**lora_config_kwargs)
 
             # adapter_name
@@ -572,6 +583,15 @@ class LoraLoaderMixin:
                     }
 
                 lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
+                if "use_dora" in lora_config_kwargs:
+                    if lora_config_kwargs["use_dora"]:
+                        if is_peft_version("<", "0.9.0"):
+                            raise ValueError(
+                                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                            )
+                    else:
+                        if is_peft_version("<", "0.9.0"):
+                            lora_config_kwargs.pop("use_dora")
                 lora_config = LoraConfig(**lora_config_kwargs)
 
                 # adapter_name
@@ -654,6 +674,13 @@ class LoraLoaderMixin:
                     rank[key] = val.shape[1]
 
             lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict)
+            if "use_dora" in lora_config_kwargs:
+                if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
+                    raise ValueError(
+                        "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+                    )
+                else:
+                    lora_config_kwargs.pop("use_dora")
             lora_config = LoraConfig(**lora_config_kwargs)
 
             # adapter_name
@@ -959,7 +986,7 @@ class LoraLoaderMixin:
         self,
         adapter_names: Union[List[str], str],
         text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
-        text_encoder_weights: List[float] = None,
+        text_encoder_weights: Optional[Union[float, List[float], List[None]]] = None,
     ):
         """
         Sets the adapter layers for the text encoder.
@@ -977,15 +1004,20 @@ class LoraLoaderMixin:
             raise ValueError("PEFT backend is required for this method.")
 
         def process_weights(adapter_names, weights):
-            if weights is None:
-                weights = [1.0] * len(adapter_names)
-            elif isinstance(weights, float):
-                weights = [weights]
+            # Expand weights into a list, one entry per adapter
+            # e.g. for 2 adapters:  7 -> [7,7] ; [3, None] -> [3, None]
+            if not isinstance(weights, list):
+                weights = [weights] * len(adapter_names)
 
             if len(adapter_names) != len(weights):
                 raise ValueError(
                     f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
                 )
+
+            # Set None values to default of 1.0
+            # e.g. [7,7] -> [7,7] ; [3, None] -> [3,1]
+            weights = [w if w is not None else 1.0 for w in weights]
+
             return weights
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
@@ -1033,17 +1065,77 @@ class LoraLoaderMixin:
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        adapter_weights: Optional[List[float]] = None,
+        adapter_weights: Optional[Union[float, Dict, List[float], List[Dict]]] = None,
     ):
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        adapter_weights = copy.deepcopy(adapter_weights)
+
+        # Expand weights into a list, one entry per adapter
+        if not isinstance(adapter_weights, list):
+            adapter_weights = [adapter_weights] * len(adapter_names)
+
+        if len(adapter_names) != len(adapter_weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(adapter_weights)}"
+            )
+
+        # Decompose weights into weights for unet, text_encoder and text_encoder_2
+        unet_lora_weights, text_encoder_lora_weights, text_encoder_2_lora_weights = [], [], []
+
+        list_adapters = self.get_list_adapters()  # eg {"unet": ["adapter1", "adapter2"], "text_encoder": ["adapter2"]}
+        all_adapters = {
+            adapter for adapters in list_adapters.values() for adapter in adapters
+        }  # eg ["adapter1", "adapter2"]
+        invert_list_adapters = {
+            adapter: [part for part, adapters in list_adapters.items() if adapter in adapters]
+            for adapter in all_adapters
+        }  # eg {"adapter1": ["unet"], "adapter2": ["unet", "text_encoder"]}
+
+        for adapter_name, weights in zip(adapter_names, adapter_weights):
+            if isinstance(weights, dict):
+                unet_lora_weight = weights.pop("unet", None)
+                text_encoder_lora_weight = weights.pop("text_encoder", None)
+                text_encoder_2_lora_weight = weights.pop("text_encoder_2", None)
+
+                if len(weights) > 0:
+                    raise ValueError(
+                        f"Got invalid key '{weights.keys()}' in lora weight dict for adapter {adapter_name}."
+                    )
+
+                if text_encoder_2_lora_weight is not None and not hasattr(self, "text_encoder_2"):
+                    logger.warning(
+                        "Lora weight dict contains text_encoder_2 weights but will be ignored because pipeline does not have text_encoder_2."
+                    )
+
+                # warn if adapter doesn't have parts specified by adapter_weights
+                for part_weight, part_name in zip(
+                    [unet_lora_weight, text_encoder_lora_weight, text_encoder_2_lora_weight],
+                    ["unet", "text_encoder", "text_encoder_2"],
+                ):
+                    if part_weight is not None and part_name not in invert_list_adapters[adapter_name]:
+                        logger.warning(
+                            f"Lora weight dict for adapter '{adapter_name}' contains {part_name}, but this will be ignored because {adapter_name} does not contain weights for {part_name}. Valid parts for {adapter_name} are: {invert_list_adapters[adapter_name]}."
+                        )
+
+            else:
+                unet_lora_weight = weights
+                text_encoder_lora_weight = weights
+                text_encoder_2_lora_weight = weights
+
+            unet_lora_weights.append(unet_lora_weight)
+            text_encoder_lora_weights.append(text_encoder_lora_weight)
+            text_encoder_2_lora_weights.append(text_encoder_2_lora_weight)
+
         unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
         # Handle the UNET
-        unet.set_adapters(adapter_names, adapter_weights)
+        unet.set_adapters(adapter_names, unet_lora_weights)
 
         # Handle the Text Encoder
         if hasattr(self, "text_encoder"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, text_encoder_lora_weights)
         if hasattr(self, "text_encoder_2"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, text_encoder_2_lora_weights)
 
     def disable_lora(self):
         if not USE_PEFT_BACKEND:
@@ -1243,7 +1335,7 @@ class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
             unet_config=self.unet.config,
             **kwargs,
         )
-        is_correct_format = all("lora" in key for key in state_dict.keys())
+        is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
 

src/diffusers/loaders/lora_conversion_utils.py

@@ -14,7 +14,7 @@
 
 import re
 
-from ..utils import logging
+from ..utils import is_peft_version, logging
 
 
 logger = logging.get_logger(__name__)
@@ -128,6 +128,15 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
     te_state_dict = {}
     te2_state_dict = {}
     network_alphas = {}
+    is_unet_dora_lora = any("dora_scale" in k and "lora_unet_" in k for k in state_dict)
+    is_te_dora_lora = any("dora_scale" in k and ("lora_te_" in k or "lora_te1_" in k) for k in state_dict)
+    is_te2_dora_lora = any("dora_scale" in k and "lora_te2_" in k for k in state_dict)
+
+    if is_unet_dora_lora or is_te_dora_lora or is_te2_dora_lora:
+        if is_peft_version("<", "0.9.0"):
+            raise ValueError(
+                "You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
+            )
 
     # every down weight has a corresponding up weight and potentially an alpha weight
     lora_keys = [k for k in state_dict.keys() if k.endswith("lora_down.weight")]
@@ -198,46 +207,19 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
                 unet_state_dict[diffusers_name] = state_dict.pop(key)
                 unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
 
-        elif lora_name.startswith("lora_te_"):
-            diffusers_name = key.replace("lora_te_", "").replace("_", ".")
-            diffusers_name = diffusers_name.replace("text.model", "text_model")
-            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
-            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
-            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
-            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
-            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
-            if "self_attn" in diffusers_name:
-                te_state_dict[diffusers_name] = state_dict.pop(key)
-                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
-            elif "mlp" in diffusers_name:
-                # Be aware that this is the new diffusers convention and the rest of the code might
-                # not utilize it yet.
-                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
-                te_state_dict[diffusers_name] = state_dict.pop(key)
-                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            if is_unet_dora_lora:
+                dora_scale_key_to_replace = "_lora.down." if "_lora.down." in diffusers_name else ".lora.down."
+                unet_state_dict[
+                    diffusers_name.replace(dora_scale_key_to_replace, ".lora_magnitude_vector.")
+                ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
 
-        # (sayakpaul): Duplicate code. Needs to be cleaned.
-        elif lora_name.startswith("lora_te1_"):
-            diffusers_name = key.replace("lora_te1_", "").replace("_", ".")
-            diffusers_name = diffusers_name.replace("text.model", "text_model")
-            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
-            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
-            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
-            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
-            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
-            if "self_attn" in diffusers_name:
-                te_state_dict[diffusers_name] = state_dict.pop(key)
-                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
-            elif "mlp" in diffusers_name:
-                # Be aware that this is the new diffusers convention and the rest of the code might
-                # not utilize it yet.
-                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
-                te_state_dict[diffusers_name] = state_dict.pop(key)
-                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+        elif lora_name.startswith(("lora_te_", "lora_te1_", "lora_te2_")):
+            if lora_name.startswith(("lora_te_", "lora_te1_")):
+                key_to_replace = "lora_te_" if lora_name.startswith("lora_te_") else "lora_te1_"
+            else:
+                key_to_replace = "lora_te2_"
 
-        # (sayakpaul): Duplicate code. Needs to be cleaned.
-        elif lora_name.startswith("lora_te2_"):
-            diffusers_name = key.replace("lora_te2_", "").replace("_", ".")
+            diffusers_name = key.replace(key_to_replace, "").replace("_", ".")
             diffusers_name = diffusers_name.replace("text.model", "text_model")
             diffusers_name = diffusers_name.replace("self.attn", "self_attn")
             diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
@@ -245,14 +227,35 @@ def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_
             diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
             diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
             if "self_attn" in diffusers_name:
-                te2_state_dict[diffusers_name] = state_dict.pop(key)
-                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                if lora_name.startswith(("lora_te_", "lora_te1_")):
+                    te_state_dict[diffusers_name] = state_dict.pop(key)
+                    te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                else:
+                    te2_state_dict[diffusers_name] = state_dict.pop(key)
+                    te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
             elif "mlp" in diffusers_name:
                 # Be aware that this is the new diffusers convention and the rest of the code might
                 # not utilize it yet.
                 diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
-                te2_state_dict[diffusers_name] = state_dict.pop(key)
-                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                if lora_name.startswith(("lora_te_", "lora_te1_")):
+                    te_state_dict[diffusers_name] = state_dict.pop(key)
+                    te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                else:
+                    te2_state_dict[diffusers_name] = state_dict.pop(key)
+                    te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+            if (is_te_dora_lora or is_te2_dora_lora) and lora_name.startswith(("lora_te_", "lora_te1_", "lora_te2_")):
+                dora_scale_key_to_replace_te = (
+                    "_lora.down." if "_lora.down." in diffusers_name else ".lora_linear_layer."
+                )
+                if lora_name.startswith(("lora_te_", "lora_te1_")):
+                    te_state_dict[
+                        diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")
+                    ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
+                elif lora_name.startswith("lora_te2_"):
+                    te2_state_dict[
+                        diffusers_name.replace(dora_scale_key_to_replace_te, ".lora_magnitude_vector.")
+                    ] = state_dict.pop(key.replace("lora_down.weight", "dora_scale"))
 
         # Rename the alphas so that they can be mapped appropriately.
         if lora_name_alpha in state_dict:

src/diffusers/loaders/single_file_utils.py

@@ -12,7 +12,7 @@
 # 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.
-""" Conversion script for the Stable Diffusion checkpoints."""
+"""Conversion script for the Stable Diffusion checkpoints."""
 
 import os
 import re
@@ -50,6 +50,8 @@ if is_transformers_available():
 if is_accelerate_available():
     from accelerate import init_empty_weights
 
+    from ..models.modeling_utils import load_model_dict_into_meta
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 CONFIG_URLS = {
@@ -977,8 +979,6 @@ def create_diffusers_controlnet_model_from_ldm(
         controlnet = ControlNetModel(**diffusers_config)
 
     if is_accelerate_available():
-        from ..models.modeling_utils import load_model_dict_into_meta
-
         unexpected_keys = load_model_dict_into_meta(
             controlnet, diffusers_format_controlnet_checkpoint, dtype=torch_dtype
         )
@@ -1155,8 +1155,6 @@ def create_text_encoder_from_ldm_clip_checkpoint(config_name, checkpoint, local_
                 text_model_dict[diffusers_key] = checkpoint[key]
 
     if is_accelerate_available():
-        from ..models.modeling_utils import load_model_dict_into_meta
-
         unexpected_keys = load_model_dict_into_meta(text_model, text_model_dict, dtype=torch_dtype)
         if text_model._keys_to_ignore_on_load_unexpected is not None:
             for pat in text_model._keys_to_ignore_on_load_unexpected:
@@ -1250,8 +1248,6 @@ def create_text_encoder_from_open_clip_checkpoint(
             text_model_dict[diffusers_key] = checkpoint[key]
 
     if is_accelerate_available():
-        from ..models.modeling_utils import load_model_dict_into_meta
-
         unexpected_keys = load_model_dict_into_meta(text_model, text_model_dict, dtype=torch_dtype)
         if text_model._keys_to_ignore_on_load_unexpected is not None:
             for pat in text_model._keys_to_ignore_on_load_unexpected:
@@ -1317,8 +1313,6 @@ def create_diffusers_unet_model_from_ldm(
         unet = UNet2DConditionModel(**unet_config)
 
     if is_accelerate_available():
-        from ..models.modeling_utils import load_model_dict_into_meta
-
         unexpected_keys = load_model_dict_into_meta(unet, diffusers_format_unet_checkpoint, dtype=torch_dtype)
         if unet._keys_to_ignore_on_load_unexpected is not None:
             for pat in unet._keys_to_ignore_on_load_unexpected:
@@ -1379,8 +1373,6 @@ def create_diffusers_vae_model_from_ldm(
         vae = AutoencoderKL(**vae_config)
 
     if is_accelerate_available():
-        from ..models.modeling_utils import load_model_dict_into_meta
-
         unexpected_keys = load_model_dict_into_meta(vae, diffusers_format_vae_checkpoint, dtype=torch_dtype)
         if vae._keys_to_ignore_on_load_unexpected is not None:
             for pat in vae._keys_to_ignore_on_load_unexpected:

src/diffusers/loaders/textual_inversion.py

@@ -18,6 +18,7 @@ import torch
 from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
+from ..models.modeling_utils import load_state_dict
 from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
 
 
@@ -100,7 +101,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     subfolder=subfolder,
                     user_agent=user_agent,
                 )
-                state_dict = torch.load(model_file, map_location="cpu")
+                state_dict = load_state_dict(model_file)
         else:
             state_dict = pretrained_model_name_or_path
 

src/diffusers/loaders/unet.py

@@ -31,7 +31,7 @@ from ..models.embeddings import (
     IPAdapterPlusImageProjection,
     MultiIPAdapterImageProjection,
 )
-from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta, load_state_dict
 from ..utils import (
     USE_PEFT_BACKEND,
     _get_model_file,
@@ -47,6 +47,7 @@ from .single_file_utils import (
     infer_stable_cascade_single_file_config,
     load_single_file_model_checkpoint,
 )
+from .unet_loader_utils import _maybe_expand_lora_scales
 from .utils import AttnProcsLayers
 
 
@@ -214,7 +215,7 @@ class UNet2DConditionLoadersMixin:
                     subfolder=subfolder,
                     user_agent=user_agent,
                 )
-                state_dict = torch.load(model_file, map_location="cpu")
+                state_dict = load_state_dict(model_file)
         else:
             state_dict = pretrained_model_name_or_path_or_dict
 
@@ -564,7 +565,7 @@ class UNet2DConditionLoadersMixin:
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        weights: Optional[Union[List[float], float]] = None,
+        weights: Optional[Union[float, Dict, List[float], List[Dict], List[None]]] = None,
     ):
         """
         Set the currently active adapters for use in the UNet.
@@ -597,9 +598,9 @@ class UNet2DConditionLoadersMixin:
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
 
-        if weights is None:
-            weights = [1.0] * len(adapter_names)
-        elif isinstance(weights, float):
+        # Expand weights into a list, one entry per adapter
+        # examples for e.g. 2 adapters:  [{...}, 7] -> [7,7] ; None -> [None, None]
+        if not isinstance(weights, list):
             weights = [weights] * len(adapter_names)
 
         if len(adapter_names) != len(weights):
@@ -607,6 +608,13 @@ class UNet2DConditionLoadersMixin:
                 f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
             )
 
+        # Set None values to default of 1.0
+        # e.g. [{...}, 7] -> [{...}, 7] ; [None, None] -> [1.0, 1.0]
+        weights = [w if w is not None else 1.0 for w in weights]
+
+        # e.g. [{...}, 7] -> [{expanded dict...}, 7]
+        weights = _maybe_expand_lora_scales(self, weights)
+
         set_weights_and_activate_adapters(self, adapter_names, weights)
 
     def disable_lora(self):

src/diffusers/loaders/unet_loader_utils.py

@@ -0,0 +1,154 @@
+# 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.
+import copy
+from typing import TYPE_CHECKING, Dict, List, Union
+
+from ..utils import logging
+
+
+if TYPE_CHECKING:
+    # import here to avoid circular imports
+    from ..models import UNet2DConditionModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def _translate_into_actual_layer_name(name):
+    """Translate user-friendly name (e.g. 'mid') into actual layer name (e.g. 'mid_block.attentions.0')"""
+    if name == "mid":
+        return "mid_block.attentions.0"
+
+    updown, block, attn = name.split(".")
+
+    updown = updown.replace("down", "down_blocks").replace("up", "up_blocks")
+    block = block.replace("block_", "")
+    attn = "attentions." + attn
+
+    return ".".join((updown, block, attn))
+
+
+def _maybe_expand_lora_scales(unet: "UNet2DConditionModel", weight_scales: List[Union[float, Dict]]):
+    blocks_with_transformer = {
+        "down": [i for i, block in enumerate(unet.down_blocks) if hasattr(block, "attentions")],
+        "up": [i for i, block in enumerate(unet.up_blocks) if hasattr(block, "attentions")],
+    }
+    transformer_per_block = {"down": unet.config.layers_per_block, "up": unet.config.layers_per_block + 1}
+
+    expanded_weight_scales = [
+        _maybe_expand_lora_scales_for_one_adapter(
+            weight_for_adapter, blocks_with_transformer, transformer_per_block, unet.state_dict()
+        )
+        for weight_for_adapter in weight_scales
+    ]
+
+    return expanded_weight_scales
+
+
+def _maybe_expand_lora_scales_for_one_adapter(
+    scales: Union[float, Dict],
+    blocks_with_transformer: Dict[str, int],
+    transformer_per_block: Dict[str, int],
+    state_dict: None,
+):
+    """
+    Expands the inputs into a more granular dictionary. See the example below for more details.
+
+    Parameters:
+        scales (`Union[float, Dict]`):
+            Scales dict to expand.
+        blocks_with_transformer (`Dict[str, int]`):
+            Dict with keys 'up' and 'down', showing which blocks have transformer layers
+        transformer_per_block (`Dict[str, int]`):
+            Dict with keys 'up' and 'down', showing how many transformer layers each block has
+
+    E.g. turns
+    ```python
+    scales = {
+        'down': 2,
+        'mid': 3,
+        'up': {
+            'block_0': 4,
+            'block_1': [5, 6, 7]
+        }
+    }
+    blocks_with_transformer = {
+        'down': [1,2],
+        'up': [0,1]
+    }
+    transformer_per_block = {
+        'down': 2,
+        'up': 3
+    }
+    ```
+    into
+    ```python
+    {
+        'down.block_1.0': 2,
+        'down.block_1.1': 2,
+        'down.block_2.0': 2,
+        'down.block_2.1': 2,
+        'mid': 3,
+        'up.block_0.0': 4,
+        'up.block_0.1': 4,
+        'up.block_0.2': 4,
+        'up.block_1.0': 5,
+        'up.block_1.1': 6,
+        'up.block_1.2': 7,
+    }
+    ```
+    """
+    if sorted(blocks_with_transformer.keys()) != ["down", "up"]:
+        raise ValueError("blocks_with_transformer needs to be a dict with keys `'down' and `'up'`")
+
+    if sorted(transformer_per_block.keys()) != ["down", "up"]:
+        raise ValueError("transformer_per_block needs to be a dict with keys `'down' and `'up'`")
+
+    if not isinstance(scales, dict):
+        # don't expand if scales is a single number
+        return scales
+
+    scales = copy.deepcopy(scales)
+
+    if "mid" not in scales:
+        scales["mid"] = 1
+
+    for updown in ["up", "down"]:
+        if updown not in scales:
+            scales[updown] = 1
+
+        # eg {"down": 1} to {"down": {"block_1": 1, "block_2": 1}}}
+        if not isinstance(scales[updown], dict):
+            scales[updown] = {f"block_{i}": scales[updown] for i in blocks_with_transformer[updown]}
+
+        # eg {"down": "block_1": 1}} to {"down": "block_1": [1, 1]}}
+        for i in blocks_with_transformer[updown]:
+            block = f"block_{i}"
+            if not isinstance(scales[updown][block], list):
+                scales[updown][block] = [scales[updown][block] for _ in range(transformer_per_block[updown])]
+
+        # eg {"down": "block_1": [1, 1]}}  to {"down.block_1.0": 1, "down.block_1.1": 1}
+        for i in blocks_with_transformer[updown]:
+            block = f"block_{i}"
+            for tf_idx, value in enumerate(scales[updown][block]):
+                scales[f"{updown}.{block}.{tf_idx}"] = value
+
+        del scales[updown]
+
+    for layer in scales.keys():
+        if not any(_translate_into_actual_layer_name(layer) in module for module in state_dict.keys()):
+            raise ValueError(
+                f"Can't set lora scale for layer {layer}. It either doesn't exist in this unet or it has no attentions."
+            )
+
+    return {_translate_into_actual_layer_name(name): weight for name, weight in scales.items()}

src/diffusers/models/attention_processor.py

@@ -424,7 +424,7 @@ class Attention(nn.Module):
         # If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
         is_lora_activated.pop("add_k_proj", None)
         is_lora_activated.pop("add_v_proj", None)
-        # 2. else it is not posssible that only some layers have LoRA activated
+        # 2. else it is not possible that only some layers have LoRA activated
         if not all(is_lora_activated.values()):
             raise ValueError(
                 f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
@@ -767,18 +767,7 @@ class AttnProcessor:
         query = attn.to_q(hidden_states)
 
         if encoder_hidden_states is None:
-            # encoder_hidden_states = hidden_states
-            batch, seq, dim = hidden_states.shape
-            height = width = seq**0.5
-            # reshape to (batch, height, width, dim)
-            encoder_hidden_states = hidden_states.view(batch, height, width, dim)
-            # reshape to (batch, dim, height, width)
-            encoder_hidden_states = encoder_hidden_states.permute(0, 3, 1, 2)
-            encoder_hidden_states = torch.nn.functional.avg_pool2d(hidden_states, kernel_size=4)
-            # reshape to (batch, dim, seq)
-            encoder_hidden_states = encoder_hidden_states.view(batch, dim, -1)
-            # reshape to (batch, seq, dim)
-            encoder_hidden_states = encoder_hidden_states.permute(0, 2, 1)
+            encoder_hidden_states = hidden_states
         elif attn.norm_cross:
             encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
 
@@ -1270,18 +1259,7 @@ class AttnProcessor2_0:
         query = attn.to_q(hidden_states)
 
         if encoder_hidden_states is None:
-            # encoder_hidden_states = hidden_states
-            batch, seq, dim = hidden_states.shape
-            height = width = seq**0.5
-            # reshape to (batch, height, width, dim)
-            encoder_hidden_states = hidden_states.view(batch, height, width, dim)
-            # reshape to (batch, dim, height, width)
-            encoder_hidden_states = encoder_hidden_states.permute(0, 3, 1, 2)
-            encoder_hidden_states = torch.nn.functional.avg_pool2d(hidden_states, kernel_size=4)
-            # reshape to (batch, dim, seq)
-            encoder_hidden_states = encoder_hidden_states.view(batch, dim, -1)
-            # reshape to (batch, seq, dim)
-            encoder_hidden_states = encoder_hidden_states.permute(0, 2, 1)
+            encoder_hidden_states = hidden_states
         elif attn.norm_cross:
             encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
 
@@ -2120,7 +2098,7 @@ class LoRAAttnAddedKVProcessor(nn.Module):
 
 class IPAdapterAttnProcessor(nn.Module):
     r"""
-    Attention processor for Multiple IP-Adapater.
+    Attention processor for Multiple IP-Adapters.
 
     Args:
         hidden_size (`int`):
@@ -2174,8 +2152,8 @@ class IPAdapterAttnProcessor(nn.Module):
                 encoder_hidden_states, ip_hidden_states = encoder_hidden_states
             else:
                 deprecation_message = (
-                    "You have passed a tensor as `encoder_hidden_states`.This is deprecated and will be removed in a future release."
-                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to supress this warning."
+                    "You have passed a tensor as `encoder_hidden_states`. This is deprecated and will be removed in a future release."
+                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to suppress this warning."
                 )
                 deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
                 end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]
@@ -2275,7 +2253,7 @@ class IPAdapterAttnProcessor(nn.Module):
 
 class IPAdapterAttnProcessor2_0(torch.nn.Module):
     r"""
-    Attention processor for IP-Adapater for PyTorch 2.0.
+    Attention processor for IP-Adapter for PyTorch 2.0.
 
     Args:
         hidden_size (`int`):
@@ -2334,8 +2312,8 @@ class IPAdapterAttnProcessor2_0(torch.nn.Module):
                 encoder_hidden_states, ip_hidden_states = encoder_hidden_states
             else:
                 deprecation_message = (
-                    "You have passed a tensor as `encoder_hidden_states`.This is deprecated and will be removed in a future release."
-                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to supress this warning."
+                    "You have passed a tensor as `encoder_hidden_states`. This is deprecated and will be removed in a future release."
+                    " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to suppress this warning."
                 )
                 deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
                 end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]

src/diffusers/models/autoencoders/consistency_decoder_vae.py

@@ -63,7 +63,8 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         ...     "runwayml/stable-diffusion-v1-5", vae=vae, torch_dtype=torch.float16
         ... ).to("cuda")
 
-        >>> pipe("horse", generator=torch.manual_seed(0)).images
+        >>> image = pipe("horse", generator=torch.manual_seed(0)).images[0]
+        >>> image
         ```
     """
 
@@ -72,6 +73,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         self,
         scaling_factor: float = 0.18215,
         latent_channels: int = 4,
+        sample_size: int = 32,
         encoder_act_fn: str = "silu",
         encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
         encoder_double_z: bool = True,
@@ -153,6 +155,16 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         self.use_slicing = False
         self.use_tiling = False
 
+        # only relevant if vae tiling is enabled
+        self.tile_sample_min_size = self.config.sample_size
+        sample_size = (
+            self.config.sample_size[0]
+            if isinstance(self.config.sample_size, (list, tuple))
+            else self.config.sample_size
+        )
+        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor = 0.25
+
     # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_tiling
     def enable_tiling(self, use_tiling: bool = True):
         r"""
@@ -272,7 +284,7 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         Args:
             x (`torch.FloatTensor`): Input batch of images.
             return_dict (`bool`, *optional*, defaults to `True`):
-                Whether to return a [`~models.consistecy_decoder_vae.ConsistencyDecoderOoutput`] instead of a plain
+                Whether to return a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] instead of a plain
                 tuple.
 
         Returns:
@@ -305,6 +317,19 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
         return_dict: bool = True,
         num_inference_steps: int = 2,
     ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        """
+        Decodes the input latent vector `z` using the consistency decoder VAE model.
+
+        Args:
+            z (torch.FloatTensor): The input latent vector.
+            generator (Optional[torch.Generator]): The random number generator. Default is None.
+            return_dict (bool): Whether to return the output as a dictionary. Default is True.
+            num_inference_steps (int): The number of inference steps. Default is 2.
+
+        Returns:
+            Union[DecoderOutput, Tuple[torch.FloatTensor]]: The decoded output.
+
+        """
         z = (z * self.config.scaling_factor - self.means) / self.stds
 
         scale_factor = 2 ** (len(self.config.block_out_channels) - 1)
@@ -345,7 +370,9 @@ class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
             b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
         return b
 
-    def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> ConsistencyDecoderVAEOutput:
+    def tiled_encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[ConsistencyDecoderVAEOutput, Tuple]:
         r"""Encode a batch of images using a tiled encoder.
 
         When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several

src/diffusers/models/controlnet.py

@@ -281,7 +281,7 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
         elif encoder_hid_dim_type == "text_image_proj":
             # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
             # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
-            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
             self.encoder_hid_proj = TextImageProjection(
                 text_embed_dim=encoder_hid_dim,
                 image_embed_dim=cross_attention_dim,
@@ -330,7 +330,7 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
         elif addition_embed_type == "text_image":
             # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
             # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
-            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
             self.add_embedding = TextImageTimeEmbedding(
                 text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
             )
@@ -509,6 +509,9 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
             if controlnet.class_embedding:
                 controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
 
+            if hasattr(controlnet, "add_embedding"):
+                controlnet.add_embedding.load_state_dict(unet.add_embedding.state_dict())
+
             controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
             controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
 

src/diffusers/models/modeling_flax_pytorch_utils.py

@@ -12,7 +12,8 @@
 # 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.
-""" PyTorch - Flax general utilities."""
+"""PyTorch - Flax general utilities."""
+
 import re
 
 import jax.numpy as jnp

src/diffusers/models/modeling_pytorch_flax_utils.py

@@ -12,7 +12,7 @@
 # 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.
-""" PyTorch - Flax general utilities."""
+"""PyTorch - Flax general utilities."""
 
 from pickle import UnpicklingError
 

src/diffusers/models/modeling_utils.py

@@ -20,6 +20,7 @@ import os
 import re
 from collections import OrderedDict
 from functools import partial
+from pathlib import Path
 from typing import Any, Callable, List, Optional, Tuple, Union
 
 import safetensors
@@ -107,7 +108,12 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[
         if file_extension == SAFETENSORS_FILE_EXTENSION:
             return safetensors.torch.load_file(checkpoint_file, device="cpu")
         else:
-            return torch.load(checkpoint_file, map_location="cpu")
+            weights_only_kwarg = {"weights_only": True} if is_torch_version(">=", "1.13") else {}
+            return torch.load(
+                checkpoint_file,
+                map_location="cpu",
+                **weights_only_kwarg,
+            )
     except Exception as e:
         try:
             with open(checkpoint_file) as f:
@@ -367,18 +373,18 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         # Save the model
         if safe_serialization:
             safetensors.torch.save_file(
-                state_dict, os.path.join(save_directory, weights_name), metadata={"format": "pt"}
+                state_dict, Path(save_directory, weights_name).as_posix(), metadata={"format": "pt"}
             )
         else:
-            torch.save(state_dict, os.path.join(save_directory, weights_name))
+            torch.save(state_dict, Path(save_directory, weights_name).as_posix())
 
-        logger.info(f"Model weights saved in {os.path.join(save_directory, weights_name)}")
+        logger.info(f"Model weights saved in {Path(save_directory, weights_name).as_posix()}")
 
         if push_to_hub:
             # Create a new empty model card and eventually tag it
             model_card = load_or_create_model_card(repo_id, token=token)
             model_card = populate_model_card(model_card)
-            model_card.save(os.path.join(save_directory, "README.md"))
+            model_card.save(Path(save_directory, "README.md").as_posix())
 
             self._upload_folder(
                 save_directory,

src/diffusers/models/transformer_temporal.py

@@ -20,15 +20,15 @@ from .transformers.transformer_temporal import (
 
 
 class TransformerTemporalModelOutput(TransformerTemporalModelOutput):
-    deprecation_message = "Importing `TransformerTemporalModelOutput` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerTemporalModelOutput`, instead."
+    deprecation_message = "Importing `TransformerTemporalModelOutput` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.transformer_temporal import TransformerTemporalModelOutput`, instead."
     deprecate("TransformerTemporalModelOutput", "0.29", deprecation_message)
 
 
 class TransformerTemporalModel(TransformerTemporalModel):
-    deprecation_message = "Importing `TransformerTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerTemporalModel`, instead."
+    deprecation_message = "Importing `TransformerTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.transformer_temporal import TransformerTemporalModel`, instead."
     deprecate("TransformerTemporalModel", "0.29", deprecation_message)
 
 
 class TransformerSpatioTemporalModel(TransformerSpatioTemporalModel):
-    deprecation_message = "Importing `TransformerSpatioTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerSpatioTemporalModel`, instead."
+    deprecation_message = "Importing `TransformerSpatioTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.transformer_temporal import TransformerSpatioTemporalModel`, instead."
     deprecate("TransformerTemporalModelOutput", "0.29", deprecation_message)

src/diffusers/models/transformers/transformer_2d.py

@@ -129,7 +129,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
         if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
             deprecation_message = (
                 f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
-                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config."
                 " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
                 " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
                 " would be very nice if you could open a Pull request for the `transformer/config.json` file"

src/diffusers/models/unets/unet_2d_condition.py

@@ -580,7 +580,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
         elif encoder_hid_dim_type == "text_image_proj":
             # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
             # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
-            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
             self.encoder_hid_proj = TextImageProjection(
                 text_embed_dim=encoder_hid_dim,
                 image_embed_dim=cross_attention_dim,
@@ -660,7 +660,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
         elif addition_embed_type == "text_image":
             # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
             # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
-            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
             self.add_embedding = TextImageTimeEmbedding(
                 text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
             )
@@ -1010,7 +1010,7 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
         if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
             encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
         elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
-            # Kadinsky 2.1 - style
+            # Kandinsky 2.1 - style
             if "image_embeds" not in added_cond_kwargs:
                 raise ValueError(
                     f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
@@ -1081,6 +1081,8 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
                 A tuple of tensors that if specified are added to the residuals of down unet blocks.
             mid_block_additional_residual: (`torch.Tensor`, *optional*):
                 A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
             encoder_attention_mask (`torch.Tensor`):
                 A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
                 `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
@@ -1088,18 +1090,6 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                 tuple.
-            cross_attention_kwargs (`dict`, *optional*):
-                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
-            added_cond_kwargs: (`dict`, *optional*):
-                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
-                are passed along to the UNet blocks.
-            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
-                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
-                example from ControlNet side model(s)
-            mid_block_additional_residual (`torch.Tensor`, *optional*):
-                additional residual to be added to UNet mid block output, for example from ControlNet side model
-            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
-                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
 
         Returns:
             [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
@@ -1185,7 +1175,14 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin,
             cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
 
         # 3. down
-        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
         if USE_PEFT_BACKEND:
             # weight the lora layers by setting `lora_scale` for each PEFT layer
             scale_lora_layers(self, lora_scale)

src/diffusers/models/unets/unet_3d_condition.py

@@ -91,6 +91,8 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         cross_attention_dim (`int`, *optional*, defaults to 1024): The dimension of the cross attention features.
         attention_head_dim (`int`, *optional*, defaults to 64): The dimension of the attention heads.
         num_attention_heads (`int`, *optional*): The number of attention heads.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
     """
 
     _supports_gradient_checkpointing = False
@@ -123,6 +125,7 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         cross_attention_dim: int = 1024,
         attention_head_dim: Union[int, Tuple[int]] = 64,
         num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        time_cond_proj_dim: Optional[int] = None,
     ):
         super().__init__()
 
@@ -174,6 +177,7 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
             timestep_input_dim,
             time_embed_dim,
             act_fn=act_fn,
+            cond_proj_dim=time_cond_proj_dim,
         )
 
         self.transformer_in = TransformerTemporalModel(

src/diffusers/models/unets/unet_stable_cascade.py

@@ -521,9 +521,11 @@ class StableCascadeUNet(ModelMixin, ConfigMixin, FromOriginalUNetMixin):
                         if isinstance(block, SDCascadeResBlock):
                             skip = level_outputs[i] if k == 0 and i > 0 else None
                             if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                                orig_type = x.dtype
                                 x = torch.nn.functional.interpolate(
                                     x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
                                 )
+                                x = x.to(orig_type)
                             x = torch.utils.checkpoint.checkpoint(
                                 create_custom_forward(block), x, skip, use_reentrant=False
                             )
@@ -547,9 +549,11 @@ class StableCascadeUNet(ModelMixin, ConfigMixin, FromOriginalUNetMixin):
                         if isinstance(block, SDCascadeResBlock):
                             skip = level_outputs[i] if k == 0 and i > 0 else None
                             if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                                orig_type = x.dtype
                                 x = torch.nn.functional.interpolate(
                                     x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
                                 )
+                                x = x.to(orig_type)
                             x = block(x, skip)
                         elif isinstance(block, SDCascadeAttnBlock):
                             x = block(x, clip)