update

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -66,32 +66,32 @@ body:
         Questions on DiffusionPipeline (Saving, Loading, From pretrained, ...):
 
         Questions on pipelines:
-        - Stable Diffusion @yiyixuxu @DN6 @sayakpaul 
-        - Stable Diffusion XL @yiyixuxu @sayakpaul @DN6 
-        - Kandinsky @yiyixuxu 
-        - ControlNet @sayakpaul @yiyixuxu @DN6 
-        - T2I Adapter @sayakpaul @yiyixuxu @DN6 
-        - IF @DN6 
-        - Text-to-Video / Video-to-Video @DN6 @sayakpaul 
-        - Wuerstchen @DN6 
+        - Stable Diffusion @yiyixuxu @DN6 @sayakpaul @patrickvonplaten
+        - Stable Diffusion XL @yiyixuxu @sayakpaul @DN6 @patrickvonplaten
+        - Kandinsky @yiyixuxu @patrickvonplaten
+        - ControlNet @sayakpaul @yiyixuxu @DN6 @patrickvonplaten
+        - T2I Adapter @sayakpaul @yiyixuxu @DN6 @patrickvonplaten
+        - IF @DN6 @patrickvonplaten
+        - Text-to-Video / Video-to-Video @DN6 @sayakpaul @patrickvonplaten
+        - Wuerstchen @DN6 @patrickvonplaten
         - Other: @yiyixuxu @DN6
 
         Questions on models:
-        - UNet @DN6 @yiyixuxu @sayakpaul 
-        - VAE @sayakpaul @DN6 @yiyixuxu 
-        - Transformers/Attention @DN6 @yiyixuxu @sayakpaul @DN6 
+        - UNet @DN6 @yiyixuxu @sayakpaul @patrickvonplaten
+        - VAE @sayakpaul @DN6 @yiyixuxu @patrickvonplaten
+        - Transformers/Attention @DN6 @yiyixuxu @sayakpaul @DN6 @patrickvonplaten
 
-        Questions on Schedulers: @yiyixuxu 
+        Questions on Schedulers: @yiyixuxu @patrickvonplaten
 
-        Questions on LoRA: @sayakpaul 
+        Questions on LoRA: @sayakpaul @patrickvonplaten
 
-        Questions on Textual Inversion: @sayakpaul 
+        Questions on Textual Inversion: @sayakpaul @patrickvonplaten
 
         Questions on Training: 
-        - DreamBooth @sayakpaul 
-        - Text-to-Image Fine-tuning @sayakpaul 
-        - Textual Inversion @sayakpaul 
-        - ControlNet @sayakpaul 
+        - DreamBooth @sayakpaul @patrickvonplaten
+        - Text-to-Image Fine-tuning @sayakpaul @patrickvonplaten
+        - Textual Inversion @sayakpaul @patrickvonplaten
+        - ControlNet @sayakpaul @patrickvonplaten
 
         Questions on Tests: @DN6 @sayakpaul @yiyixuxu 
 
@@ -99,7 +99,7 @@ body:
 
         Questions on JAX- and MPS-related things: @pcuenca
 
-        Questions on audio pipelines: @DN6 
+        Questions on audio pipelines: @DN6 @patrickvonplaten
         
 
         

.github/PULL_REQUEST_TEMPLATE.md

@@ -38,13 +38,13 @@ members/contributors who may be interested in your PR.
 
 Core library:
 
-- Schedulers: @yiyixuxu 
-- Pipelines:  @sayakpaul @yiyixuxu @DN6
-- Training examples: @sayakpaul 
-- Docs: @stevhliu and @sayakpaul
+- Schedulers: @yiyixuxu and @patrickvonplaten
+- Pipelines:  @patrickvonplaten and @sayakpaul
+- Training examples: @sayakpaul and @patrickvonplaten
+- Docs: @stevhliu and @yiyixuxu
 - JAX and MPS: @pcuenca
 - Audio: @sanchit-gandhi
-- General functionalities: @sayakpaul @yiyixuxu @DN6
+- General functionalities: @patrickvonplaten and @sayakpaul
 
 Integrations:
 

.github/workflows/benchmark.yml

@@ -1,7 +1,6 @@
 name: Benchmarking tests
 
 on:
-  workflow_dispatch:
   schedule:
     - cron: "30 1 1,15 * *" # every 2 weeks on the 1st and the 15th of every month at 1:30 AM
 
@@ -32,9 +31,8 @@ jobs:
       - 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 pandas peft
+          python -m pip install -e .[quality,test]
+          python -m pip install pandas peft
       - name: Environment
         run: |
           python utils/print_env.py

.github/workflows/build_docker_images.yml

@@ -11,7 +11,6 @@ concurrency:
 
 env:
   REGISTRY: diffusers
-  CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
 
 jobs:
   build-docker-images:
@@ -51,27 +50,3 @@ jobs:
           context: ./docker/${{ matrix.image-name }}
           push: true
           tags: ${{ env.REGISTRY }}/${{ matrix.image-name }}:latest
-
-      - name: Post to a Slack channel
-        id: slack
-        uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-        with:
-          # Slack channel id, channel name, or user id to post message.
-          # See also: https://api.slack.com/methods/chat.postMessage#channels
-          channel-id: ${{ env.CI_SLACK_CHANNEL }}
-          # For posting a rich message using Block Kit
-          payload: |
-            {
-              "text": "${{ matrix.image-name }} Docker Image build result: ${{ job.status }}\n${{ github.event.head_commit.url }}",
-              "blocks": [
-                {
-                  "type": "section",
-                  "text": {
-                    "type": "mrkdwn",
-                    "text": "${{ matrix.image-name }} Docker Image build result: ${{ job.status }}\n${{ github.event.head_commit.url }}"
-                  }
-                }
-              ]
-            }
-        env:
-          SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/build_documentation.yml

@@ -7,10 +7,6 @@ on:
       - doc-builder*
       - v*-release
       - v*-patch
-    paths:
-      - "src/diffusers/**.py"
-      - "examples/**"
-      - "docs/**"
 
 jobs:
   build:

.github/workflows/build_pr_documentation.yml

@@ -2,10 +2,6 @@ name: Build PR Documentation
 
 on:
   pull_request:
-    paths:
-      - "src/diffusers/**.py"
-      - "examples/**"
-      - "docs/**"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}

.github/workflows/nightly_tests.yml

@@ -60,10 +60,9 @@ jobs:
 
       - name: Install dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m uv pip install -e [quality,test]
-          python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers
-          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
+          python -m pip install -e .[quality,test]
+          python -m pip install -U git+https://github.com/huggingface/transformers
+          python -m pip install git+https://github.com/huggingface/accelerate
 
       - name: Environment
         run: |
@@ -74,7 +73,6 @@ jobs:
         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 "not Flax and not Onnx" \
             --make-reports=tests_${{ matrix.config.report }} \
@@ -85,7 +83,6 @@ jobs:
         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 }} \
@@ -96,7 +93,6 @@ jobs:
         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 }} \
@@ -136,10 +132,10 @@ jobs:
       - name: Install dependencies
         shell: arch -arch arm64 bash {0}
         run: |
-          ${CONDA_RUN} python -m pip install --upgrade pip uv
-          ${CONDA_RUN} python -m uv pip install -e [quality,test]
-          ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
-          ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
+          ${CONDA_RUN} python -m pip install --upgrade pip
+          ${CONDA_RUN} python -m pip install -e .[quality,test]
+          ${CONDA_RUN} python -m pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
+          ${CONDA_RUN} python -m pip install git+https://github.com/huggingface/accelerate
 
       - name: Environment
         shell: arch -arch arm64 bash {0}

.github/workflows/pr_dependency_test.yml

@@ -4,8 +4,6 @@ on:
   pull_request:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
   push:
     branches:
       - main
@@ -25,12 +23,10 @@ jobs:
           python-version: "3.8"
       - name: Install dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m pip install --upgrade pip uv
-          python -m uv pip install -e .
-          python -m uv pip install pytest
+          python -m pip install --upgrade pip
+          pip install -e .
+          pip install pytest
       - name: Check for soft dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           pytest tests/others/test_dependencies.py
       

.github/workflows/pr_flax_dependency_test.yml

@@ -4,8 +4,6 @@ on:
   pull_request:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
   push:
     branches:
       - main
@@ -25,14 +23,12 @@ jobs:
           python-version: "3.8"
       - name: Install dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m pip install --upgrade pip uv
-          python -m uv pip install -e .
-          python -m uv pip install "jax[cpu]>=0.2.16,!=0.3.2"
-          python -m uv pip install "flax>=0.4.1"
-          python -m uv pip install "jaxlib>=0.1.65"
-          python -m uv pip install pytest
+          python -m pip install --upgrade pip
+          pip install -e .
+          pip install "jax[cpu]>=0.2.16,!=0.3.2"
+          pip install "flax>=0.4.1"
+          pip install "jaxlib>=0.1.65"
+          pip install pytest
       - name: Check for soft dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           pytest tests/others/test_dependencies.py

.github/workflows/pr_quality.yml

@@ -0,0 +1,49 @@
+name: Run code quality checks
+
+on:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - main
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  check_code_quality:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: |
+          ruff check examples tests src utils scripts
+          ruff format examples tests src utils scripts --check
+
+  check_repository_consistency:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: |
+          python utils/check_copies.py
+          python utils/check_dummies.py
+          make deps_table_check_updated

.github/workflows/pr_test_fetcher.yml

@@ -33,8 +33,7 @@ jobs:
     - name: Install dependencies
       run: |
         apt-get update && apt-get install libsndfile1-dev libgl1 -y
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
+        python -m pip install -e .[quality,test]
     - name: Environment
       run: |
         python utils/print_env.py
@@ -90,18 +89,15 @@ jobs:
     - name: Install dependencies
       run: |
         apt-get update && apt-get install libsndfile1-dev libgl1 -y
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m pip install -e [quality,test]
+        python -m pip install -e .[quality,test]
         python -m pip install accelerate
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run all selected tests on CPU
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --dist=loadfile -v --make-reports=${{ matrix.modules }}_tests_cpu ${{ fromJson(needs.setup_pr_tests.outputs.test_map)[matrix.modules] }}
 
     - name: Failure short reports
@@ -148,18 +144,15 @@ jobs:
     - 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 pip install -e [quality,test]
+        python -m pip install -e .[quality,test]
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run Hub tests for models, schedulers, and pipelines on a staging env
       if: ${{ matrix.config.framework == 'hub_tests_pytorch' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         HUGGINGFACE_CO_STAGING=true python -m pytest \
           -m "is_staging_test" \
           --make-reports=tests_${{ matrix.config.report }} \

.github/workflows/pr_test_peft_backend.yml

@@ -4,9 +4,6 @@ on:
   pull_request:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
-      - "tests/**.py"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@@ -19,44 +16,7 @@ env:
   PYTEST_TIMEOUT: 60
 
 jobs:
-  check_code_quality:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: |
-          ruff check examples tests src utils scripts
-          ruff format examples tests src utils scripts --check
-
-  check_repository_consistency:
-    needs: check_code_quality
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: |
-          python utils/check_copies.py
-          python utils/check_dummies.py
-          make deps_table_check_updated
-
   run_fast_tests:
-    needs: [check_code_quality, check_repository_consistency]
     strategy:
       fail-fast: false
       matrix:
@@ -84,24 +44,21 @@ jobs:
     - name: Install dependencies
       run: |
         apt-get update && apt-get install libsndfile1-dev libgl1 -y
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
+        python -m pip install -e .[quality,test]
         if [ "${{ matrix.lib-versions }}" == "main" ]; then
-            python -m uv pip install -U peft@git+https://github.com/huggingface/peft.git
-            python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git
-            python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+            python -m pip install -U git+https://github.com/huggingface/peft.git
+            python -m pip install -U git+https://github.com/huggingface/transformers.git
+            python -m pip install -U git+https://github.com/huggingface/accelerate.git
         else
-            python -m uv pip install -U peft transformers accelerate
+            python -m pip install -U peft transformers accelerate
         fi
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run fast PyTorch LoRA CPU tests with PEFT backend
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
           -s -v \
           --make-reports=tests_${{ matrix.config.report }} \

.github/workflows/pr_tests.yml

@@ -4,14 +4,6 @@ on:
   pull_request:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
-      - "benchmarks/**.py"
-      - "examples/**.py"
-      - "scripts/**.py"
-      - "tests/**.py"
-      - ".github/**.yml"
-      - "utils/**.py"
   push:
     branches:
       - ci-*
@@ -27,44 +19,7 @@ env:
   PYTEST_TIMEOUT: 60
 
 jobs:
-  check_code_quality:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: |
-          ruff check examples tests src utils scripts
-          ruff format examples tests src utils scripts --check
-
-  check_repository_consistency:
-    needs: check_code_quality
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: |
-          python utils/check_copies.py
-          python utils/check_dummies.py
-          make deps_table_check_updated
-
   run_fast_tests:
-    needs: [check_code_quality, check_repository_consistency]
     strategy:
       fail-fast: false
       matrix:
@@ -111,19 +66,16 @@ jobs:
     - 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
+        python -m pip install -e .[quality,test]
+        python -m pip install accelerate
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run fast PyTorch Pipeline CPU tests
       if: ${{ matrix.config.framework == 'pytorch_pipelines' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "not Flax and not Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -132,7 +84,6 @@ jobs:
     - name: Run fast PyTorch Model Scheduler CPU tests
       if: ${{ matrix.config.framework == 'pytorch_models' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "not Flax and not Onnx and not Dependency" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -141,7 +92,6 @@ jobs:
     - name: Run fast Flax TPU tests
       if: ${{ matrix.config.framework == 'flax' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "Flax" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -150,8 +100,7 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install peft
+        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples
@@ -168,7 +117,6 @@ jobs:
         path: reports
 
   run_staging_tests:
-    needs: [check_code_quality, check_repository_consistency]
     strategy:
       fail-fast: false
       matrix:
@@ -200,18 +148,15 @@ jobs:
     - 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 pip install -e .[quality,test]
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run Hub tests for models, schedulers, and pipelines on a staging env
       if: ${{ matrix.config.framework == 'hub_tests_pytorch' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         HUGGINGFACE_CO_STAGING=true python -m pytest \
           -m "is_staging_test" \
           --make-reports=tests_${{ matrix.config.report }} \

.github/workflows/pr_torch_dependency_test.yml

@@ -4,8 +4,6 @@ on:
   pull_request:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
   push:
     branches:
       - main
@@ -25,12 +23,10 @@ jobs:
           python-version: "3.8"
       - name: Install dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m pip install --upgrade pip uv
-          python -m uv pip install -e .
-          python -m uv pip install torch torchvision torchaudio
-          python -m uv pip install pytest
+          python -m pip install --upgrade pip
+          pip install -e .
+          pip install torch torchvision torchaudio
+          pip install pytest
       - name: Check for soft dependencies
         run: |
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           pytest tests/others/test_dependencies.py

.github/workflows/push_tests.yml

@@ -4,10 +4,7 @@ on:
   push:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
-      - "examples/**.py"
-      - "tests/**.py"
+
 
 env:
   DIFFUSERS_IS_CI: yes
@@ -21,7 +18,7 @@ env:
 jobs:
   setup_torch_cuda_pipeline_matrix:
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    runs-on: docker-gpu
     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
@@ -35,9 +32,8 @@ jobs:
       - 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 pip install -e .[quality,test]
+          python -m pip install git+https://github.com/huggingface/accelerate.git
 
       - name: Environment
         run: |
@@ -62,6 +58,7 @@ jobs:
     needs: setup_torch_cuda_pipeline_matrix
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         module: ${{ fromJson(needs.setup_torch_cuda_pipeline_matrix.outputs.pipeline_test_matrix) }}
     runs-on: [single-gpu, nvidia-gpu, t4, ci]
@@ -79,9 +76,8 @@ jobs:
       - name: Install dependencies
         run: |
           apt-get update && apt-get install libsndfile1-dev libgl1 -y
-          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-          python -m uv pip install -e [quality,test]
-          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+          python -m pip install -e .[quality,test]
+          python -m pip install git+https://github.com/huggingface/accelerate.git
       - name: Environment
         run: |
           python utils/print_env.py
@@ -129,9 +125,8 @@ jobs:
     - 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 pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate.git
 
     - name: Environment
       run: |
@@ -179,10 +174,9 @@ jobs:
     - 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 pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate.git
+        python -m pip install git+https://github.com/huggingface/peft.git
 
     - name: Environment
       run: |
@@ -230,9 +224,8 @@ jobs:
     - 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 pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate.git
 
     - name: Environment
       run: |
@@ -278,9 +271,8 @@ jobs:
     - 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 pip install -e .[quality,test]
+        python -m pip install git+https://github.com/huggingface/accelerate.git
 
     - name: Environment
       run: |
@@ -328,8 +320,7 @@ jobs:
         nvidia-smi
     - name: Install dependencies
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test,training]
+        python -m pip install -e .[quality,test,training]
     - name: Environment
       run: |
         python utils/print_env.py
@@ -369,8 +360,7 @@ jobs:
         nvidia-smi
     - name: Install dependencies
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test,training]
+        python -m pip install -e .[quality,test,training]
     - name: Environment
       run: |
         python utils/print_env.py
@@ -411,19 +401,16 @@ jobs:
 
     - name: Install dependencies
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test,training]
+        python -m pip install -e .[quality,test,training]
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run example tests on GPU
       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 --make-reports=examples_torch_cuda examples/
 
     - name: Failure short reports

.github/workflows/push_tests_fast.yml

@@ -4,10 +4,6 @@ on:
   push:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
-      - "examples/**.py"
-      - "tests/**.py"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@@ -69,18 +65,15 @@ jobs:
     - 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 pip install -e .[quality,test]
 
     - name: Environment
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python utils/print_env.py
 
     - name: Run fast PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "not Flax and not Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -89,7 +82,6 @@ jobs:
     - name: Run fast Flax TPU tests
       if: ${{ matrix.config.framework == 'flax' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "Flax" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -98,7 +90,6 @@ jobs:
     - name: Run fast ONNXRuntime CPU tests
       if: ${{ matrix.config.framework == 'onnxruntime' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
@@ -107,8 +98,7 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install peft
+        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples

.github/workflows/push_tests_mps.yml

@@ -4,9 +4,6 @@ on:
   push:
     branches:
       - main
-    paths:
-      - "src/diffusers/**.py"
-      - "tests/**.py"
 
 env:
   DIFFUSERS_IS_CI: yes
@@ -44,11 +41,11 @@ jobs:
     - name: Install dependencies
       shell: arch -arch arm64 bash {0}
       run: |
-        ${CONDA_RUN} python -m pip install --upgrade pip uv
-        ${CONDA_RUN} python -m uv pip install -e [quality,test]
-        ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio
-        ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
-        ${CONDA_RUN} python -m uv pip install transformers --upgrade
+        ${CONDA_RUN} python -m pip install --upgrade pip
+        ${CONDA_RUN} python -m pip install -e .[quality,test]
+        ${CONDA_RUN} python -m pip install torch torchvision torchaudio
+        ${CONDA_RUN} python -m pip install git+https://github.com/huggingface/accelerate.git
+        ${CONDA_RUN} python -m pip install transformers --upgrade
 
     - name: Environment
       shell: arch -arch arm64 bash {0}

benchmarks/base_classes.py

@@ -141,7 +141,6 @@ class LCMLoRATextToImageBenchmark(TextToImageBenchmark):
         super().__init__(args)
         self.pipe.load_lora_weights(self.lora_id)
         self.pipe.fuse_lora()
-        self.pipe.unload_lora_weights()
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
 
     def get_result_filepath(self, args):
@@ -236,35 +235,6 @@ class InpaintingBenchmark(ImageToImageBenchmark):
         )
 
 
-class IPAdapterTextToImageBenchmark(TextToImageBenchmark):
-    url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png"
-    image = load_image(url)
-
-    def __init__(self, args):
-        pipe = self.pipeline_class.from_pretrained(args.ckpt, torch_dtype=torch.float16).to("cuda")
-        pipe.load_ip_adapter(
-            args.ip_adapter_id[0],
-            subfolder="models" if "sdxl" not in args.ip_adapter_id[1] else "sdxl_models",
-            weight_name=args.ip_adapter_id[1],
-        )
-
-        if args.run_compile:
-            pipe.unet.to(memory_format=torch.channels_last)
-            print("Run torch compile")
-            pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-        pipe.set_progress_bar_config(disable=True)
-        self.pipe = pipe
-
-    def run_inference(self, pipe, args):
-        _ = pipe(
-            prompt=PROMPT,
-            ip_adapter_image=self.image,
-            num_inference_steps=args.num_inference_steps,
-            num_images_per_prompt=args.batch_size,
-        )
-
-
 class ControlNetBenchmark(TextToImageBenchmark):
     pipeline_class = StableDiffusionControlNetPipeline
     aux_network_class = ControlNetModel

benchmarks/benchmark_ip_adapters.py

@@ -1,32 +0,0 @@
-import argparse
-import sys
-
-
-sys.path.append(".")
-from base_classes import IPAdapterTextToImageBenchmark  # noqa: E402
-
-
-IP_ADAPTER_CKPTS = {
-    "runwayml/stable-diffusion-v1-5": ("h94/IP-Adapter", "ip-adapter_sd15.bin"),
-    "stabilityai/stable-diffusion-xl-base-1.0": ("h94/IP-Adapter", "ip-adapter_sdxl.bin"),
-}
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--ckpt",
-        type=str,
-        default="runwayml/stable-diffusion-v1-5",
-        choices=list(IP_ADAPTER_CKPTS.keys()),
-    )
-    parser.add_argument("--batch_size", type=int, default=1)
-    parser.add_argument("--num_inference_steps", type=int, default=50)
-    parser.add_argument("--model_cpu_offload", action="store_true")
-    parser.add_argument("--run_compile", action="store_true")
-    args = parser.parse_args()
-
-    args.ip_adapter_id = IP_ADAPTER_CKPTS[args.ckpt]
-    benchmark_pipe = IPAdapterTextToImageBenchmark(args)
-    args.ckpt = f"{args.ckpt} (IP-Adapter)"
-    benchmark_pipe.benchmark(args)

benchmarks/run_all.py

@@ -72,7 +72,7 @@ def main():
                 command += " --run_compile"
                 run_command(command.split())
 
-        elif file in ["benchmark_sd_inpainting.py", "benchmark_ip_adapters.py"]:
+        elif file == "benchmark_sd_inpainting.py":
             sdxl_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
             command = f"python {file} --ckpt {sdxl_ckpt}"
             run_command(command.split())

docker/diffusers-flax-cpu/Dockerfile

@@ -23,13 +23,13 @@ ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 # follow the instructions here: https://cloud.google.com/tpu/docs/run-in-container#train_a_jax_model_in_a_docker_container
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3 -m uv pip install --upgrade --no-cache-dir \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --upgrade --no-cache-dir \
         clu \
         "jax[cpu]>=0.2.16,!=0.3.2" \
         "flax>=0.4.1" \
         "jaxlib>=0.1.65" && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docker/diffusers-flax-tpu/Dockerfile

@@ -23,15 +23,15 @@ ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 # follow the instructions here: https://cloud.google.com/tpu/docs/run-in-container#train_a_jax_model_in_a_docker_container
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
         "jax[tpu]>=0.2.16,!=0.3.2" \
         -f https://storage.googleapis.com/jax-releases/libtpu_releases.html && \
-    python3 -m uv pip install --upgrade --no-cache-dir \
+    python3 -m pip install --upgrade --no-cache-dir \
         clu \
         "flax>=0.4.1" \
         "jaxlib>=0.1.65" && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docker/diffusers-onnxruntime-cpu/Dockerfile

@@ -22,14 +22,14 @@ RUN python3 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3 -m uv pip install --no-cache-dir \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
         torch==2.1.2 \
         torchvision==0.16.2 \
         torchaudio==2.1.2 \
         onnxruntime \
         --extra-index-url https://download.pytorch.org/whl/cpu && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docker/diffusers-onnxruntime-cuda/Dockerfile

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
+FROM nvidia/cuda:11.6.2-cudnn8-devel-ubuntu20.04
 LABEL maintainer="Hugging Face"
 LABEL repository="diffusers"
 
@@ -22,14 +22,14 @@ RUN python3 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3 -m uv pip install --no-cache-dir \
-        torch \
-        torchvision \
-        torchaudio \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
+        torch==2.1.2 \
+        torchvision==0.16.2 \
+        torchaudio==2.1.2 \
         "onnxruntime-gpu>=1.13.1" \
         --extra-index-url https://download.pytorch.org/whl/cu117 && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -24,8 +24,8 @@ RUN python3.9 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3.9 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3.9 -m uv pip install --no-cache-dir \
+RUN python3.9 -m pip install --no-cache-dir --upgrade pip && \
+    python3.9 -m pip install --no-cache-dir \
     torch \
     torchvision \
     torchaudio \

docker/diffusers-pytorch-cpu/Dockerfile

@@ -23,14 +23,14 @@ RUN python3 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3 -m uv pip install --no-cache-dir \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
         torch \
         torchvision \
         torchaudio \
         invisible_watermark \
         --extra-index-url https://download.pytorch.org/whl/cpu && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docker/diffusers-pytorch-cuda/Dockerfile

@@ -23,8 +23,8 @@ RUN python3 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
-    python3 -m uv pip install --no-cache-dir \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
+    python3 -m pip install --no-cache-dir \
     torch \
     torchvision \
     torchaudio \

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -23,13 +23,13 @@ RUN python3 -m venv /opt/venv
 ENV PATH="/opt/venv/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
-RUN python3 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
+RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
         torch \
         torchvision \
         torchaudio \
         invisible_watermark && \
-    python3 -m uv pip install --no-cache-dir \
+    python3 -m pip install --no-cache-dir \
         accelerate \
         datasets \
         hf-doc-builder \

docs/source/en/_toctree.yml

@@ -318,8 +318,6 @@
       title: Semantic Guidance
     - local: api/pipelines/shap_e
       title: Shap-E
-    - local: api/pipelines/stable_cascade
-      title: Stable Cascade
     - sections:
       - local: api/pipelines/stable_diffusion/overview
         title: Overview
@@ -420,8 +418,6 @@
       title: ScoreSdeVeScheduler
     - local: api/schedulers/score_sde_vp
       title: ScoreSdeVpScheduler
-    - local: api/schedulers/tcd
-      title: TCDScheduler
     - local: api/schedulers/unipc
       title: UniPCMultistepScheduler
     - local: api/schedulers/vq_diffusion

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

@@ -1,18 +1,6 @@
-<!--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.
--->
-
 # Consistency Decoder
 
-Consistency decoder can be used to decode the latents from the denoising UNet in the [`StableDiffusionPipeline`]. This decoder was introduced in the [DALL-E 3 technical report](https://openai.com/dall-e-3).
+Consistency decoder can be used to decode the latents from the denoising UNet in the [`StableDiffusionPipeline`]. This decoder was introduced in the [DALL-E 3 technical report](https://openai.com/dall-e-3). 
 
 The original codebase can be found at [openai/consistencydecoder](https://github.com/openai/consistencydecoder).
 

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

@@ -1,88 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# Stable Cascade
-
-This model is built upon the [Würstchen](https://openreview.net/forum?id=gU58d5QeGv) architecture and its main 
-difference to other models like Stable Diffusion is that it is working at a much smaller latent space. Why is this 
-important? The smaller the latent space, the **faster** you can run inference and the **cheaper** the training becomes. 
-How small is the latent space? Stable Diffusion uses a compression factor of 8, resulting in a 1024x1024 image being 
-encoded to 128x128. Stable Cascade achieves a compression factor of 42, meaning that it is possible to encode a 
-1024x1024 image to 24x24, while maintaining crisp reconstructions. The text-conditional model is then trained in the 
-highly compressed latent space. Previous versions of this architecture, achieved a 16x cost reduction over Stable 
-Diffusion 1.5.
-
-Therefore, this kind of model is well suited for usages where efficiency is important. Furthermore, all known extensions
-like finetuning, LoRA, ControlNet, IP-Adapter, LCM etc. are possible with this method as well.
-
-The original codebase can be found at [Stability-AI/StableCascade](https://github.com/Stability-AI/StableCascade).
-
-## Model Overview
-Stable Cascade consists of three models: Stage A, Stage B and Stage C, representing a cascade to generate images,
-hence the name "Stable Cascade".
-
-Stage A & B are used to compress images, similar to what the job of the VAE is in Stable Diffusion. 
-However, with this setup, a much higher compression of images can be achieved. While the Stable Diffusion models use a 
-spatial compression factor of 8, encoding an image with resolution of 1024 x 1024 to 128 x 128, Stable Cascade achieves 
-a compression factor of 42. This encodes a 1024 x 1024 image to 24 x 24, while being able to accurately decode the 
-image. This comes with the great benefit of cheaper training and inference. Furthermore, Stage C is responsible 
-for generating the small 24 x 24 latents given a text prompt.
-
-## Uses
-
-### Direct Use
-
-The model is intended for research purposes for now. Possible research areas and tasks include
-
-- Research on generative models.
-- Safe deployment of models which have the potential to generate harmful content.
-- Probing and understanding the limitations and biases of generative models.
-- Generation of artworks and use in design and other artistic processes.
-- Applications in educational or creative tools.
-
-Excluded uses are described below.
-
-### Out-of-Scope Use
-
-The model was not trained to be factual or true representations of people or events, 
-and therefore using the model to generate such content is out-of-scope for the abilities of this model.
-The model should not be used in any way that violates Stability AI's [Acceptable Use Policy](https://stability.ai/use-policy).
-
-## Limitations and Bias
-
-### Limitations
-- Faces and people in general may not be generated properly.
-- The autoencoding part of the model is lossy.
-
-
-## StableCascadeCombinedPipeline
-
-[[autodoc]] StableCascadeCombinedPipeline
-	- all
-	- __call__
-
-## StableCascadePriorPipeline
-
-[[autodoc]] StableCascadePriorPipeline
-	- all
-	- __call__
-
-## StableCascadePriorPipelineOutput
-
-[[autodoc]] pipelines.stable_cascade.pipeline_stable_cascade_prior.StableCascadePriorPipelineOutput
-
-## StableCascadeDecoderPipeline
-
-[[autodoc]] StableCascadeDecoderPipeline
-	- all
-	- __call__
-

docs/source/en/api/pipelines/stable_diffusion/sdxl_turbo.md

@@ -21,7 +21,7 @@ The abstract from the paper is:
 ## Tips
 
 - SDXL Turbo uses the exact same architecture as [SDXL](./stable_diffusion_xl), which means it also has the same API. Please refer to the [SDXL](./stable_diffusion_xl) API reference for more details.
-- SDXL Turbo should disable guidance scale by setting `guidance_scale=0.0`.
+- SDXL Turbo should disable guidance scale by setting `guidance_scale=0.0`
 - SDXL Turbo should use `timestep_spacing='trailing'` for the scheduler and use between 1 and 4 steps.
 - SDXL Turbo has been trained to generate images of size 512x512.
 - SDXL Turbo is open-access, but not open-source meaning that one might have to buy a model license in order to use it for commercial applications. Make sure to read the [official model card](https://huggingface.co/stabilityai/sdxl-turbo) to learn more.

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

@@ -1,21 +1,9 @@
-<!--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.
--->
-
 # ConsistencyDecoderScheduler
 
-This scheduler is a part of the [`ConsistencyDecoderPipeline`] and was introduced in [DALL-E 3](https://openai.com/dall-e-3).
+This scheduler is a part of the [`ConsistencyDecoderPipeline`] and was introduced in [DALL-E 3](https://openai.com/dall-e-3). 
 
 The original codebase can be found at [openai/consistency_models](https://github.com/openai/consistency_models).
 
 
 ## ConsistencyDecoderScheduler
-[[autodoc]] schedulers.scheduling_consistency_decoder.ConsistencyDecoderScheduler
+[[autodoc]] schedulers.scheduling_consistency_decoder.ConsistencyDecoderScheduler

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

@@ -1,29 +0,0 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# TCDScheduler 
-
-[Trajectory Consistency Distillation](https://huggingface.co/papers/2402.19159) by Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao and Tat-Jen Cham introduced a Strategic Stochastic Sampling (Algorithm 4) that is capable of generating good samples in a small number of steps. Distinguishing it as an advanced iteration of the multistep scheduler (Algorithm 1) in the [Consistency Models](https://huggingface.co/papers/2303.01469), Strategic Stochastic Sampling specifically tailored for the trajectory consistency function.
-
-The abstract from the paper is:
-
-*Latent Consistency Model (LCM) extends the Consistency Model to the latent space and leverages the guided consistency distillation technique to achieve impressive performance in accelerating text-to-image synthesis. However, we observed that LCM struggles to generate images with both clarity and detailed intricacy. To address this limitation, we initially delve into and elucidate the underlying causes. Our investigation identifies that the primary issue stems from errors in three distinct areas. Consequently, we introduce Trajectory Consistency Distillation (TCD), which encompasses trajectory consistency function and strategic stochastic sampling. The trajectory consistency function diminishes the distillation errors by broadening the scope of the self-consistency boundary condition and endowing the TCD with the ability to accurately trace the entire trajectory of the Probability Flow ODE. Additionally, strategic stochastic sampling is specifically designed to circumvent the accumulated errors inherent in multi-step consistency sampling, which is meticulously tailored to complement the TCD model. Experiments demonstrate that TCD not only significantly enhances image quality at low NFEs but also yields more detailed results compared to the teacher model at high NFEs.*
-
-The original codebase can be found at [jabir-zheng/TCD](https://github.com/jabir-zheng/TCD).
-
-## TCDScheduler
-[[autodoc]] TCDScheduler
-
-
-## TCDSchedulerOutput
-[[autodoc]] schedulers.scheduling_tcd.TCDSchedulerOutput
-

docs/source/en/training/lora.md

@@ -77,7 +77,7 @@ accelerate config default
 
 Or if your environment doesn't support an interactive shell, like a notebook, you can use:
 
-```py
+```bash
 from accelerate.utils import write_basic_config
 
 write_basic_config()
@@ -170,7 +170,7 @@ Aside from setting up the LoRA layers, the training script is more or less the s
 
 Once you've made all your changes or you're okay with the default configuration, you're ready to launch the training script! 🚀
 
-Let's train on the [Pokémon BLIP captions](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions) dataset to generate our own Pokémon. Set the environment variables `MODEL_NAME` and `DATASET_NAME` to the model and dataset respectively. You should also specify where to save the model in `OUTPUT_DIR`, and the name of the model to save to on the Hub with `HUB_MODEL_ID`. The script creates and saves the following files to your repository:
+Let's train on the [Pokémon BLIP captions](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions) dataset to generate our yown Pokémon. Set the environment variables `MODEL_NAME` and `DATASET_NAME` to the model and dataset respectively. You should also specify where to save the model in `OUTPUT_DIR`, and the name of the model to save to on the Hub with `HUB_MODEL_ID`. The script creates and saves the following files to your repository:
 
 - saved model checkpoints
 - `pytorch_lora_weights.safetensors` (the trained LoRA weights)

docs/source/en/tutorials/using_peft_for_inference.md

@@ -169,7 +169,7 @@ list_adapters_component_wise
 
 If you want to compile your model with `torch.compile` make sure to first fuse the LoRA weights into the base model and unload them.
 
-```diff
+```py
 pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
 pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
 
@@ -178,16 +178,12 @@ pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
 pipe.fuse_lora()
 pipe.unload_lora_weights()
 
-+ pipe.unet.to(memory_format=torch.channels_last)
-+ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+pipe = torch.compile(pipe)
 
 prompt = "toy_face of a hacker with a hoodie, pixel art"
 image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
 ```
 
-> [!TIP]
-> You can refer to the `torch.compile()` section [here](https://huggingface.co/docs/diffusers/main/en/optimization/torch2.0#torchcompile) and [here](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) for more elaborate examples.
-
 ## Fusing adapters into the model
 
 You can use PEFT to easily fuse/unfuse multiple adapters directly into the model weights (both UNet and text encoder) using the [`~diffusers.loaders.LoraLoaderMixin.fuse_lora`] method, which can lead to a speed-up in inference and lower VRAM usage.

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

@@ -31,9 +31,9 @@ def callback_dynamic_cfg(pipe, step_index, timestep, callback_kwargs):
                 prompt_embeds = callback_kwargs["prompt_embeds"]
                 prompt_embeds = prompt_embeds.chunk(2)[-1]
 
-                # update guidance_scale and prompt_embeds
-                pipe._guidance_scale = 0.0
-                callback_kwargs["prompt_embeds"] = prompt_embeds
+        # update guidance_scale and prompt_embeds
+        pipe._guidance_scale = 0.0
+        callback_kwargs["prompt_embeds"] = prompt_embeds
         return callback_kwargs
 ```
 

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

@@ -128,7 +128,7 @@ seed = 2023
 # The values come from
 # https://github.com/lyn-rgb/FreeU_Diffusers#video-pipelines
 pipe.enable_freeu(b1=1.2, b2=1.4, s1=0.9, s2=0.2)
-video_frames = pipe(prompt, height=320, width=576, num_frames=30, generator=torch.manual_seed(seed)).frames[0]
+video_frames = pipe(prompt, height=320, width=576, num_frames=30, generator=torch.manual_seed(seed)).frames
 export_to_video(video_frames, "astronaut_rides_horse.mp4")
 ```
 

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

@@ -48,10 +48,10 @@ Create a text prompt and load an image prompt before passing them to the pipelin
 image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ip_adapter_diner.png")
 generator = torch.Generator(device="cpu").manual_seed(0)
 images = pipeline(
-    prompt="a polar bear sitting in a chair drinking a milkshake",
+    prompt="a polar bear sitting in a chair drinking a milkshake", 
     ip_adapter_image=image,
     negative_prompt="deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
-    num_inference_steps=100,
+    num_inference_steps=100, 
     generator=generator,
 ).images
 images[0]
@@ -234,39 +234,6 @@ export_to_gif(frames, "gummy_bear.gif")
 > [!TIP]
 > While calling `load_ip_adapter()`, pass `low_cpu_mem_usage=True` to speed up the loading time.
 
-All the pipelines supporting IP-Adapter accept a `ip_adapter_image_embeds` argument. If you need to run the IP-Adapter multiple times with the same image, you can encode the image once and save the embedding to the disk.
-
-```py
-image_embeds = pipeline.prepare_ip_adapter_image_embeds(
-    ip_adapter_image=image,
-    ip_adapter_image_embeds=None,
-    device="cuda",
-    num_images_per_prompt=1,
-    do_classifier_free_guidance=True,
-)
-
-torch.save(image_embeds, "image_embeds.ipadpt")
-```
-
-Load the image embedding and pass it to the pipeline as `ip_adapter_image_embeds`
-
-> [!TIP]
-> ComfyUI image embeddings for IP-Adapters are fully compatible in Diffusers and should work out-of-box.
-
-```py
-image_embeds = torch.load("image_embeds.ipadpt")
-images = pipeline(
-    prompt="a polar bear sitting in a chair drinking a milkshake",
-    ip_adapter_image_embeds=image_embeds,
-    negative_prompt="deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
-    num_inference_steps=100,
-    generator=generator,
-).images
-```
-
-> [!TIP]
-> If you use IP-Adapter with `ip_adapter_image_embedding` instead of `ip_adapter_image`, you can choose not to load an image encoder by passing `image_encoder_folder=None` to `load_ip_adapter()`. 
-
 ## Specific use cases
 
 IP-Adapter's image prompting and compatibility with other adapters and models makes it a versatile tool for a variety of use cases. This section covers some of the more popular applications of IP-Adapter, and we can't wait to see what you come up with!
@@ -303,7 +270,7 @@ generator = torch.Generator(device="cpu").manual_seed(26)
 image = pipeline(
     prompt="A photo of Einstein as a chef, wearing an apron, cooking in a French restaurant",
     ip_adapter_image=image,
-    negative_prompt="lowres, bad anatomy, worst quality, low quality",
+    negative_prompt="lowres, bad anatomy, worst quality, low quality", 
     num_inference_steps=100,
     generator=generator,
 ).images[0]
@@ -337,7 +304,7 @@ from transformers import CLIPVisionModelWithProjection
 from diffusers.utils import load_image
 
 image_encoder = CLIPVisionModelWithProjection.from_pretrained(
-    "h94/IP-Adapter",
+    "h94/IP-Adapter", 
     subfolder="models/image_encoder",
     torch_dtype=torch.float16,
 )
@@ -356,8 +323,8 @@ pipeline = AutoPipelineForText2Image.from_pretrained(
 )
 pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
 pipeline.load_ip_adapter(
-  "h94/IP-Adapter",
-  subfolder="sdxl_models",
+  "h94/IP-Adapter", 
+  subfolder="sdxl_models", 
   weight_name=["ip-adapter-plus_sdxl_vit-h.safetensors", "ip-adapter-plus-face_sdxl_vit-h.safetensors"]
 )
 pipeline.set_ip_adapter_scale([0.7, 0.3])
@@ -369,7 +336,7 @@ Load an image prompt and a folder containing images of a certain style you want
 ```py
 face_image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/women_input.png")
 style_folder = "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/style_ziggy"
-style_images = [load_image(f"{style_folder}/img{i}.png") for i in range(10)]
+style_images =  [load_image(f"{style_folder}/img{i}.png") for i in range(10)]
 ```
 
 <div class="flex flex-row gap-4">
@@ -391,11 +358,10 @@ generator = torch.Generator(device="cpu").manual_seed(0)
 image = pipeline(
     prompt="wonderwoman",
     ip_adapter_image=[style_images, face_image],
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
     num_inference_steps=50, num_images_per_prompt=1,
     generator=generator,
 ).images[0]
-image
 ```
 
 <div class="flex justify-center">
@@ -413,14 +379,14 @@ from diffusers import DiffusionPipeline, LCMScheduler
 import torch
 from diffusers.utils import load_image
 
-model_id = "sd-dreambooth-library/herge-style"
+model_id =  "sd-dreambooth-library/herge-style"
 lcm_lora_id = "latent-consistency/lcm-lora-sdv1-5"
 
 pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
 
 pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
 pipeline.load_lora_weights(lcm_lora_id)
-pipeline.scheduler = LCMScheduler.from_config(pipeline.scheduler.config)
+pipeline.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
 pipeline.enable_model_cpu_offload()
 ```
 
@@ -489,13 +455,13 @@ Pass the depth map and IP-Adapter image to the pipeline to generate an image.
 ```py
 generator = torch.Generator(device="cpu").manual_seed(33)
 image = pipeline(
-    prompt="best quality, high quality",
+    prompt="best quality, high quality", 
     image=depth_map,
     ip_adapter_image=ip_adapter_image,
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
     num_inference_steps=50,
     generator=generator,
-).images[0]
+).image[0]
 image
 ```
 
@@ -545,7 +511,8 @@ If you have more than one IP-Adapter image, load them into a list, ensuring each
 face_image1 = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ip_mask_girl1.png")
 face_image2 = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ip_mask_girl2.png")
 
-ip_images = [[face_image1], [face_image2]]
+ip_images =[[image1], [image2]]
+
 ```
 
 <div class="flex flex-row gap-4">
@@ -562,19 +529,19 @@ ip_images = [[face_image1], [face_image2]]
 Pass preprocessed masks to the pipeline using `cross_attention_kwargs` as shown below:
 
 ```py
+
 pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name=["ip-adapter-plus-face_sdxl_vit-h.safetensors"] * 2)
 pipeline.set_ip_adapter_scale([0.7] * 2)
 generator = torch.Generator(device="cpu").manual_seed(0)
-num_images = 1
+num_images=1
 
 image = pipeline(
     prompt="2 girls",
     ip_adapter_image=ip_images,
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
-    num_inference_steps=20, num_images_per_prompt=num_images,
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
+    num_inference_steps=20, num_images_per_prompt=num_images, 
     generator=generator, cross_attention_kwargs={"ip_adapter_masks": masks}
 ).images[0]
-image
 ```
 
 <div class="flex justify-center">

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

@@ -340,9 +340,9 @@ Once loaded, you can use the pipeline with an image and text prompt to guide the
 image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png")
 generator = torch.Generator(device="cpu").manual_seed(33)
 images = pipeline(
-    prompt='best quality, high quality, wearing sunglasses',
+    prompt='best quality, high quality, wearing sunglasses', 
     ip_adapter_image=image,
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
     num_inference_steps=50,
     generator=generator,
 ).images[0]
@@ -355,13 +355,11 @@ images
 
 ### IP-Adapter Plus
 
-IP-Adapter relies on an image encoder to generate image features. If the IP-Adapter repository contains an `image_encoder` subfolder, the image encoder is automatically loaded and registered to the pipeline. Otherwise, you'll need to explicitly load the image encoder with a [`~transformers.CLIPVisionModelWithProjection`] model and pass it to the pipeline.
+IP-Adapter relies on an image encoder to generate image features. If the IP-Adapter repository contains a `image_encoder` subfolder, the image encoder is automatically loaded and registed to the pipeline. Otherwise, you'll need to explicitly load the image encoder with a [`~transformers.CLIPVisionModelWithProjection`] model and pass it to the pipeline.
 
 This is the case for *IP-Adapter Plus* checkpoints which use the ViT-H image encoder.
 
 ```py
-from transformers import CLIPVisionModelWithProjection
-
 image_encoder = CLIPVisionModelWithProjection.from_pretrained(
     "h94/IP-Adapter",
     subfolder="models/image_encoder",

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

@@ -31,31 +31,29 @@ Before you begin, make sure you have the following libraries installed:
 Model weights may be stored in separate subfolders on the Hub or locally, in which case, you should use the [`~StableDiffusionXLPipeline.from_pretrained`] method:
 
 ```py
-from diffusers import AutoPipelineForText2Image
+from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
 import torch
 
 pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
 pipeline = pipeline.to("cuda")
 ```
 
-You can also use the [`~StableDiffusionXLPipeline.from_single_file`] method to load a model checkpoint stored in a single file format (`.ckpt` or `.safetensors`) from the Hub or locally. For this loading method, you need to set `timestep_spacing="trailing"` (feel free to experiment with the other scheduler config values to get better results):
+You can also use the [`~StableDiffusionXLPipeline.from_single_file`] method to load a model checkpoint stored in a single file format (`.ckpt` or `.safetensors`) from the Hub or locally:
 
 ```py
-from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
+from diffusers import StableDiffusionXLPipeline
 import torch
 
 pipeline = StableDiffusionXLPipeline.from_single_file(
-    "https://huggingface.co/stabilityai/sdxl-turbo/blob/main/sd_xl_turbo_1.0_fp16.safetensors",
-    torch_dtype=torch.float16, variant="fp16")
+    "https://huggingface.co/stabilityai/sdxl-turbo/blob/main/sd_xl_turbo_1.0_fp16.safetensors", torch_dtype=torch.float16)
 pipeline = pipeline.to("cuda")
-pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing")
 ```
 
 ## Text-to-image
 
 For text-to-image, pass a text prompt. By default, SDXL Turbo generates a 512x512 image, and that resolution gives the best results. You can try setting the `height` and `width` parameters to 768x768 or 1024x1024, but you should expect quality degradations when doing so.
 
-Make sure to set `guidance_scale` to 0.0 to disable, as the model was trained without it. A single inference step is enough to generate high quality images.
+Make sure to set `guidance_scale` to 0.0 to disable, as the model was trained without it. A single inference step is enough to generate high quality images. 
 Increasing the number of steps to 2, 3 or 4 should improve image quality.
 
 ```py
@@ -77,7 +75,7 @@ image
 
 ## Image-to-image
 
-For image-to-image generation, make sure that `num_inference_steps * strength` is larger or equal to 1.
+For image-to-image generation, make sure that `num_inference_steps * strength` is larger or equal to 1. 
 The image-to-image pipeline will run for `int(num_inference_steps * strength)` steps, e.g. `0.5 * 2.0 = 1` step in
 our example below.
 
@@ -86,14 +84,14 @@ from diffusers import AutoPipelineForImage2Image
 from diffusers.utils import load_image, make_image_grid
 
 # use from_pipe to avoid consuming additional memory when loading a checkpoint
-pipeline_image2image = AutoPipelineForImage2Image.from_pipe(pipeline_text2image).to("cuda")
+pipeline = AutoPipelineForImage2Image.from_pipe(pipeline_text2image).to("cuda")
 
 init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
 init_image = init_image.resize((512, 512))
 
 prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k"
 
-image = pipeline_image2image(prompt, image=init_image, strength=0.5, guidance_scale=0.0, num_inference_steps=2).images[0]
+image = pipeline(prompt, image=init_image, strength=0.5, guidance_scale=0.0, num_inference_steps=2).images[0]
 make_image_grid([init_image, image], rows=1, cols=2)
 ```
 
@@ -103,7 +101,7 @@ make_image_grid([init_image, image], rows=1, cols=2)
 
 ## Speed-up SDXL Turbo even more
 
-- Compile the UNet if you are using PyTorch version 2.0 or higher. The first inference run will be very slow, but subsequent ones will be much faster.
+- Compile the UNet if you are using PyTorch version 2 or better. The first inference run will be very slow, but subsequent ones will be much faster.
 
 ```py
 pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

examples/advanced_diffusion_training/README.md

@@ -80,7 +80,8 @@ To do so, just specify `--train_text_encoder_ti` while launching training (for r
 Please keep the following points in mind:
 
 * SDXL has two text encoders. So, we fine-tune both using LoRA.
-* When not fine-tuning the text encoders, we ALWAYS precompute the text embeddings to save memory.
+* When not fine-tuning the text encoders, we ALWAYS precompute the text embeddings to save memoםהקרry.
+
 
 ### 3D icon example
 
@@ -233,32 +234,6 @@ In ComfyUI we will load a LoRA and a textual embedding at the same time.
 
 SDXL's VAE is known to suffer from numerical instability issues. This is why we also expose a CLI argument namely `--pretrained_vae_model_name_or_path` that lets you specify the location of a better VAE (such as [this one](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).
 
-### DoRA training 
-The advanced script now supports DoRA training too!
-> Proposed in [DoRA: Weight-Decomposed Low-Rank Adaptation](https://arxiv.org/abs/2402.09353), 
-**DoRA** is very similar to LoRA, except it decomposes the pre-trained weight into two components, **magnitude** and **direction** and employs LoRA for _directional_ updates to efficiently minimize the number of trainable parameters. 
-The authors found that by using DoRA, both the learning capacity and training stability of LoRA are enhanced without any additional overhead during inference. 
-
-> [!NOTE]
-> 💡DoRA training is still _experimental_  
-> and is likely to require different hyperparameter values to perform best compared to a LoRA.
-> Specifically, we've noticed 2 differences to take into account your training: 
-> 1. **LoRA seem to converge faster than DoRA** (so a set of parameters that may lead to overfitting when training a LoRA may be working well for a DoRA)
-> 2. **DoRA quality superior to LoRA especially in lower ranks** the difference in quality of DoRA of rank 8 and LoRA of rank 8 appears to be more significant than when training ranks of 32 or 64 for example.  
-> 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: 
-```bash
-pip install git+https://github.com/huggingface/peft.git
-```
-2. Enable DoRA training by adding this flag
-```bash
---use_dora
-```
-**Inference** 
-The inference is the same as if you train a regular LoRA 🤗
-
 
 ### Tips and Tricks
 Check out [these recommended practices](https://huggingface.co/blog/sdxl_lora_advanced_script#additional-good-practices)

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -77,7 +77,6 @@ logger = get_logger(__name__)
 
 def save_model_card(
     repo_id: str,
-    use_dora: bool,
     images=None,
     base_model=str,
     train_text_encoder=False,
@@ -89,7 +88,6 @@ def save_model_card(
     vae_path=None,
 ):
     img_str = "widget:\n"
-    lora = "lora" if not use_dora else "dora"
     for i, image in enumerate(images):
         image.save(os.path.join(repo_folder, f"image_{i}.png"))
         img_str += f"""
@@ -141,10 +139,9 @@ to trigger concept `{key}` → use `{tokens}` in your prompt \n
 tags:
 - stable-diffusion
 - stable-diffusion-diffusers
-- diffusers-training
 - text-to-image
 - diffusers
-- {lora}
+- lora
 - template:sd-lora
 {img_str}
 base_model: {base_model}
@@ -654,16 +651,6 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
-    parser.add_argument(
-        "--use_dora",
-        type=bool,
-        action="store_true",
-        default=False,
-        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
-            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
-        ),
-    )
     parser.add_argument(
         "--cache_latents",
         action="store_true",
@@ -1232,7 +1219,6 @@ def main(args):
     unet_lora_config = LoraConfig(
         r=args.rank,
         lora_alpha=args.rank,
-        use_dora=args.use_dora,
         init_lora_weights="gaussian",
         target_modules=["to_k", "to_q", "to_v", "to_out.0"],
     )
@@ -1244,7 +1230,6 @@ def main(args):
         text_lora_config = LoraConfig(
             r=args.rank,
             lora_alpha=args.rank,
-            use_dora=args.use_dora,
             init_lora_weights="gaussian",
             target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
         )
@@ -1970,7 +1955,6 @@ def main(args):
 
         save_model_card(
             model_id if not args.push_to_hub else repo_id,
-            use_dora=args.use_dora,
             images=images,
             base_model=args.pretrained_model_name_or_path,
             train_text_encoder=args.train_text_encoder,

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -81,7 +81,6 @@ logger = get_logger(__name__)
 
 def save_model_card(
     repo_id: str,
-    use_dora: bool,
     images=None,
     base_model=str,
     train_text_encoder=False,
@@ -93,7 +92,6 @@ def save_model_card(
     vae_path=None,
 ):
     img_str = "widget:\n"
-    lora = "lora" if not use_dora else "dora"
     for i, image in enumerate(images):
         image.save(os.path.join(repo_folder, f"image_{i}.png"))
         img_str += f"""
@@ -146,10 +144,9 @@ to trigger concept `{key}` → use `{tokens}` in your prompt \n
 tags:
 - stable-diffusion-xl
 - stable-diffusion-xl-diffusers
-- diffusers-training
 - text-to-image
 - diffusers
-- {lora}
+- lora
 - template:sd-lora
 {img_str}
 base_model: {base_model}
@@ -664,15 +661,6 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
-    parser.add_argument(
-        "--use_dora",
-        action="store_true",
-        default=False,
-        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
-            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
-        ),
-    )
     parser.add_argument(
         "--cache_latents",
         action="store_true",
@@ -1335,7 +1323,6 @@ def main(args):
     unet_lora_config = LoraConfig(
         r=args.rank,
         lora_alpha=args.rank,
-        use_dora=args.use_dora,
         init_lora_weights="gaussian",
         target_modules=["to_k", "to_q", "to_v", "to_out.0"],
     )
@@ -1347,7 +1334,6 @@ def main(args):
         text_lora_config = LoraConfig(
             r=args.rank,
             lora_alpha=args.rank,
-            use_dora=args.use_dora,
             init_lora_weights="gaussian",
             target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
         )
@@ -2206,7 +2192,6 @@ def main(args):
 
         save_model_card(
             model_id if not args.push_to_hub else repo_id,
-            use_dora=args.use_dora,
             images=images,
             base_model=args.pretrained_model_name_or_path,
             train_text_encoder=args.train_text_encoder,

examples/community/README.md

@@ -750,7 +750,7 @@ This example produces the following images:
 ![image](https://user-images.githubusercontent.com/4313860/198328706-295824a4-9856-4ce5-8e66-278ceb42fd29.png)
 
 ### GlueGen Stable Diffusion Pipeline
-GlueGen is a minimal adapter that allow alignment between any encoder (Text Encoder of different language, Multilingual Roberta, AudioClip) and CLIP text encoder used in standard Stable Diffusion model. This method allows easy language adaptation to available english Stable Diffusion checkpoints without the need of an image captioning dataset as well as long training hours.
+GlueGen is a minimal adapter that allow alignment between any encoder (Text Encoder of different language, Multilingual Roberta, AudioClip) and CLIP text encoder used in standard Stable Diffusion model. This method allows easy language adaptation to available english Stable Diffusion checkpoints without the need of an image captioning dataset as well as long training hours. 
 
 Make sure you downloaded `gluenet_French_clip_overnorm_over3_noln.ckpt` for French (there are also pre-trained weights for Chinese, Italian, Japanese, Spanish or train your own) at [GlueGen's official repo](https://github.com/salesforce/GlueGen/tree/main)
 
@@ -782,9 +782,9 @@ if __name__ == "__main__":
     ).to(device)
     pipeline.load_language_adapter("gluenet_French_clip_overnorm_over3_noln.ckpt", num_token=token_max_length, dim=1024, dim_out=768, tensor_norm=tensor_norm)
 
-    prompt = "une voiture sur la plage"
+    prompt = "une voiture sur la plage" 
 
-    generator = torch.Generator(device=device).manual_seed(42)
+    generator = torch.Generator(device=device).manual_seed(42) 
     image = pipeline(prompt, generator=generator).images[0]
     image.save("gluegen_output_fr.png")
 ```
@@ -1755,7 +1755,7 @@ with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
 ```
 
 The following code compares the performance of the original stable diffusion xl pipeline with the ipex-optimized pipeline.
-By using this optimized pipeline, we can get about 1.4-2 times performance boost with BFloat16 on fourth generation of Intel Xeon CPUs,
+By using this optimized pipeline, we can get about 1.4-2 times performance boost with BFloat16 on fourth generation of Intel Xeon CPUs, 
 code-named Sapphire Rapids.
 
 ```python
@@ -1826,7 +1826,7 @@ This approach is using (optional) CoCa model to avoid writing image description.
 
 This SDXL pipeline support unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style.
 
-You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline.
+You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline. 
 
 ```python
 from diffusers import DiffusionPipeline
@@ -3397,7 +3397,7 @@ invert_prompt = "A lying cat"
 input_image = "siamese.jpg"
 steps = 50
 
-# Provide prompt used for generation. Same if reconstruction
+# Provide prompt used for generation. Same if reconstruction 
 prompt = "A lying cat"
 # or different if editing.
 prompt = "A lying dog"
@@ -3493,7 +3493,7 @@ output_frames = pipe(
     mask_end=0.8,
     mask_strength=0.5,
     negative_prompt='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
-).frames[0]
+).frames
 
 export_to_video(
     output_frames, "/path/to/video.mp4", 5)
@@ -3636,8 +3636,8 @@ image = torch.from_numpy(faces[0].normed_embedding).unsqueeze(0)
 images = pipeline(
     prompt="A photo of a girl wearing a black dress, holding red roses in hand, upper body, behind is the Eiffel Tower",
     image_embeds=image,
-    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
-    num_inference_steps=20, num_images_per_prompt=num_images, width=512, height=704,
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality", 
+    num_inference_steps=20, num_images_per_prompt=num_images, width=512, height=704, 
     generator=generator
 ).images
 

examples/community/clip_guided_images_mixing_stable_diffusion.py

@@ -12,12 +12,12 @@ from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTok
 from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
+    DiffusionPipeline,
     DPMSolverMultistepScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     UNet2DConditionModel,
 )
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.utils import PIL_INTERPOLATION
 from diffusers.utils.torch_utils import randn_tensor
@@ -77,7 +77,7 @@ def set_requires_grad(model, value):
         param.requires_grad = value
 
 
-class CLIPGuidedImagesMixingStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
+class CLIPGuidedImagesMixingStableDiffusion(DiffusionPipeline):
     def __init__(
         self,
         vae: AutoencoderKL,
@@ -113,6 +113,16 @@ class CLIPGuidedImagesMixingStableDiffusion(DiffusionPipeline, StableDiffusionMi
         set_requires_grad(self.text_encoder, False)
         set_requires_grad(self.clip_model, False)
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        self.enable_attention_slicing(None)
+
     def freeze_vae(self):
         set_requires_grad(self.vae, False)
 

examples/community/clip_guided_stable_diffusion.py

@@ -10,12 +10,12 @@ from transformers import CLIPImageProcessor, CLIPModel, CLIPTextModel, CLIPToken
 from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
+    DiffusionPipeline,
     DPMSolverMultistepScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     UNet2DConditionModel,
 )
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 
 
@@ -51,7 +51,7 @@ def set_requires_grad(model, value):
         param.requires_grad = value
 
 
-class CLIPGuidedStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
+class CLIPGuidedStableDiffusion(DiffusionPipeline):
     """CLIP guided stable diffusion based on the amazing repo by @crowsonkb and @Jack000
     - https://github.com/Jack000/glid-3-xl
     - https://github.dev/crowsonkb/k-diffusion
@@ -89,6 +89,16 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
         set_requires_grad(self.text_encoder, False)
         set_requires_grad(self.clip_model, False)
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        self.enable_attention_slicing(None)
+
     def freeze_vae(self):
         set_requires_grad(self.vae, False)
 

examples/community/clip_guided_stable_diffusion_img2img.py

@@ -12,12 +12,12 @@ from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextModel, CLIPTok
 from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
+    DiffusionPipeline,
     DPMSolverMultistepScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     UNet2DConditionModel,
 )
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.utils import PIL_INTERPOLATION, deprecate
 from diffusers.utils.torch_utils import randn_tensor
@@ -125,7 +125,7 @@ def set_requires_grad(model, value):
         param.requires_grad = value
 
 
-class CLIPGuidedStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
+class CLIPGuidedStableDiffusion(DiffusionPipeline):
     """CLIP guided stable diffusion based on the amazing repo by @crowsonkb and @Jack000
     - https://github.com/Jack000/glid-3-xl
     - https://github.dev/crowsonkb/k-diffusion
@@ -163,6 +163,16 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
         set_requires_grad(self.text_encoder, False)
         set_requires_grad(self.clip_model, False)
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        self.enable_attention_slicing(None)
+
     def freeze_vae(self):
         set_requires_grad(self.vae, False)
 

examples/community/composable_stable_diffusion.py

@@ -22,7 +22,6 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import (
@@ -33,13 +32,13 @@ from diffusers.schedulers import (
     LMSDiscreteScheduler,
     PNDMScheduler,
 )
-from diffusers.utils import deprecate, logging
+from diffusers.utils import deprecate, is_accelerate_available, logging
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class ComposableStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class ComposableStableDiffusionPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -165,6 +164,62 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin)
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            # TODO(Patrick) - there is currently a bug with cpu offload of nn.Parameter in accelerate
+            # fix by only offloading self.safety_checker for now
+            cpu_offload(self.safety_checker.vision_model, device)
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
         r"""
         Encodes the prompt into text encoder hidden states.

examples/community/gluegen.py

@@ -10,7 +10,6 @@ from diffusers.image_processor import VaeImageProcessor
 from diffusers.loaders import LoraLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -194,7 +193,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class GlueGenStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin, LoraLoaderMixin):
+class GlueGenStableDiffusionPipeline(DiffusionPipeline, LoraLoaderMixin):
     def __init__(
         self,
         vae: AutoencoderKL,
@@ -242,6 +241,35 @@ class GlueGenStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin, Lo
         )
         self.language_adapter.load_state_dict(torch.load(model_path))
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def _adapt_language(self, prompt_embeds: torch.FloatTensor):
         prompt_embeds = prompt_embeds / 3
         prompt_embeds = self.language_adapter(prompt_embeds) * (self.tensor_norm / 2)
@@ -516,6 +544,32 @@ class GlueGenStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin, Lo
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

examples/community/imagic_stable_diffusion.py

@@ -19,7 +19,6 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from diffusers import DiffusionPipeline
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -57,7 +56,7 @@ def preprocess(image):
     return 2.0 * image - 1.0
 
 
-class ImagicStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class ImagicStableDiffusionPipeline(DiffusionPipeline):
     r"""
     Pipeline for imagic image editing.
     See paper here: https://arxiv.org/pdf/2210.09276.pdf
@@ -106,6 +105,31 @@ class ImagicStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     def train(
         self,
         prompt: Union[str, List[str]],

examples/community/img2img_inpainting.py

@@ -129,6 +129,33 @@ class ImageToImageInpaintingPipeline(DiffusionPipeline):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/instaflow_one_step.py

@@ -24,7 +24,7 @@ from diffusers.image_processor import VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -52,9 +52,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
-class InstaFlowPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
-):
+class InstaFlowPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin):
     r"""
     Pipeline for text-to-image generation using Rectified Flow and Euler discretization.
     This customized pipeline is based on StableDiffusionPipeline from the official Diffusers library (0.21.4)
@@ -182,6 +180,35 @@ class InstaFlowPipeline(
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def _encode_prompt(
         self,
         prompt,

examples/community/interpolate_stable_diffusion.py

@@ -7,9 +7,9 @@ import numpy as np
 import torch
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
+from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -46,7 +46,7 @@ def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
     return v2
 
 
-class StableDiffusionWalkPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionWalkPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -120,6 +120,33 @@ class StableDiffusionWalkPipeline(DiffusionPipeline, StableDiffusionMixin):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/ip_adapter_face_id.py

@@ -26,8 +26,9 @@ from diffusers.configuration_utils import FrozenDict
 from diffusers.image_processor import VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.attention_processor import FusedAttnProcessor2_0
 from diffusers.models.lora import LoRALinearLayer, adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -414,12 +415,7 @@ def retrieve_timesteps(
 
 
 class IPAdapterFaceIDStableDiffusionPipeline(
-    DiffusionPipeline,
-    StableDiffusionMixin,
-    TextualInversionLoaderMixin,
-    LoraLoaderMixin,
-    IPAdapterMixin,
-    FromSingleFileMixin,
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
@@ -731,6 +727,35 @@ class IPAdapterFaceIDStableDiffusionPipeline(
             if isinstance(attn_processor, (LoRAIPAdapterAttnProcessor, LoRAIPAdapterAttnProcessor2_0)):
                 attn_processor.scale = scale
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def _encode_prompt(
         self,
         prompt,
@@ -1055,6 +1080,93 @@ class IPAdapterFaceIDStableDiffusionPipeline(
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

examples/community/latent_consistency_interpolate.py

@@ -9,7 +9,7 @@ from diffusers.image_processor import VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import LCMScheduler
 from diffusers.utils import (
@@ -190,7 +190,7 @@ def slerp(
 
 
 class LatentConsistencyModelWalkPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image generation using a latent consistency model.
@@ -273,6 +273,67 @@ class LatentConsistencyModelWalkPipeline(
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
     def encode_prompt(
         self,

examples/community/llm_grounded_diffusion.py

@@ -35,7 +35,6 @@ from diffusers.models.attention import Attention, GatedSelfAttentionDense
 from diffusers.models.attention_processor import AttnProcessor2_0
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines import DiffusionPipeline
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -268,12 +267,7 @@ class AttnProcessorWithHook(AttnProcessor2_0):
 
 
 class LLMGroundedDiffusionPipeline(
-    DiffusionPipeline,
-    StableDiffusionMixin,
-    TextualInversionLoaderMixin,
-    LoraLoaderMixin,
-    IPAdapterMixin,
-    FromSingleFileMixin,
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for layout-grounded text-to-image generation using LLM-grounded Diffusion (LMD+): https://arxiv.org/pdf/2305.13655.pdf.
@@ -1186,6 +1180,39 @@ class LLMGroundedDiffusionPipeline(
     # Below are methods copied from StableDiffusionPipeline
     # The design choice of not inheriting from StableDiffusionPipeline is discussed here: https://github.com/huggingface/diffusers/pull/5993#issuecomment-1834258517
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,
@@ -1495,6 +1522,34 @@ class LLMGroundedDiffusionPipeline(
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

examples/community/lpw_stable_diffusion.py

@@ -13,12 +13,13 @@ from diffusers.configuration_utils import FrozenDict
 from diffusers.image_processor import VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
     deprecate,
+    is_accelerate_available,
+    is_accelerate_version,
     logging,
 )
 from diffusers.utils.torch_utils import randn_tensor
@@ -409,7 +410,7 @@ def preprocess_mask(mask, batch_size, scale_factor=8):
 
 
 class StableDiffusionLongPromptWeightingPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
@@ -533,6 +534,112 @@ class StableDiffusionLongPromptWeightingPipeline(
             requires_safety_checker=requires_safety_checker,
         )
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/lpw_stable_diffusion_xl.py

@@ -26,11 +26,11 @@ from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMix
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from diffusers.models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
 )
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
@@ -545,12 +545,7 @@ def retrieve_timesteps(
 
 
 class SDXLLongPromptWeightingPipeline(
-    DiffusionPipeline,
-    StableDiffusionMixin,
-    FromSingleFileMixin,
-    IPAdapterMixin,
-    LoraLoaderMixin,
-    TextualInversionLoaderMixin,
+    DiffusionPipeline, FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion XL.
@@ -654,6 +649,39 @@ class SDXLLongPromptWeightingPipeline(
         else:
             self.watermark = None
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
@@ -1002,6 +1030,95 @@ class SDXLLongPromptWeightingPipeline(
                 "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
             )
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
         # get the original timestep using init_timestep
         if denoising_start is None:

examples/community/mixture_canvas.py

@@ -12,7 +12,7 @@ from tqdm.auto import tqdm
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
 
@@ -264,7 +264,7 @@ class MaskWeightsBuilder:
         return torch.tile(torch.tensor(weights), (self.nbatch, self.latent_space_dim, 1, 1))
 
 
-class StableDiffusionCanvasPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionCanvasPipeline(DiffusionPipeline):
     """Stable Diffusion pipeline that mixes several diffusers in the same canvas"""
 
     def __init__(

examples/community/multilingual_stable_diffusion.py

@@ -11,9 +11,9 @@ from transformers import (
     pipeline,
 )
 
+from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -48,7 +48,7 @@ def translate_prompt(prompt, translation_tokenizer, translation_model, device):
     return en_trans[0]
 
 
-class MultilingualStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
+class MultilingualStableDiffusion(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion in different languages.
 
@@ -135,6 +135,33 @@ class MultilingualStableDiffusion(DiffusionPipeline, StableDiffusionMixin):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/pipeline_animatediff_controlnet.py

@@ -28,7 +28,7 @@ from diffusers.models import AutoencoderKL, ControlNetModel, ImageProjection, UN
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.models.unets.unet_motion_model import MotionAdapter
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers import (
     DDIMScheduler,
     DPMSolverMultistepScheduler,
@@ -111,9 +111,7 @@ class AnimateDiffControlNetPipelineOutput(BaseOutput):
     frames: Union[torch.Tensor, np.ndarray]
 
 
-class AnimateDiffControlNetPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin
-):
+class AnimateDiffControlNetPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-to-video generation.
 
@@ -443,6 +441,67 @@ class AnimateDiffControlNetPipeline(
         video = video.float()
         return video
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
@@ -799,10 +858,8 @@ class AnimateDiffControlNetPipeline(
             ip_adapter_image (`PipelineImageInput`, *optional*):
                 Optional image input to work with IP Adapters.
             ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
-                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
-                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
-                if `do_classifier_free_guidance` is set to `True`.
-                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+                Pre-generated image embeddings for IP-Adapter. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
             conditioning_frames (`List[PipelineImageInput]`, *optional*):
                 The ControlNet input condition to provide guidance to the `unet` for generation. If multiple ControlNets
                 are specified, images must be passed as a list such that each element of the list can be correctly

examples/community/pipeline_animatediff_img2video.py

@@ -30,9 +30,9 @@ from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.models.unet_motion_model import MotionAdapter
+from diffusers.models.unets.unet_motion_model import MotionAdapter
 from diffusers.pipelines.animatediff.pipeline_output import AnimateDiffPipelineOutput
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers import (
     DDIMScheduler,
     DPMSolverMultistepScheduler,
@@ -232,9 +232,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class AnimateDiffImgToVideoPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin
-):
+class AnimateDiffImgToVideoPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin):
     r"""
     Pipeline for image-to-video generation.
 
@@ -566,6 +564,67 @@ class AnimateDiffImgToVideoPipeline(
         video = video.float()
         return video
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
         # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
@@ -798,10 +857,8 @@ class AnimateDiffImgToVideoPipeline(
             ip_adapter_image: (`PipelineImageInput`, *optional*):
                 Optional image input to work with IP Adapters.
             ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
-                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
-                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
-                if `do_classifier_free_guidance` is set to `True`.
-                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+                Pre-generated image embeddings for IP-Adapter. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
                 `np.array`.

examples/community/pipeline_demofusion_sdxl.py

@@ -23,7 +23,7 @@ from diffusers.models.attention_processor import (
     XFormersAttnProcessor,
 )
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     is_accelerate_available,
@@ -93,9 +93,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
-class DemoFusionSDXLPipeline(
-    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-):
+class DemoFusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion XL.
 
@@ -178,6 +176,39 @@ class DemoFusionSDXLPipeline(
         else:
             self.watermark = None
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def encode_prompt(
         self,
         prompt: str,

examples/community/pipeline_prompt2prompt.py

@@ -15,46 +15,17 @@
 from __future__ import annotations
 
 import abc
-import inspect
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 import torch.nn.functional as F
-from packaging import version
-from transformers import (
-    CLIPImageProcessor,
-    CLIPTextModel,
-    CLIPTokenizer,
-    CLIPVisionModelWithProjection,
-)
 
-from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
-from diffusers.configuration_utils import FrozenDict, deprecate
-from diffusers.image_processor import VaeImageProcessor
-from diffusers.loaders import (
-    FromSingleFileMixin,
-    IPAdapterMixin,
-    LoraLoaderMixin,
-    TextualInversionLoaderMixin,
-)
 from diffusers.models.attention import Attention
-from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.pipelines.stable_diffusion.safety_checker import (
-    StableDiffusionSafetyChecker,
+from diffusers.pipelines.stable_diffusion import (
+    StableDiffusionPipeline,
+    StableDiffusionPipelineOutput,
 )
-from diffusers.schedulers import KarrasDiffusionSchedulers
-from diffusers.utils import (
-    USE_PEFT_BACKEND,
-    logging,
-    scale_lora_layers,
-    unscale_lora_layers,
-)
-from diffusers.utils.torch_utils import randn_tensor
-
-
-logger = logging.get_logger(__name__)
 
 
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
@@ -72,486 +43,34 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
-class Prompt2PromptPipeline(
-    DiffusionPipeline,
-    TextualInversionLoaderMixin,
-    LoraLoaderMixin,
-    IPAdapterMixin,
-    FromSingleFileMixin,
-):
+class Prompt2PromptPipeline(StableDiffusionPipeline):
     r"""
-    Pipeline for text-to-image generation using Stable Diffusion.
-
-    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
-    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
-
-    The pipeline also inherits the following loading methods:
-        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
-        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
-        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
-        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
-        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
-
     Args:
+    Prompt-to-Prompt-Pipeline for text-to-image generation using Stable Diffusion. This model inherits from
+    [`StableDiffusionPipeline`]. Check the superclass documentation for the generic methods the library implements for
+    all the pipelines (such as downloading or saving, running on a particular device, etc.)
         vae ([`AutoencoderKL`]):
-            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
-        text_encoder ([`~transformers.CLIPTextModel`]):
-            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
-        tokenizer ([`~transformers.CLIPTokenizer`]):
-            A `CLIPTokenizer` to tokenize text.
-        unet ([`UNet2DConditionModel`]):
-            A `UNet2DConditionModel` to denoise the encoded image latents.
-        scheduler ([`SchedulerMixin`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. scheduler
+        ([`SchedulerMixin`]):
             A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
             [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
         safety_checker ([`StableDiffusionSafetyChecker`]):
             Classification module that estimates whether generated images could be considered offensive or harmful.
-            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
-            about a model's potential harms.
-        feature_extractor ([`~transformers.CLIPImageProcessor`]):
-            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
 
-    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
-    _exclude_from_cpu_offload = ["safety_checker"]
-    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
     _optional_components = ["safety_checker", "feature_extractor"]
 
-    def __init__(
-        self,
-        vae: AutoencoderKL,
-        text_encoder: CLIPTextModel,
-        tokenizer: CLIPTokenizer,
-        unet: UNet2DConditionModel,
-        scheduler: KarrasDiffusionSchedulers,
-        safety_checker: StableDiffusionSafetyChecker,
-        feature_extractor: CLIPImageProcessor,
-        image_encoder: CLIPVisionModelWithProjection = None,
-        requires_safety_checker: bool = True,
-    ):
-        super().__init__()
-
-        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
-            deprecation_message = (
-                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
-                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
-                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
-                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
-                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
-                " file"
-            )
-            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
-            new_config = dict(scheduler.config)
-            new_config["steps_offset"] = 1
-            scheduler._internal_dict = FrozenDict(new_config)
-
-        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
-            deprecation_message = (
-                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
-                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
-                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
-                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
-                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
-            )
-            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
-            new_config = dict(scheduler.config)
-            new_config["clip_sample"] = False
-            scheduler._internal_dict = FrozenDict(new_config)
-
-        if safety_checker is None and requires_safety_checker:
-            logger.warning(
-                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
-                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
-                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
-                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
-                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
-                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
-            )
-
-        if safety_checker is not None and feature_extractor is None:
-            raise ValueError(
-                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
-                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
-            )
-
-        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
-            version.parse(unet.config._diffusers_version).base_version
-        ) < version.parse("0.9.0.dev0")
-        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
-        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
-            deprecation_message = (
-                "The configuration file of the unet has set the default `sample_size` to smaller than"
-                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
-                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
-                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
-                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
-                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
-                " in the config might lead to incorrect results in future versions. If you have downloaded this"
-                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
-                " the `unet/config.json` file"
-            )
-            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
-            new_config = dict(unet.config)
-            new_config["sample_size"] = 64
-            unet._internal_dict = FrozenDict(new_config)
-
-        self.register_modules(
-            vae=vae,
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            unet=unet,
-            scheduler=scheduler,
-            safety_checker=safety_checker,
-            feature_extractor=feature_extractor,
-            image_encoder=image_encoder,
-        )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
-        self.register_to_config(requires_safety_checker=requires_safety_checker)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
-    def _encode_prompt(
-        self,
-        prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
-        negative_prompt=None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        lora_scale: Optional[float] = None,
-        **kwargs,
-    ):
-        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
-        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
-
-        prompt_embeds_tuple = self.encode_prompt(
-            prompt=prompt,
-            device=device,
-            num_images_per_prompt=num_images_per_prompt,
-            do_classifier_free_guidance=do_classifier_free_guidance,
-            negative_prompt=negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=lora_scale,
-            **kwargs,
-        )
-
-        # concatenate for backwards comp
-        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
-
-        return prompt_embeds
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
-    def encode_prompt(
-        self,
-        prompt,
-        device,
-        num_images_per_prompt,
-        do_classifier_free_guidance,
-        negative_prompt=None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        lora_scale: Optional[float] = None,
-        clip_skip: Optional[int] = None,
-    ):
-        r"""
-        Encodes the prompt into text encoder hidden states.
-
-        Args:
-            prompt (`str` or `List[str]`, *optional*):
-                prompt to be encoded
-            device: (`torch.device`):
-                torch device
-            num_images_per_prompt (`int`):
-                number of images that should be generated per prompt
-            do_classifier_free_guidance (`bool`):
-                whether to use classifier free guidance or not
-            negative_prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts not to guide the image generation. If not defined, one has to pass
-                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
-                less than `1`).
-            prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
-                provided, text embeddings will be generated from `prompt` input argument.
-            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
-                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
-                argument.
-            lora_scale (`float`, *optional*):
-                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
-            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.
-        """
-        # set lora scale so that monkey patched LoRA
-        # function of text encoder can correctly access it
-        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
-            self._lora_scale = lora_scale
-
-            # dynamically adjust the LoRA scale
-            if not USE_PEFT_BACKEND:
-                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
-            else:
-                scale_lora_layers(self.text_encoder, lora_scale)
-
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        if prompt_embeds is None:
-            # textual inversion: process multi-vector tokens if necessary
-            if isinstance(self, TextualInversionLoaderMixin):
-                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
-
-            text_inputs = self.tokenizer(
-                prompt,
-                padding="max_length",
-                max_length=self.tokenizer.model_max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-            text_input_ids = text_inputs.input_ids
-            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
-
-            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
-                text_input_ids, untruncated_ids
-            ):
-                removed_text = self.tokenizer.batch_decode(
-                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
-                )
-                logger.warning(
-                    "The following part of your input was truncated because CLIP can only handle sequences up to"
-                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
-                )
-
-            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
-                attention_mask = text_inputs.attention_mask.to(device)
-            else:
-                attention_mask = None
-
-            if clip_skip is None:
-                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
-                prompt_embeds = prompt_embeds[0]
-            else:
-                prompt_embeds = self.text_encoder(
-                    text_input_ids.to(device),
-                    attention_mask=attention_mask,
-                    output_hidden_states=True,
-                )
-                # Access the `hidden_states` first, that contains a tuple of
-                # all the hidden states from the encoder layers. Then index into
-                # the tuple to access the hidden states from the desired layer.
-                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
-                # We also need to apply the final LayerNorm here to not mess with the
-                # representations. The `last_hidden_states` that we typically use for
-                # obtaining the final prompt representations passes through the LayerNorm
-                # layer.
-                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
-
-        if self.text_encoder is not None:
-            prompt_embeds_dtype = self.text_encoder.dtype
-        elif self.unet is not None:
-            prompt_embeds_dtype = self.unet.dtype
-        else:
-            prompt_embeds_dtype = prompt_embeds.dtype
-
-        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
-        bs_embed, seq_len, _ = prompt_embeds.shape
-        # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
-        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
-
-        # get unconditional embeddings for classifier free guidance
-        if do_classifier_free_guidance and negative_prompt_embeds is None:
-            uncond_tokens: List[str]
-            if negative_prompt is None:
-                uncond_tokens = [""] * batch_size
-            elif prompt is not None and type(prompt) is not type(negative_prompt):
-                raise TypeError(
-                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
-                    f" {type(prompt)}."
-                )
-            elif isinstance(negative_prompt, str):
-                uncond_tokens = [negative_prompt]
-            elif batch_size != len(negative_prompt):
-                raise ValueError(
-                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
-                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
-                    " the batch size of `prompt`."
-                )
-            else:
-                uncond_tokens = negative_prompt
-
-            # textual inversion: process multi-vector tokens if necessary
-            if isinstance(self, TextualInversionLoaderMixin):
-                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
-
-            max_length = prompt_embeds.shape[1]
-            uncond_input = self.tokenizer(
-                uncond_tokens,
-                padding="max_length",
-                max_length=max_length,
-                truncation=True,
-                return_tensors="pt",
-            )
-
-            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
-                attention_mask = uncond_input.attention_mask.to(device)
-            else:
-                attention_mask = None
-
-            negative_prompt_embeds = self.text_encoder(
-                uncond_input.input_ids.to(device),
-                attention_mask=attention_mask,
-            )
-            negative_prompt_embeds = negative_prompt_embeds[0]
-
-        if do_classifier_free_guidance:
-            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
-            seq_len = negative_prompt_embeds.shape[1]
-
-            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
-            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
-            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
-
-        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
-            # Retrieve the original scale by scaling back the LoRA layers
-            unscale_lora_layers(self.text_encoder, lora_scale)
-
-        return prompt_embeds, negative_prompt_embeds
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
-    def run_safety_checker(self, image, device, dtype):
-        if self.safety_checker is None:
-            has_nsfw_concept = None
-        else:
-            if torch.is_tensor(image):
-                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
-            else:
-                feature_extractor_input = self.image_processor.numpy_to_pil(image)
-            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
-            image, has_nsfw_concept = self.safety_checker(
-                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
-            )
-        return image, has_nsfw_concept
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
-    def prepare_extra_step_kwargs(self, generator, eta):
-        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
-        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
-        # and should be between [0, 1]
-
-        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        extra_step_kwargs = {}
-        if accepts_eta:
-            extra_step_kwargs["eta"] = eta
-
-        # check if the scheduler accepts generator
-        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        if accepts_generator:
-            extra_step_kwargs["generator"] = generator
-        return extra_step_kwargs
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
-    def check_inputs(
-        self,
-        prompt,
-        height,
-        width,
-        callback_steps,
-        negative_prompt=None,
-        prompt_embeds=None,
-        negative_prompt_embeds=None,
-        ip_adapter_image=None,
-        ip_adapter_image_embeds=None,
-        callback_on_step_end_tensor_inputs=None,
-    ):
-        if height % 8 != 0 or width % 8 != 0:
-            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
-
-        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
-            raise ValueError(
-                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
-                f" {type(callback_steps)}."
-            )
-        if callback_on_step_end_tensor_inputs is not None and not all(
-            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
-        ):
-            raise ValueError(
-                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
-            )
-
-        if prompt is not None and prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
-                " only forward one of the two."
-            )
-        elif prompt is None and prompt_embeds is None:
-            raise ValueError(
-                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
-            )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
-
-        if negative_prompt is not None and negative_prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
-                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
-            )
-
-        if prompt_embeds is not None and negative_prompt_embeds is not None:
-            if prompt_embeds.shape != negative_prompt_embeds.shape:
-                raise ValueError(
-                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
-                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
-                    f" {negative_prompt_embeds.shape}."
-                )
-
-        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
-            raise ValueError(
-                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
-            )
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
-    def prepare_latents(
-        self,
-        batch_size,
-        num_channels_latents,
-        height,
-        width,
-        dtype,
-        device,
-        generator,
-        latents=None,
-    ):
-        shape = (
-            batch_size,
-            num_channels_latents,
-            height // self.vae_scale_factor,
-            width // self.vae_scale_factor,
-        )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
-
-        if latents is None:
-            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
-        else:
-            latents = latents.to(device)
-
-        # scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * self.scheduler.init_noise_sigma
-        return latents
-
     @torch.no_grad()
     def __call__(
         self,

examples/community/pipeline_sdxl_style_aligned.py

@@ -51,7 +51,7 @@ from diffusers.models.attention_processor import (
     XFormersAttnProcessor,
 )
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
@@ -389,7 +389,6 @@ def retrieve_latents(
 
 class StyleAlignedSDXLPipeline(
     DiffusionPipeline,
-    StableDiffusionMixin,
     FromSingleFileMixin,
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
@@ -505,6 +504,39 @@ class StyleAlignedSDXLPipeline(
         else:
             self.watermark = None
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def encode_prompt(
         self,
         prompt: str,
@@ -1155,6 +1187,34 @@ class StyleAlignedSDXLPipeline(
             self.vae.decoder.conv_in.to(dtype)
             self.vae.decoder.mid_block.to(dtype)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     def _enable_shared_attention_processors(
         self,
         share_attention: bool,
@@ -1301,6 +1361,65 @@ class StyleAlignedSDXLPipeline(
             self._style_aligned_norm_layers = None
             self._disable_shared_attention_processors()
 
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py

@@ -33,7 +33,7 @@ from diffusers.models.attention_processor import (
 )
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
@@ -158,11 +158,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
 
 
 class StableDiffusionXLControlNetAdapterPipeline(
-    DiffusionPipeline,
-    StableDiffusionMixin,
-    FromSingleFileMixin,
-    StableDiffusionXLLoraLoaderMixin,
-    TextualInversionLoaderMixin,
+    DiffusionPipeline, FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
@@ -238,6 +234,39 @@ class StableDiffusionXLControlNetAdapterPipeline(
         )
         self.default_sample_size = self.unet.config.sample_size
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
         self,
@@ -834,6 +863,34 @@ class StableDiffusionXLControlNetAdapterPipeline(
 
         return height, width
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     def prepare_control_image(
         self,
         image,

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -52,7 +52,6 @@ from diffusers.models.attention_processor import (
 )
 from diffusers.models.lora import adjust_lora_scale_text_encoder
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
@@ -304,9 +303,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     return noise_cfg
 
 
-class StableDiffusionXLControlNetAdapterInpaintPipeline(
-    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, LoraLoaderMixin
-):
+class StableDiffusionXLControlNetAdapterInpaintPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
     https://arxiv.org/abs/2302.08453
@@ -386,6 +383,39 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
         )
         self.default_sample_size = self.unet.config.sample_size
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
         self,
@@ -1177,6 +1207,34 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(
 
         return height, width
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     def prepare_control_image(
         self,
         image,

examples/community/pipeline_stable_diffusion_xl_ipex.py

@@ -267,6 +267,39 @@ class StableDiffusionXLPipelineIpex(
         else:
             self.watermark = None
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     def encode_prompt(
         self,
         prompt: str,
@@ -668,6 +701,34 @@ class StableDiffusionXLPipelineIpex(
             self.vae.decoder.conv_in.to(dtype)
             self.vae.decoder.mid_block.to(dtype)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """

examples/community/pipeline_zero1to3.py

@@ -22,16 +22,18 @@ from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection
 #     randn_tensor,
 #     replace_example_docstring,
 # )
-# from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+# from ..pipeline_utils import DiffusionPipeline
 # from . import StableDiffusionPipelineOutput
 # from .safety_checker import StableDiffusionSafetyChecker
-from diffusers import AutoencoderKL, DiffusionPipeline, StableDiffusionMixin, UNet2DConditionModel
+from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
 from diffusers.configuration_utils import ConfigMixin, FrozenDict
 from diffusers.models.modeling_utils import ModelMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     deprecate,
+    is_accelerate_available,
+    is_accelerate_version,
     logging,
     replace_example_docstring,
 )
@@ -66,7 +68,7 @@ class CCProjection(ModelMixin, ConfigMixin):
         return self.projection(x)
 
 
-class Zero1to3StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class Zero1to3StableDiffusionPipeline(DiffusionPipeline):
     r"""
     Pipeline for single view conditioned novel view generation using Zero1to3.
 
@@ -185,6 +187,109 @@ class Zero1to3StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
         self.register_to_config(requires_safety_checker=requires_safety_checker)
         # self.model_mode = None
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/sd_text2img_k_diffusion.py

@@ -19,9 +19,9 @@ from typing import Callable, List, Optional, Union
 import torch
 from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser
 
-from diffusers import DiffusionPipeline, LMSDiscreteScheduler, StableDiffusionMixin
+from diffusers import DiffusionPipeline, LMSDiscreteScheduler
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import logging
+from diffusers.utils import is_accelerate_available, logging
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -41,7 +41,7 @@ class ModelWrapper:
         return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample
 
 
-class StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -120,6 +120,68 @@ class StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
         sampling = getattr(library, "sampling")
         self.sampler = getattr(sampling, scheduler_type)
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
         r"""
         Encodes the prompt into text encoder hidden states.

examples/community/seed_resize_stable_diffusion.py

@@ -9,7 +9,6 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from diffusers import DiffusionPipeline
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -19,7 +18,7 @@ from diffusers.utils import logging
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class SeedResizeStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class SeedResizeStableDiffusionPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -68,6 +67,33 @@ class SeedResizeStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin)
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/speech_to_image_diffusion.py

@@ -18,7 +18,6 @@ from diffusers import (
     PNDMScheduler,
     UNet2DConditionModel,
 )
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.utils import logging
@@ -27,7 +26,7 @@ from diffusers.utils import logging
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class SpeechToImagePipeline(DiffusionPipeline, StableDiffusionMixin):
+class SpeechToImagePipeline(DiffusionPipeline):
     def __init__(
         self,
         speech_model: WhisperForConditionalGeneration,
@@ -63,6 +62,14 @@ class SpeechToImagePipeline(DiffusionPipeline, StableDiffusionMixin):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        if slice_size == "auto":
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/stable_diffusion_comparison.py

@@ -12,7 +12,6 @@ from diffusers import (
     StableDiffusionPipeline,
     UNet2DConditionModel,
 )
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
@@ -23,7 +22,7 @@ pipe3_model_id = "CompVis/stable-diffusion-v1-3"
 pipe4_model_id = "CompVis/stable-diffusion-v1-4"
 
 
-class StableDiffusionComparisonPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionComparisonPipeline(DiffusionPipeline):
     r"""
     Pipeline for parallel comparison of Stable Diffusion v1-v4
     This pipeline inherits from DiffusionPipeline and depends on the use of an Auth Token for
@@ -84,6 +83,31 @@ class StableDiffusionComparisonPipeline(DiffusionPipeline, StableDiffusionMixin)
     def layers(self) -> Dict[str, Any]:
         return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def text2img_sd1_1(
         self,

examples/community/stable_diffusion_controlnet_img2img.py

@@ -8,13 +8,14 @@ import PIL.Image
 import torch
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
-from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
+from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
+    is_accelerate_available,
+    is_accelerate_version,
     replace_example_docstring,
 )
 from diffusers.utils.torch_utils import randn_tensor
@@ -129,7 +130,7 @@ def prepare_controlnet_conditioning_image(
     return controlnet_conditioning_image
 
 
-class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline):
     """
     Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
     """
@@ -182,6 +183,89 @@ class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline, StableDiffusio
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            # the safety checker can offload the vae again
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # control net hook has be manually offloaded as it alternates with unet
+        cpu_offload_with_hook(self.controlnet, device)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/stable_diffusion_controlnet_inpaint.py

@@ -9,13 +9,14 @@ import torch
 import torch.nn.functional as F
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
-from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
+from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
+    is_accelerate_available,
+    is_accelerate_version,
     replace_example_docstring,
 )
 from diffusers.utils.torch_utils import randn_tensor
@@ -227,7 +228,7 @@ def prepare_controlnet_conditioning_image(
     return controlnet_conditioning_image
 
 
-class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline):
     """
     Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
     """
@@ -281,6 +282,89 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, StableDiffusio
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            # the safety checker can offload the vae again
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # control net hook has be manually offloaded as it alternates with unet
+        cpu_offload_with_hook(self.controlnet, device)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/stable_diffusion_controlnet_inpaint_img2img.py

@@ -9,12 +9,13 @@ import torch
 import torch.nn.functional as F
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
-from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     PIL_INTERPOLATION,
+    is_accelerate_available,
+    is_accelerate_version,
     replace_example_docstring,
 )
 from diffusers.utils.torch_utils import randn_tensor
@@ -216,7 +217,7 @@ def prepare_controlnet_conditioning_image(
     return controlnet_conditioning_image
 
 
-class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline):
     """
     Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
     """
@@ -266,6 +267,89 @@ class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline, StableD
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            # the safety checker can offload the vae again
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # control net hook has be manually offloaded as it alternates with unet
+        cpu_offload_with_hook(self.controlnet, device)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/stable_diffusion_ipex.py

@@ -23,12 +23,14 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 from diffusers.configuration_utils import FrozenDict
 from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     deprecate,
+    is_accelerate_available,
+    is_accelerate_version,
     logging,
     replace_example_docstring,
 )
@@ -60,9 +62,7 @@ EXAMPLE_DOC_STRING = """
 """
 
 
-class StableDiffusionIPEXPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
-):
+class StableDiffusionIPEXPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion on IPEX.
 
@@ -304,6 +304,109 @@ class StableDiffusionIPEXPipeline(
             ave_decoder_trace_model = torch.jit.freeze(ave_decoder_trace_model)
         self.vae.decoder.forward = ave_decoder_trace_model.forward
 
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
+        several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     def _encode_prompt(
         self,
         prompt,

examples/community/stable_diffusion_mega.py

@@ -16,7 +16,6 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.configuration_utils import FrozenDict
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.utils import deprecate, logging
 
@@ -24,7 +23,7 @@ from diffusers.utils import deprecate, logging
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class StableDiffusionMegaPipeline(DiffusionPipeline, StableDiffusionMixin):
+class StableDiffusionMegaPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 
@@ -95,6 +94,33 @@ class StableDiffusionMegaPipeline(DiffusionPipeline, StableDiffusionMixin):
     def components(self) -> Dict[str, Any]:
         return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
     @torch.no_grad()
     def inpaint(
         self,

examples/community/stable_diffusion_repaint.py

@@ -24,13 +24,14 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
 from diffusers.configuration_utils import FrozenDict, deprecate
 from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import (
     StableDiffusionSafetyChecker,
 )
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
+    is_accelerate_available,
+    is_accelerate_version,
     logging,
 )
 from diffusers.utils.torch_utils import randn_tensor
@@ -139,9 +140,7 @@ def prepare_mask_and_masked_image(image, mask):
     return mask, masked_image
 
 
-class StableDiffusionRepaintPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
-):
+class StableDiffusionRepaintPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*.
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
@@ -277,6 +276,80 @@ class StableDiffusionRepaintPipeline(
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        Note that offloading happens on a submodule basis. Memory savings are higher than with
+        `enable_model_cpu_offload`, but performance is lower.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            cpu_offload(cpu_offloaded_model, device)
+
+        if self.safety_checker is not None:
+            cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
+    def enable_model_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        """
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate import cpu_offload_with_hook
+        else:
+            raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        if self.device.type != "cpu":
+            self.to("cpu", silence_dtype_warnings=True)
+            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+
+        hook = None
+        for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+            _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+        if self.safety_checker is not None:
+            _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+        # We'll offload the last model manually.
+        self.final_offload_hook = hook
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,

examples/community/text_inpainting.py

@@ -13,17 +13,16 @@ from transformers import (
 from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
-from diffusers.utils import deprecate, logging
+from diffusers.utils import deprecate, is_accelerate_available, logging
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class TextInpainting(DiffusionPipeline, StableDiffusionMixin):
+class TextInpainting(DiffusionPipeline):
     r"""
     Pipeline for text based inpainting using Stable Diffusion.
     Uses CLIPSeg to get a mask from the given text, then calls the Inpainting pipeline with the generated mask
@@ -121,6 +120,69 @@ class TextInpainting(DiffusionPipeline, StableDiffusionMixin):
             feature_extractor=feature_extractor,
         )
 
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = self.unet.config.attention_head_dim // 2
+        self.unet.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device("cuda")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/unclip_image_interpolation.py

@@ -19,7 +19,7 @@ from diffusers import (
     UNet2DModel,
 )
 from diffusers.pipelines.unclip import UnCLIPTextProjModel
-from diffusers.utils import logging
+from diffusers.utils import is_accelerate_available, logging
 from diffusers.utils.torch_utils import randn_tensor
 
 
@@ -204,6 +204,50 @@ class UnCLIPImageInterpolationPipeline(DiffusionPipeline):
 
         return image_embeddings
 
+    # Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline.enable_sequential_cpu_offload
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
+        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
+        when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        models = [
+            self.decoder,
+            self.text_proj,
+            self.text_encoder,
+            self.super_res_first,
+            self.super_res_last,
+        ]
+        for cpu_offloaded_model in models:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.decoder, "_hf_hook"):
+            return self.device
+        for module in self.decoder.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/unclip_text_interpolation.py

@@ -15,7 +15,7 @@ from diffusers import (
     UNet2DModel,
 )
 from diffusers.pipelines.unclip import UnCLIPTextProjModel
-from diffusers.utils import logging
+from diffusers.utils import is_accelerate_available, logging
 from diffusers.utils.torch_utils import randn_tensor
 
 
@@ -212,6 +212,51 @@ class UnCLIPTextInterpolationPipeline(DiffusionPipeline):
 
         return prompt_embeds, text_encoder_hidden_states, text_mask
 
+    # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.enable_sequential_cpu_offload
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
+        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
+        when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        # TODO: self.prior.post_process_latents is not covered by the offload hooks, so it fails if added to the list
+        models = [
+            self.decoder,
+            self.text_proj,
+            self.text_encoder,
+            self.super_res_first,
+            self.super_res_last,
+        ]
+        for cpu_offloaded_model in models:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.decoder, "_hf_hook"):
+            return self.device
+        for module in self.decoder.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
     @torch.no_grad()
     def __call__(
         self,

examples/community/wildcard_stable_diffusion.py

@@ -8,9 +8,9 @@ from typing import Callable, Dict, List, Optional, Union
 import torch
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
+from diffusers import DiffusionPipeline
 from diffusers.configuration_utils import FrozenDict
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
@@ -63,7 +63,7 @@ class WildcardStableDiffusionOutput(StableDiffusionPipelineOutput):
     prompts: List[str]
 
 
-class WildcardStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+class WildcardStableDiffusionPipeline(DiffusionPipeline):
     r"""
     Example Usage:
         pipe = WildcardStableDiffusionPipeline.from_pretrained(

examples/controlnet/README_sdxl.md

@@ -113,7 +113,7 @@ pipe.enable_xformers_memory_efficient_attention()
 # memory optimization.
 pipe.enable_model_cpu_offload()
 
-control_image = load_image("./conditioning_image_1.png").resize((1024, 1024))
+control_image = load_image("./conditioning_image_1.png")
 prompt = "pale golden rod circle with old lace background"
 
 # generate image
@@ -128,14 +128,4 @@ image.save("./output.png")
 
 ### Specifying a better VAE
 
-SDXL's VAE is known to suffer from numerical instability issues. This is why we also expose a CLI argument namely `--pretrained_vae_model_name_or_path` that lets you specify the location of an alternative VAE (such as [`madebyollin/sdxl-vae-fp16-fix`](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).
-
-If you're using this VAE during training, you need to ensure you're using it during inference too. You do so by:
-
-```diff
-+ vae = AutoencoderKL.from_pretrained(vae_path_or_repo_id, torch_dtype=torch.float16)
-controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
-pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
-    base_model_path, controlnet=controlnet, torch_dtype=torch.float16,
-+   vae=vae,
-)
+SDXL's VAE is known to suffer from numerical instability issues. This is why we also expose a CLI argument namely `--pretrained_vae_model_name_or_path` that lets you specify the location of a better VAE (such as [this one](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).

examples/controlnet/train_controlnet.py

@@ -14,8 +14,6 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
-import gc
 import logging
 import math
 import os
@@ -76,15 +74,10 @@ def image_grid(imgs, rows, cols):
     return grid
 
 
-def log_validation(
-    vae, text_encoder, tokenizer, unet, controlnet, args, accelerator, weight_dtype, step, is_final_validation=False
-):
+def log_validation(vae, text_encoder, tokenizer, unet, controlnet, args, accelerator, weight_dtype, step):
     logger.info("Running validation... ")
 
-    if not is_final_validation:
-        controlnet = accelerator.unwrap_model(controlnet)
-    else:
-        controlnet = ControlNetModel.from_pretrained(args.output_dir, torch_dtype=weight_dtype)
+    controlnet = accelerator.unwrap_model(controlnet)
 
     pipeline = StableDiffusionControlNetPipeline.from_pretrained(
         args.pretrained_model_name_or_path,
@@ -125,7 +118,6 @@ def log_validation(
         )
 
     image_logs = []
-    inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast("cuda")
 
     for validation_prompt, validation_image in zip(validation_prompts, validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
@@ -133,7 +125,7 @@ def log_validation(
         images = []
 
         for _ in range(args.num_validation_images):
-            with inference_ctx:
+            with torch.autocast("cuda"):
                 image = pipeline(
                     validation_prompt, validation_image, num_inference_steps=20, generator=generator
                 ).images[0]
@@ -144,7 +136,6 @@ def log_validation(
             {"validation_image": validation_image, "images": images, "validation_prompt": validation_prompt}
         )
 
-    tracker_key = "test" if is_final_validation else "validation"
     for tracker in accelerator.trackers:
         if tracker.name == "tensorboard":
             for log in image_logs:
@@ -176,14 +167,10 @@ def log_validation(
                     image = wandb.Image(image, caption=validation_prompt)
                     formatted_images.append(image)
 
-            tracker.log({tracker_key: formatted_images})
+            tracker.log({"validation": formatted_images})
         else:
             logger.warn(f"image logging not implemented for {tracker.name}")
 
-        del pipeline
-        gc.collect()
-        torch.cuda.empty_cache()
-
         return image_logs
 
 
@@ -210,7 +197,7 @@ def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: st
 def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):
     img_str = ""
     if image_logs is not None:
-        img_str = "You can find some example images below.\n\n"
+        img_str = "You can find some example images below.\n"
         for i, log in enumerate(image_logs):
             images = log["images"]
             validation_prompt = log["validation_prompt"]
@@ -242,7 +229,6 @@ These are controlnet weights trained on {base_model} with new type of conditioni
         "text-to-image",
         "diffusers",
         "controlnet",
-        "diffusers-training",
     ]
     model_card = populate_model_card(model_card, tags=tags)
 
@@ -1145,22 +1131,6 @@ def main(args):
         controlnet = unwrap_model(controlnet)
         controlnet.save_pretrained(args.output_dir)
 
-        # Run a final round of validation.
-        image_logs = None
-        if args.validation_prompt is not None:
-            image_logs = log_validation(
-                vae=vae,
-                text_encoder=text_encoder,
-                tokenizer=tokenizer,
-                unet=unet,
-                controlnet=None,
-                args=args,
-                accelerator=accelerator,
-                weight_dtype=weight_dtype,
-                step=global_step,
-                is_final_validation=True,
-            )
-
         if args.push_to_hub:
             save_model_card(
                 repo_id,

examples/controlnet/train_controlnet_flax.py

@@ -169,7 +169,6 @@ These are controlnet weights trained on {base_model} with new type of conditioni
         "diffusers",
         "controlnet",
         "jax-diffusers-event",
-        "diffusers-training",
     ]
     model_card = populate_model_card(model_card, tags=tags)
 

examples/controlnet/train_controlnet_sdxl.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
 import functools
 import gc
 import logging
@@ -66,38 +65,20 @@ check_min_version("0.27.0.dev0")
 logger = get_logger(__name__)
 
 
-def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step, is_final_validation=False):
+def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step):
     logger.info("Running validation... ")
 
-    if not is_final_validation:
-        controlnet = accelerator.unwrap_model(controlnet)
-        pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
-            args.pretrained_model_name_or_path,
-            vae=vae,
-            unet=unet,
-            controlnet=controlnet,
-            revision=args.revision,
-            variant=args.variant,
-            torch_dtype=weight_dtype,
-        )
-    else:
-        controlnet = ControlNetModel.from_pretrained(args.output_dir, torch_dtype=weight_dtype)
-        if args.pretrained_vae_model_name_or_path is not None:
-            vae = AutoencoderKL.from_pretrained(args.pretrained_vae_model_name_or_path, torch_dtype=weight_dtype)
-        else:
-            vae = AutoencoderKL.from_pretrained(
-                args.pretrained_model_name_or_path, subfolder="vae", torch_dtype=weight_dtype
-            )
-
-        pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
-            args.pretrained_model_name_or_path,
-            vae=vae,
-            controlnet=controlnet,
-            revision=args.revision,
-            variant=args.variant,
-            torch_dtype=weight_dtype,
-        )
+    controlnet = accelerator.unwrap_model(controlnet)
 
+    pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        vae=vae,
+        unet=unet,
+        controlnet=controlnet,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
     pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)
     pipeline = pipeline.to(accelerator.device)
     pipeline.set_progress_bar_config(disable=True)
@@ -125,7 +106,6 @@ 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")
 
     for validation_prompt, validation_image in zip(validation_prompts, validation_images):
         validation_image = Image.open(validation_image).convert("RGB")
@@ -134,7 +114,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
         images = []
 
         for _ in range(args.num_validation_images):
-            with inference_ctx:
+            with torch.autocast("cuda"):
                 image = pipeline(
                     prompt=validation_prompt, image=validation_image, num_inference_steps=20, generator=generator
                 ).images[0]
@@ -144,7 +124,6 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
             {"validation_image": validation_image, "images": images, "validation_prompt": validation_prompt}
         )
 
-    tracker_key = "test" if is_final_validation else "validation"
     for tracker in accelerator.trackers:
         if tracker.name == "tensorboard":
             for log in image_logs:
@@ -176,7 +155,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
                     image = wandb.Image(image, caption=validation_prompt)
                     formatted_images.append(image)
 
-            tracker.log({tracker_key: formatted_images})
+            tracker.log({"validation": formatted_images})
         else:
             logger.warn(f"image logging not implemented for {tracker.name}")
 
@@ -210,7 +189,7 @@ def import_model_class_from_model_name_or_path(
 def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):
     img_str = ""
     if image_logs is not None:
-        img_str = "You can find some example images below.\n\n"
+        img_str = "You can find some example images below.\n"
         for i, log in enumerate(image_logs):
             images = log["images"]
             validation_prompt = log["validation_prompt"]
@@ -243,7 +222,6 @@ These are controlnet weights trained on {base_model} with new type of conditioni
         "text-to-image",
         "diffusers",
         "controlnet",
-        "diffusers-training",
     ]
     model_card = populate_model_card(model_card, tags=tags)
 
@@ -1250,13 +1228,7 @@ def main(args):
 
                     if args.validation_prompt is not None and global_step % args.validation_steps == 0:
                         image_logs = log_validation(
-                            vae=vae,
-                            unet=unet,
-                            controlnet=controlnet,
-                            args=args,
-                            accelerator=accelerator,
-                            weight_dtype=weight_dtype,
-                            step=global_step,
+                            vae, unet, controlnet, args, accelerator, weight_dtype, global_step
                         )
 
             logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
@@ -1272,21 +1244,6 @@ def main(args):
         controlnet = unwrap_model(controlnet)
         controlnet.save_pretrained(args.output_dir)
 
-        # Run a final round of validation.
-        # Setting `vae`, `unet`, and `controlnet` to None to load automatically from `args.output_dir`.
-        image_logs = None
-        if args.validation_prompt is not None:
-            image_logs = log_validation(
-                vae=None,
-                unet=None,
-                controlnet=None,
-                args=args,
-                accelerator=accelerator,
-                weight_dtype=weight_dtype,
-                step=global_step,
-                is_final_validation=True,
-            )
-
         if args.push_to_hub:
             save_model_card(
                 repo_id,

examples/custom_diffusion/train_custom_diffusion.py

@@ -97,14 +97,7 @@ These are Custom Diffusion adaption weights for {base_model}. The weights were t
         inference=True,
     )
 
-    tags = [
-        "text-to-image",
-        "diffusers",
-        "stable-diffusion",
-        "stable-diffusion-diffusers",
-        "custom-diffusion",
-        "diffusers-training",
-    ]
+    tags = ["text-to-image", "diffusers", "stable-diffusion", "stable-diffusion-diffusers", "custom-diffusion"]
     model_card = populate_model_card(model_card, tags=tags)
 
     model_card.save(os.path.join(repo_folder, "README.md"))

examples/dreambooth/README_sdxl.md

@@ -206,66 +206,3 @@ You can explore the results from a couple of our internal experiments by checkin
 ## Running on a free-tier Colab Notebook
 
 Check out [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_DreamBooth_LoRA_.ipynb). 
-
-## Conducting EDM-style training
-
-It's now possible to perform EDM-style training as proposed in [Elucidating the Design Space of Diffusion-Based Generative Models](https://arxiv.org/abs/2206.00364). 
-
-For the SDXL model, simple set:
-
-```diff
-+  --do_edm_style_training \
-```
-
-Other SDXL-like models that use the EDM formulation, such as [playgroundai/playground-v2.5-1024px-aesthetic](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic), can also be DreamBooth'd with the script. Below is an example command:
-
-```bash
-accelerate launch train_dreambooth_lora_sdxl.py \
-  --pretrained_model_name_or_path="playgroundai/playground-v2.5-1024px-aesthetic"  \
-  --instance_data_dir="dog" \
-  --output_dir="dog-playground-lora" \
-  --mixed_precision="fp16" \
-  --instance_prompt="a photo of sks dog" \
-  --resolution=1024 \
-  --train_batch_size=1 \
-  --gradient_accumulation_steps=4 \
-  --learning_rate=1e-4 \
-  --use_8bit_adam \
-  --report_to="wandb" \
-  --lr_scheduler="constant" \
-  --lr_warmup_steps=0 \
-  --max_train_steps=500 \
-  --validation_prompt="A photo of sks dog in a bucket" \
-  --validation_epochs=25 \
-  --seed="0" \
-  --push_to_hub
-```
-
-> [!CAUTION]
-> Min-SNR gamma is not supported with the EDM-style training yet. When training with the PlaygroundAI model, it's recommended to not pass any "variant".
-
-### DoRA training 
-The script now supports DoRA training too!
-> Proposed in [DoRA: Weight-Decomposed Low-Rank Adaptation](https://arxiv.org/abs/2402.09353), 
-**DoRA** is very similar to LoRA, except it decomposes the pre-trained weight into two components, **magnitude** and **direction** and employs LoRA for _directional_ updates to efficiently minimize the number of trainable parameters. 
-The authors found that by using DoRA, both the learning capacity and training stability of LoRA are enhanced without any additional overhead during inference. 
-
-> [!NOTE]
-> 💡DoRA training is still _experimental_  
-> and is likely to require different hyperparameter values to perform best compared to a LoRA.
-> Specifically, we've noticed 2 differences to take into account your training: 
-> 1. **LoRA seem to converge faster than DoRA** (so a set of parameters that may lead to overfitting when training a LoRA may be working well for a DoRA)
-> 2. **DoRA quality superior to LoRA especially in lower ranks** the difference in quality of DoRA of rank 8 and LoRA of rank 8 appears to be more significant than when training ranks of 32 or 64 for example.  
-> 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: 
-```bash
-pip install git+https://github.com/huggingface/peft.git
-```
-2. Enable DoRA training by adding this flag
-```bash
---use_dora
-```
-**Inference** 
-The inference is the same as if you train a regular LoRA 🤗

examples/dreambooth/test_dreambooth_lora_edm.py

@@ -1,99 +0,0 @@
-# coding=utf-8
-# Copyright 2024 HuggingFace Inc.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-import os
-import sys
-import tempfile
-
-import safetensors
-
-
-sys.path.append("..")
-from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
-
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-stream_handler = logging.StreamHandler(sys.stdout)
-logger.addHandler(stream_handler)
-
-
-class DreamBoothLoRASDXLWithEDM(ExamplesTestsAccelerate):
-    def test_dreambooth_lora_sdxl_with_edm(self):
-        with tempfile.TemporaryDirectory() as tmpdir:
-            test_args = f"""
-                examples/dreambooth/train_dreambooth_lora_sdxl.py
-                --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-xl-pipe
-                --do_edm_style_training
-                --instance_data_dir docs/source/en/imgs
-                --instance_prompt photo
-                --resolution 64
-                --train_batch_size 1
-                --gradient_accumulation_steps 1
-                --max_train_steps 2
-                --learning_rate 5.0e-04
-                --scale_lr
-                --lr_scheduler constant
-                --lr_warmup_steps 0
-                --output_dir {tmpdir}
-                """.split()
-
-            run_command(self._launch_args + test_args)
-            # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
-
-            # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
-            is_lora = all("lora" in k for k in lora_state_dict.keys())
-            self.assertTrue(is_lora)
-
-            # when not training the text encoder, all the parameters in the state dict should start
-            # with `"unet"` in their names.
-            starts_with_unet = all(key.startswith("unet") for key in lora_state_dict.keys())
-            self.assertTrue(starts_with_unet)
-
-    def test_dreambooth_lora_playground(self):
-        with tempfile.TemporaryDirectory() as tmpdir:
-            test_args = f"""
-                examples/dreambooth/train_dreambooth_lora_sdxl.py
-                --pretrained_model_name_or_path hf-internal-testing/tiny-playground-v2-5-pipe
-                --instance_data_dir docs/source/en/imgs
-                --instance_prompt photo
-                --resolution 64
-                --train_batch_size 1
-                --gradient_accumulation_steps 1
-                --max_train_steps 2
-                --learning_rate 5.0e-04
-                --scale_lr
-                --lr_scheduler constant
-                --lr_warmup_steps 0
-                --output_dir {tmpdir}
-                """.split()
-
-            run_command(self._launch_args + test_args)
-            # save_pretrained smoke test
-            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
-
-            # make sure the state_dict has the correct naming in the parameters.
-            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
-            is_lora = all("lora" in k for k in lora_state_dict.keys())
-            self.assertTrue(is_lora)
-
-            # when not training the text encoder, all the parameters in the state dict should start
-            # with `"unet"` in their names.
-            starts_with_unet = all(key.startswith("unet") for key in lora_state_dict.keys())
-            self.assertTrue(starts_with_unet)

examples/dreambooth/train_dreambooth.py

@@ -102,7 +102,7 @@ DreamBooth for the text encoder was enabled: {train_text_encoder}.
         inference=True,
     )
 
-    tags = ["text-to-image", "dreambooth", "diffusers-training"]
+    tags = ["text-to-image", "dreambooth"]
     if isinstance(pipeline, StableDiffusionPipeline):
         tags.extend(["stable-diffusion", "stable-diffusion-diffusers"])
     else:

examples/dreambooth/train_dreambooth_lora.py

@@ -106,7 +106,7 @@ LoRA for the text encoder was enabled: {train_text_encoder}.
         model_description=model_description,
         inference=True,
     )
-    tags = ["text-to-image", "diffusers", "lora", "diffusers-training"]
+    tags = ["text-to-image", "diffusers", "lora"]
     if isinstance(pipeline, StableDiffusionPipeline):
         tags.extend(["stable-diffusion", "stable-diffusion-diffusers"])
     else:

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -14,10 +14,8 @@
 # See the License for the specific language governing permissions and
 
 import argparse
-import contextlib
 import gc
 import itertools
-import json
 import logging
 import math
 import os
@@ -34,7 +32,7 @@ import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
 from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
-from huggingface_hub import create_repo, hf_hub_download, upload_folder
+from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
 from peft import LoraConfig, set_peft_model_state_dict
@@ -52,8 +50,6 @@ from diffusers import (
     AutoencoderKL,
     DDPMScheduler,
     DPMSolverMultistepScheduler,
-    EDMEulerScheduler,
-    EulerDiscreteScheduler,
     StableDiffusionXLPipeline,
     UNet2DConditionModel,
 )
@@ -80,23 +76,8 @@ check_min_version("0.27.0.dev0")
 logger = get_logger(__name__)
 
 
-def determine_scheduler_type(pretrained_model_name_or_path, revision):
-    model_index_filename = "model_index.json"
-    if os.path.isdir(pretrained_model_name_or_path):
-        model_index = os.path.join(pretrained_model_name_or_path, model_index_filename)
-    else:
-        model_index = hf_hub_download(
-            repo_id=pretrained_model_name_or_path, filename=model_index_filename, revision=revision
-        )
-
-    with open(model_index, "r") as f:
-        scheduler_type = json.load(f)["scheduler"][1]
-    return scheduler_type
-
-
 def save_model_card(
     repo_id: str,
-    use_dora: bool,
     images=None,
     base_model: str = None,
     train_text_encoder=False,
@@ -114,7 +95,7 @@ def save_model_card(
             )
 
     model_description = f"""
-# {'SDXL' if 'playgroundai' not in base_model else 'Playground'} LoRA DreamBooth - {repo_id}
+# SDXL LoRA DreamBooth - {repo_id}
 
 <Gallery />
 
@@ -138,17 +119,11 @@ Weights for this model are available in Safetensors format.
 
 [Download]({repo_id}/tree/main) them in the Files & versions tab.
 
-"""
-    if "playgroundai" in args.pretrained_model_name_or_path:
-        model_description += """\n
-## License
-
-Please adhere to the licensing terms as described [here](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic/blob/main/LICENSE.md).
 """
     model_card = load_or_create_model_card(
         repo_id_or_path=repo_id,
         from_training=True,
-        license="openrail++" if "playgroundai" not in base_model else "playground-v2dot5-community",
+        license="openrail++",
         base_model=base_model,
         prompt=instance_prompt,
         model_description=model_description,
@@ -156,18 +131,15 @@ Please adhere to the licensing terms as described [here](https://huggingface.co/
     )
     tags = [
         "text-to-image",
+        "stable-diffusion-xl",
+        "stable-diffusion-xl-diffusers",
         "text-to-image",
-        "diffusers-training",
         "diffusers",
-        "lora" if not use_dora else "dora",
+        "lora",
         "template:sd-lora",
     ]
-    if "playgroundai" in base_model:
-        tags.extend(["playground", "playground-diffusers"])
-    else:
-        tags.extend(["stable-diffusion-xl", "stable-diffusion-xl-diffusers"])
-
     model_card = populate_model_card(model_card, tags=tags)
+
     model_card.save(os.path.join(repo_folder, "README.md"))
 
 
@@ -187,29 +159,23 @@ def log_validation(
     # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
     scheduler_args = {}
 
-    if not args.do_edm_style_training:
-        if "variance_type" in pipeline.scheduler.config:
-            variance_type = pipeline.scheduler.config.variance_type
+    if "variance_type" in pipeline.scheduler.config:
+        variance_type = pipeline.scheduler.config.variance_type
 
-            if variance_type in ["learned", "learned_range"]:
-                variance_type = "fixed_small"
+        if variance_type in ["learned", "learned_range"]:
+            variance_type = "fixed_small"
 
-            scheduler_args["variance_type"] = variance_type
+        scheduler_args["variance_type"] = variance_type
 
-        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
+    pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
 
     pipeline = pipeline.to(accelerator.device)
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
     generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
-    # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
-    # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
-    inference_ctx = (
-        contextlib.nullcontext() if "playgroundai" in args.pretrained_model_name_or_path else torch.cuda.amp.autocast()
-    )
 
-    with inference_ctx:
+    with torch.cuda.amp.autocast():
         images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
 
     for tracker in accelerator.trackers:
@@ -368,12 +334,6 @@ def parse_args(input_args=None):
             " `args.validation_prompt` multiple times: `args.num_validation_images`."
         ),
     )
-    parser.add_argument(
-        "--do_edm_style_training",
-        default=False,
-        action="store_true",
-        help="Flag to conduct training using the EDM formulation as introduced in https://arxiv.org/abs/2206.00364.",
-    )
     parser.add_argument(
         "--with_prior_preservation",
         default=False,
@@ -647,15 +607,6 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
-    parser.add_argument(
-        "--use_dora",
-        action="store_true",
-        default=False,
-        help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
-            "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
-        ),
-    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -877,8 +828,6 @@ def collate_fn(examples, with_prior_preservation=False):
     if with_prior_preservation:
         pixel_values += [example["class_images"] for example in examples]
         prompts += [example["class_prompt"] for example in examples]
-        original_sizes += [example["original_size"] for example in examples]
-        crop_top_lefts += [example["crop_top_left"] for example in examples]
 
     pixel_values = torch.stack(pixel_values)
     pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
@@ -956,9 +905,6 @@ def main(args):
             " Please use `huggingface-cli login` to authenticate with the Hub."
         )
 
-    if args.do_edm_style_training and args.snr_gamma is not None:
-        raise ValueError("Min-SNR formulation is not supported when conducting EDM-style training.")
-
     logging_dir = Path(args.output_dir, args.logging_dir)
 
     accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
@@ -1072,19 +1018,7 @@ def main(args):
     )
 
     # Load scheduler and models
-    scheduler_type = determine_scheduler_type(args.pretrained_model_name_or_path, args.revision)
-    if "EDM" in scheduler_type:
-        args.do_edm_style_training = True
-        noise_scheduler = EDMEulerScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
-        logger.info("Performing EDM-style training!")
-    elif args.do_edm_style_training:
-        noise_scheduler = EulerDiscreteScheduler.from_pretrained(
-            args.pretrained_model_name_or_path, subfolder="scheduler"
-        )
-        logger.info("Performing EDM-style training!")
-    else:
-        noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
-
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
     text_encoder_one = text_encoder_cls_one.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
     )
@@ -1102,12 +1036,6 @@ def main(args):
         revision=args.revision,
         variant=args.variant,
     )
-    latents_mean = latents_std = None
-    if hasattr(vae.config, "latents_mean") and vae.config.latents_mean is not None:
-        latents_mean = torch.tensor(vae.config.latents_mean).view(1, 4, 1, 1)
-    if hasattr(vae.config, "latents_std") and vae.config.latents_std is not None:
-        latents_std = torch.tensor(vae.config.latents_std).view(1, 4, 1, 1)
-
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
@@ -1158,7 +1086,6 @@ def main(args):
     # now we will add new LoRA weights to the attention layers
     unet_lora_config = LoraConfig(
         r=args.rank,
-        use_dora=args.use_dora,
         lora_alpha=args.rank,
         init_lora_weights="gaussian",
         target_modules=["to_k", "to_q", "to_v", "to_out.0"],
@@ -1170,7 +1097,6 @@ def main(args):
     if args.train_text_encoder:
         text_lora_config = LoraConfig(
             r=args.rank,
-            use_dora=args.use_dora,
             lora_alpha=args.rank,
             init_lora_weights="gaussian",
             target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
@@ -1252,7 +1178,7 @@ def main(args):
             _set_state_dict_into_text_encoder(lora_state_dict, prefix="text_encoder.", text_encoder=text_encoder_one_)
 
             _set_state_dict_into_text_encoder(
-                lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_two_
+                lora_state_dict, prefix="text_encoder_2.", text_encoder=text_encoder_one_
             )
 
         # Make sure the trainable params are in float32. This is again needed since the base models
@@ -1507,12 +1433,7 @@ def main(args):
     # We need to initialize the trackers we use, and also store our configuration.
     # The trackers initializes automatically on the main process.
     if accelerator.is_main_process:
-        tracker_name = (
-            "dreambooth-lora-sd-xl"
-            if "playgroundai" not in args.pretrained_model_name_or_path
-            else "dreambooth-lora-playground"
-        )
-        accelerator.init_trackers(tracker_name, config=vars(args))
+        accelerator.init_trackers("dreambooth-lora-sd-xl", config=vars(args))
 
     # Train!
     total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
@@ -1564,18 +1485,6 @@ def main(args):
         disable=not accelerator.is_local_main_process,
     )
 
-    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
-        sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype)
-        schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device)
-        timesteps = timesteps.to(accelerator.device)
-
-        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
-
-        sigma = sigmas[step_indices].flatten()
-        while len(sigma.shape) < n_dim:
-            sigma = sigma.unsqueeze(-1)
-        return sigma
-
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         if args.train_text_encoder:
@@ -1603,46 +1512,22 @@ def main(args):
 
                 # Convert images to latent space
                 model_input = vae.encode(pixel_values).latent_dist.sample()
-
-                if latents_mean is None and latents_std is None:
-                    model_input = model_input * vae.config.scaling_factor
-                    if args.pretrained_vae_model_name_or_path is None:
-                        model_input = model_input.to(weight_dtype)
-                else:
-                    latents_mean = latents_mean.to(device=model_input.device, dtype=model_input.dtype)
-                    latents_std = latents_std.to(device=model_input.device, dtype=model_input.dtype)
-                    model_input = (model_input - latents_mean) * vae.config.scaling_factor / latents_std
-                    model_input = model_input.to(dtype=weight_dtype)
+                model_input = model_input * vae.config.scaling_factor
+                if args.pretrained_vae_model_name_or_path is None:
+                    model_input = model_input.to(weight_dtype)
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(model_input)
                 bsz = model_input.shape[0]
-
                 # Sample a random timestep for each image
-                if not args.do_edm_style_training:
-                    timesteps = torch.randint(
-                        0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
-                    )
-                    timesteps = timesteps.long()
-                else:
-                    # in EDM formulation, the model is conditioned on the pre-conditioned noise levels
-                    # instead of discrete timesteps, so here we sample indices to get the noise levels
-                    # from `scheduler.timesteps`
-                    indices = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,))
-                    timesteps = noise_scheduler.timesteps[indices].to(device=model_input.device)
+                timesteps = torch.randint(
+                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
+                )
+                timesteps = timesteps.long()
 
                 # Add noise to the model input according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
                 noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
-                # For EDM-style training, we first obtain the sigmas based on the continuous timesteps.
-                # We then precondition the final model inputs based on these sigmas instead of the timesteps.
-                # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
-                if args.do_edm_style_training:
-                    sigmas = get_sigmas(timesteps, len(noisy_model_input.shape), noisy_model_input.dtype)
-                    if "EDM" in scheduler_type:
-                        inp_noisy_latents = noise_scheduler.precondition_inputs(noisy_model_input, sigmas)
-                    else:
-                        inp_noisy_latents = noisy_model_input / ((sigmas**2 + 1) ** 0.5)
 
                 # time ids
                 add_time_ids = torch.cat(
@@ -1666,7 +1551,7 @@ def main(args):
                     }
                     prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
                     model_pred = unet(
-                        inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
+                        noisy_model_input,
                         timesteps,
                         prompt_embeds_input,
                         added_cond_kwargs=unet_added_conditions,
@@ -1685,43 +1570,18 @@ def main(args):
                     )
                     prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
                     model_pred = unet(
-                        inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
+                        noisy_model_input,
                         timesteps,
                         prompt_embeds_input,
                         added_cond_kwargs=unet_added_conditions,
                         return_dict=False,
                     )[0]
 
-                weighting = None
-                if args.do_edm_style_training:
-                    # Similar to the input preconditioning, the model predictions are also preconditioned
-                    # on noised model inputs (before preconditioning) and the sigmas.
-                    # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
-                    if "EDM" in scheduler_type:
-                        model_pred = noise_scheduler.precondition_outputs(noisy_model_input, model_pred, sigmas)
-                    else:
-                        if noise_scheduler.config.prediction_type == "epsilon":
-                            model_pred = model_pred * (-sigmas) + noisy_model_input
-                        elif noise_scheduler.config.prediction_type == "v_prediction":
-                            model_pred = model_pred * (-sigmas / (sigmas**2 + 1) ** 0.5) + (
-                                noisy_model_input / (sigmas**2 + 1)
-                            )
-                    # We are not doing weighting here because it tends result in numerical problems.
-                    # See: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
-                    # There might be other alternatives for weighting as well:
-                    # https://github.com/huggingface/diffusers/pull/7126#discussion_r1505404686
-                    if "EDM" not in scheduler_type:
-                        weighting = (sigmas**-2.0).float()
-
                 # Get the target for loss depending on the prediction type
                 if noise_scheduler.config.prediction_type == "epsilon":
-                    target = model_input if args.do_edm_style_training else noise
+                    target = noise
                 elif noise_scheduler.config.prediction_type == "v_prediction":
-                    target = (
-                        model_input
-                        if args.do_edm_style_training
-                        else noise_scheduler.get_velocity(model_input, noise, timesteps)
-                    )
+                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
                 else:
                     raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
@@ -1731,28 +1591,10 @@ def main(args):
                     target, target_prior = torch.chunk(target, 2, dim=0)
 
                     # Compute prior loss
-                    if weighting is not None:
-                        prior_loss = torch.mean(
-                            (weighting.float() * (model_pred_prior.float() - target_prior.float()) ** 2).reshape(
-                                target_prior.shape[0], -1
-                            ),
-                            1,
-                        )
-                        prior_loss = prior_loss.mean()
-                    else:
-                        prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
+                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
 
                 if args.snr_gamma is None:
-                    if weighting is not None:
-                        loss = torch.mean(
-                            (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(
-                                target.shape[0], -1
-                            ),
-                            1,
-                        )
-                        loss = loss.mean()
-                    else:
-                        loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 else:
                     # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
@@ -1854,6 +1696,7 @@ def main(args):
                     variant=args.variant,
                     torch_dtype=weight_dtype,
                 )
+
                 pipeline_args = {"prompt": args.validation_prompt}
 
                 images = log_validation(
@@ -1921,13 +1764,12 @@ def main(args):
                 accelerator,
                 pipeline_args,
                 epoch,
-                is_final_validation=True,
+                final_validation=True,
             )
 
         if args.push_to_hub:
             save_model_card(
                 repo_id,
-                use_dora=args.use_dora,
                 images=images,
                 base_model=args.pretrained_model_name_or_path,
                 train_text_encoder=args.train_text_encoder,

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -81,7 +81,6 @@ tags:
 - kandinsky
 - text-to-image
 - diffusers
-- diffusers-training
 inference: true
 ---
     """

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -65,7 +65,6 @@ tags:
 - kandinsky
 - text-to-image
 - diffusers
-- diffusers-training
 - lora
 inference: true
 ---

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -65,7 +65,6 @@ tags:
 - kandinsky
 - text-to-image
 - diffusers
-- diffusers-training
 - lora
 inference: true
 ---

examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py

@@ -82,7 +82,6 @@ tags:
 - kandinsky
 - text-to-image
 - diffusers
-- diffusers-training
 inference: true
 ---
     """

examples/research_projects/controlnet/train_controlnet_webdataset.py

@@ -460,8 +460,6 @@ tags:
 - text-to-image
 - diffusers
 - controlnet
-- diffusers-training
-- webdataset
 inference: true
 ---
     """

examples/research_projects/controlnetxs/pipeline_controlnet_xs.py

@@ -26,7 +26,7 @@ from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from diffusers.models import AutoencoderKL, UNet2DConditionModel
 from diffusers.models.lora import adjust_lora_scale_text_encoder
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 from diffusers.schedulers import KarrasDiffusionSchedulers
@@ -44,7 +44,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 class StableDiffusionControlNetXSPipeline(
-    DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion with ControlNet-XS guidance.
@@ -139,6 +139,39 @@ class StableDiffusionControlNetXSPipeline(
         )
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
     def _encode_prompt(
         self,
@@ -563,6 +596,34 @@ class StableDiffusionControlNetXSPipeline(
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
     @torch.no_grad()
     def __call__(
         self,