[chore] remove torch.save from remnant code. (#12717)

remove torch.save from remnant code.

.github/workflows/nightly_tests.yml

@@ -73,6 +73,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
           uv pip install pytest-reportlog
       - name: Environment
         run: |
@@ -84,7 +86,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx" \
+             -k "not Flax and not Onnx" \
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             --report-log=tests_pipeline_${{ matrix.module }}_cuda.log \
             tests/pipelines/${{ matrix.module }}
@@ -126,6 +128,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
         uv pip install pytest-reportlog
     - name: Environment
       run: python utils/print_env.py
@@ -138,7 +142,7 @@ jobs:
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
         pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_torch_${{ matrix.module }}_cuda \
           --report-log=tests_torch_${{ matrix.module }}_cuda.log \
           tests/${{ matrix.module }}
@@ -151,7 +155,7 @@ jobs:
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
         pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v --make-reports=examples_torch_cuda \
+          --make-reports=examples_torch_cuda \
           --report-log=examples_torch_cuda.log \
           examples/
 
@@ -190,6 +194,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
     - name: Environment
       run: |
         python utils/print_env.py
@@ -198,7 +204,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         RUN_COMPILE: yes
       run: |
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile -k "compile" --make-reports=tests_torch_compile_cuda tests/
     - name: Failure short reports
       if: ${{ failure() }}
       run: cat reports/tests_torch_compile_cuda_failures_short.txt
@@ -232,6 +238,8 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip install peft@git+https://github.com/huggingface/peft.git
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
           uv pip install pytest-reportlog
       - name: Environment
         run: |
@@ -281,6 +289,8 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip install peft@git+https://github.com/huggingface/peft.git
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
 
       - name: Environment
         run: |
@@ -293,7 +303,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx" \
+            -k "not Flax and not Onnx" \
             --make-reports=tests_torch_minimum_version_cuda \
             tests/models/test_modeling_common.py \
             tests/pipelines/test_pipelines_common.py \
@@ -358,6 +368,8 @@ jobs:
               uv pip install ${{ join(matrix.config.additional_deps, ' ') }}
           fi
           uv pip install pytest-reportlog
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
       - name: Environment
         run: |
           python utils/print_env.py
@@ -405,6 +417,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip install -U bitsandbytes optimum_quanto
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1 
           uv pip install pytest-reportlog
       - name: Environment
         run: |
@@ -531,7 +545,7 @@ jobs:
 #          HF_HOME: /System/Volumes/Data/mnt/cache
 #          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
 #        run: |
-#          ${CONDA_RUN} pytest -n 1 -s -v --make-reports=tests_torch_mps \
+#          ${CONDA_RUN} pytest -n 1  --make-reports=tests_torch_mps \
 #            --report-log=tests_torch_mps.log \
 #            tests/
 #      - name: Failure short reports
@@ -587,7 +601,7 @@ jobs:
 #          HF_HOME: /System/Volumes/Data/mnt/cache
 #          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
 #        run: |
-#          ${CONDA_RUN} pytest -n 1 -s -v --make-reports=tests_torch_mps \
+#          ${CONDA_RUN} pytest -n 1  --make-reports=tests_torch_mps \
 #            --report-log=tests_torch_mps.log \
 #            tests/
 #      - name: Failure short reports

.github/workflows/pr_modular_tests.yml

@@ -109,7 +109,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality]"
-        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
 
     - name: Environment
@@ -120,7 +121,7 @@ jobs:
       if: ${{ matrix.config.framework == 'pytorch_pipelines' }}
       run: |
         pytest -n 8 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
           tests/modular_pipelines
 

.github/workflows/pr_tests.yml

@@ -115,7 +115,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality]"
-        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
 
     - name: Environment
@@ -126,7 +127,7 @@ jobs:
       if: ${{ matrix.config.framework == 'pytorch_pipelines' }}
       run: |
         pytest -n 8 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
           tests/pipelines
 
@@ -134,7 +135,7 @@ jobs:
       if: ${{ matrix.config.framework == 'pytorch_models' }}
       run: |
         pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx and not Dependency" \
+          -k "not Flax and not Onnx and not Dependency" \
           --make-reports=tests_${{ matrix.config.report }} \
           tests/models tests/schedulers tests/others
 
@@ -246,7 +247,8 @@ jobs:
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
-        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -255,11 +257,11 @@ jobs:
     - name: Run fast PyTorch LoRA tests with PEFT
       run: |
         pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
+          \
           --make-reports=tests_peft_main \
           tests/lora/
         pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
+          \
           --make-reports=tests_models_lora_peft_main \
           tests/models/ -k "lora"
 

.github/workflows/pr_tests_gpu.yml

@@ -1,4 +1,4 @@
-name: Fast GPU Tests on PR 
+name: Fast GPU Tests on PR
 
 on:
   pull_request:
@@ -71,7 +71,7 @@ jobs:
         if: ${{ failure() }}
         run: |
           echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
-  
+
   setup_torch_cuda_pipeline_matrix:
     needs: [check_code_quality, check_repository_consistency]
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
@@ -131,7 +131,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-          uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
       - name: Environment
         run: |
@@ -149,18 +150,18 @@ jobs:
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
-          if [ "${{ matrix.module }}" = "ip_adapters" ]; then 
+          if [ "${{ matrix.module }}" = "ip_adapters" ]; then
               pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-              -s -v -k "not Flax and not Onnx" \
+              -k "not Flax and not Onnx" \
               --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
               tests/pipelines/${{ matrix.module }}
-          else 
+          else
               pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
               pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-              -s -v -k "not Flax and not Onnx and $pattern" \
+              -k "not Flax and not Onnx and $pattern" \
               --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
               tests/pipelines/${{ matrix.module }}
-          fi 
+          fi
 
       - name: Failure short reports
         if: ${{ failure() }}
@@ -201,7 +202,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -222,11 +224,11 @@ jobs:
       run: |
         pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
         if [ -z "$pattern" ]; then
-          pytest -n 1 -sv --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx" tests/${{ matrix.module }} \
-          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+          pytest -n 1  --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx" tests/${{ matrix.module }} \
+          --make-reports=tests_torch_cuda_${{ matrix.module }}
         else
-          pytest -n 1 -sv --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx and $pattern" tests/${{ matrix.module }} \
-          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+          pytest -n 1  --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx and $pattern" tests/${{ matrix.module }} \
+          --make-reports=tests_torch_cuda_${{ matrix.module }}
         fi
 
     - name: Failure short reports
@@ -262,7 +264,8 @@ jobs:
         nvidia-smi
     - name: Install dependencies
       run: |
-        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip install -e ".[quality,training]"
 
     - name: Environment
@@ -274,7 +277,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
         uv pip install ".[training]"
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile --make-reports=examples_torch_cuda examples/
 
     - name: Failure short reports
       if: ${{ failure() }}

.github/workflows/push_tests.yml

@@ -76,6 +76,8 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
       - name: Environment
         run: |
           python utils/print_env.py
@@ -86,7 +88,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx" \
+            -k "not Flax and not Onnx" \
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             tests/pipelines/${{ matrix.module }}
       - name: Failure short reports
@@ -127,6 +129,8 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -139,7 +143,7 @@ jobs:
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
         pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_torch_cuda_${{ matrix.module }} \
           tests/${{ matrix.module }}
 
@@ -178,6 +182,8 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
+        #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
     - name: Environment
       run: |
         python utils/print_env.py
@@ -186,7 +192,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         RUN_COMPILE: yes
       run: |
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile -k "compile" --make-reports=tests_torch_compile_cuda tests/
     - name: Failure short reports
       if: ${{ failure() }}
       run: cat reports/tests_torch_compile_cuda_failures_short.txt
@@ -227,7 +233,7 @@ jobs:
       env:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "xformers" --make-reports=tests_torch_xformers_cuda tests/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile -k "xformers" --make-reports=tests_torch_xformers_cuda tests/
     - name: Failure short reports
       if: ${{ failure() }}
       run: cat reports/tests_torch_xformers_cuda_failures_short.txt
@@ -270,7 +276,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
         uv pip install ".[training]"
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile --make-reports=examples_torch_cuda examples/
 
     - name: Failure short reports
       if: ${{ failure() }}

.github/workflows/push_tests_fast.yml

@@ -70,7 +70,7 @@ jobs:
       if: ${{ matrix.config.framework == 'pytorch' }}
       run: |
         pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_${{ matrix.config.report }} \
           tests/
 

.github/workflows/push_tests_mps.yml

@@ -57,7 +57,7 @@ jobs:
         HF_HOME: /System/Volumes/Data/mnt/cache
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |
-        ${CONDA_RUN} python -m pytest -n 0 -s -v --make-reports=tests_torch_mps tests/
+        ${CONDA_RUN} python -m pytest -n 0 --make-reports=tests_torch_mps tests/
 
     - name: Failure short reports
       if: ${{ failure() }}

.github/workflows/release_tests_fast.yml

@@ -84,7 +84,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx" \
+            -k "not Flax and not Onnx" \
             --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
             tests/pipelines/${{ matrix.module }}
       - name: Failure short reports
@@ -137,7 +137,7 @@ jobs:
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
         pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-          -s -v -k "not Flax and not Onnx" \
+          -k "not Flax and not Onnx" \
           --make-reports=tests_torch_${{ matrix.module }}_cuda \
           tests/${{ matrix.module }}
 
@@ -187,7 +187,7 @@ jobs:
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
           pytest -n 1 --max-worker-restart=0 --dist=loadfile \
-            -s -v -k "not Flax and not Onnx" \
+            -k "not Flax and not Onnx" \
             --make-reports=tests_torch_minimum_cuda \
             tests/models/test_modeling_common.py \
             tests/pipelines/test_pipelines_common.py \
@@ -240,7 +240,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         RUN_COMPILE: yes
       run: |
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile -k "compile" --make-reports=tests_torch_compile_cuda tests/
     - name: Failure short reports
       if: ${{ failure() }}
       run: cat reports/tests_torch_compile_cuda_failures_short.txt
@@ -281,7 +281,7 @@ jobs:
       env:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "xformers" --make-reports=tests_torch_xformers_cuda tests/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile -k "xformers" --make-reports=tests_torch_xformers_cuda tests/
     - name: Failure short reports
       if: ${{ failure() }}
       run: cat reports/tests_torch_xformers_cuda_failures_short.txt
@@ -326,7 +326,7 @@ jobs:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
         uv pip install ".[training]"
-        pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
+        pytest -n 1 --max-worker-restart=0 --dist=loadfile --make-reports=examples_torch_cuda examples/
 
     - name: Failure short reports
       if: ${{ failure() }}

docs/source/en/_toctree.yml

@@ -22,6 +22,8 @@
     title: Reproducibility
   - local: using-diffusers/schedulers
     title: Schedulers
+  - local: using-diffusers/automodel
+    title: AutoModel
   - local: using-diffusers/other-formats
     title: Model formats
   - local: using-diffusers/push_to_hub
@@ -119,6 +121,8 @@
     title: ComponentsManager
   - local: modular_diffusers/guiders
     title: Guiders
+  - local: modular_diffusers/custom_blocks
+    title: Building Custom Blocks
   title: Modular Diffusers
 - isExpanded: false
   sections:
@@ -323,10 +327,14 @@
         title: AllegroTransformer3DModel
       - local: api/models/aura_flow_transformer2d
         title: AuraFlowTransformer2DModel
+      - local: api/models/transformer_bria_fibo
+        title: BriaFiboTransformer2DModel
       - local: api/models/bria_transformer
         title: BriaTransformer2DModel
       - local: api/models/chroma_transformer
         title: ChromaTransformer2DModel
+      - local: api/models/chronoedit_transformer_3d
+        title: ChronoEditTransformer3DModel
       - local: api/models/cogvideox_transformer3d
         title: CogVideoXTransformer3DModel
       - local: api/models/cogview3plus_transformer2d
@@ -341,6 +349,8 @@
         title: DiTTransformer2DModel
       - local: api/models/easyanimate_transformer3d
         title: EasyAnimateTransformer3DModel
+      - local: api/models/flux2_transformer
+        title: Flux2Transformer2DModel
       - local: api/models/flux_transformer
         title: FluxTransformer2DModel
       - local: api/models/hidream_image_transformer
@@ -371,6 +381,8 @@
         title: QwenImageTransformer2DModel
       - local: api/models/sana_transformer2d
         title: SanaTransformer2DModel
+      - local: api/models/sana_video_transformer3d
+        title: SanaVideoTransformer3DModel
       - local: api/models/sd3_transformer2d
         title: SD3Transformer2DModel
       - local: api/models/skyreels_v2_transformer_3d
@@ -381,6 +393,8 @@
         title: Transformer2DModel
       - local: api/models/transformer_temporal
         title: TransformerTemporalModel
+      - local: api/models/wan_animate_transformer_3d
+        title: WanAnimateTransformer3DModel
       - local: api/models/wan_transformer_3d
         title: WanTransformer3DModel
       title: Transformers
@@ -442,6 +456,8 @@
   - sections:
     - local: api/pipelines/overview
       title: Overview
+    - local: api/pipelines/auto_pipeline
+      title: AutoPipeline
     - sections:
       - local: api/pipelines/audioldm
         title: AudioLDM
@@ -454,8 +470,6 @@
       - local: api/pipelines/stable_audio
         title: Stable Audio
       title: Audio
-    - local: api/pipelines/auto_pipeline
-      title: AutoPipeline
     - sections:
       - local: api/pipelines/amused
         title: aMUSEd
@@ -469,6 +483,8 @@
         title: BLIP-Diffusion
       - local: api/pipelines/bria_3_2
         title: Bria 3.2
+      - local: api/pipelines/bria_fibo
+        title: Bria Fibo
       - local: api/pipelines/chroma
         title: Chroma
       - local: api/pipelines/cogview3
@@ -511,12 +527,16 @@
         title: EasyAnimate
       - local: api/pipelines/flux
         title: Flux
+      - local: api/pipelines/flux2
+        title: Flux2
       - local: api/pipelines/control_flux_inpaint
         title: FluxControlInpaint
       - local: api/pipelines/hidream
         title: HiDream-I1
       - local: api/pipelines/hunyuandit
         title: Hunyuan-DiT
+      - local: api/pipelines/hunyuanimage21
+        title: HunyuanImage2.1
       - local: api/pipelines/pix2pix
         title: InstructPix2Pix
       - local: api/pipelines/kandinsky
@@ -525,8 +545,6 @@
         title: Kandinsky 2.2
       - local: api/pipelines/kandinsky3
         title: Kandinsky 3
-      - local: api/pipelines/kandinsky5
-        title: Kandinsky 5
       - local: api/pipelines/kolors
         title: Kolors
       - local: api/pipelines/latent_consistency_models
@@ -561,6 +579,8 @@
         title: Sana
       - local: api/pipelines/sana_sprint
         title: Sana Sprint
+      - local: api/pipelines/sana_video
+        title: Sana Video
       - local: api/pipelines/self_attention_guidance
         title: Self-Attention Guidance
       - local: api/pipelines/semantic_stable_diffusion
@@ -622,18 +642,20 @@
     - sections:
       - local: api/pipelines/allegro
         title: Allegro
+      - local: api/pipelines/chronoedit
+        title: ChronoEdit
       - local: api/pipelines/cogvideox
         title: CogVideoX
       - local: api/pipelines/consisid
         title: ConsisID
       - local: api/pipelines/framepack
         title: Framepack
-      - local: api/pipelines/hunyuanimage21
-        title: HunyuanImage2.1
       - local: api/pipelines/hunyuan_video
         title: HunyuanVideo
       - local: api/pipelines/i2vgenxl
         title: I2VGen-XL
+      - local: api/pipelines/kandinsky5_video
+        title: Kandinsky 5.0 Video
       - local: api/pipelines/latte
         title: Latte
       - local: api/pipelines/ltx_video

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

@@ -30,7 +30,8 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`CogView4LoraLoaderMixin`] provides similar functions for [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`HiDreamImageLoraLoaderMixin`] provides similar functions for [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream)
-- [`QwenImageLoraLoaderMixin`] provides similar functions for [Qwen Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/qwen)
+- [`QwenImageLoraLoaderMixin`] provides similar functions for [Qwen Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/qwen).
+- [`Flux2LoraLoaderMixin`] provides similar functions for [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux2).
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.
 
 > [!TIP]
@@ -56,6 +57,10 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 
 [[autodoc]] loaders.lora_pipeline.FluxLoraLoaderMixin
 
+## Flux2LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.Flux2LoraLoaderMixin
+
 ## CogVideoXLoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.CogVideoXLoraLoaderMixin

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

@@ -12,15 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # AutoModel
 
-The `AutoModel` is designed to make it easy to load a checkpoint without needing to know the specific model class. `AutoModel` automatically retrieves the correct model class from the checkpoint `config.json` file.
-
-```python
-from diffusers import AutoModel, AutoPipelineForText2Image
-
-unet = AutoModel.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="unet")
-pipe = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", unet=unet)
-```
-
+[`AutoModel`] automatically retrieves the correct model class from the checkpoint `config.json` file.
 
 ## AutoModel
 

docs/source/en/api/models/chronoedit_transformer_3d.md

@@ -0,0 +1,32 @@
+<!-- Copyright 2025 The ChronoEdit Team and 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. -->
+
+# ChronoEditTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data from [ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation](https://huggingface.co/papers/2510.04290) from NVIDIA and University of Toronto, by Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M. Alvarez, Jun Gao, Sanja Fidler, Zian Wang, Huan Ling.
+
+> **TL;DR:** ChronoEdit reframes image editing as a video generation task, using input and edited images as start/end frames to leverage pretrained video models with temporal consistency. A temporal reasoning stage introduces reasoning tokens to ensure physically plausible edits and visualize the editing trajectory.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import ChronoEditTransformer3DModel
+
+transformer = ChronoEditTransformer3DModel.from_pretrained("nvidia/ChronoEdit-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## ChronoEditTransformer3DModel
+
+[[autodoc]] ChronoEditTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/models/flux2_transformer.md

@@ -0,0 +1,19 @@
+<!--Copyright 2025 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.
+-->
+
+# Flux2Transformer2DModel
+
+A Transformer model for image-like data from [Flux2](https://hf.co/black-forest-labs/FLUX.2-dev).
+
+## Flux2Transformer2DModel
+
+[[autodoc]] Flux2Transformer2DModel

docs/source/en/api/models/sana_video_transformer3d.md

@@ -0,0 +1,36 @@
+<!-- Copyright 2025 The SANA-Video Authors and 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. -->
+
+# SanaVideoTransformer3DModel
+
+A Diffusion Transformer model for 3D data (video) from [SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer](https://huggingface.co/papers/2509.24695) from NVIDIA and MIT HAN Lab, by Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie.
+
+The abstract from the paper is:
+
+*We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation.*
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import SanaVideoTransformer3DModel
+import torch
+
+transformer = SanaVideoTransformer3DModel.from_pretrained("Efficient-Large-Model/SANA-Video_2B_480p_diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## SanaVideoTransformer3DModel
+
+[[autodoc]] SanaVideoTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
+

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

@@ -0,0 +1,19 @@
+<!--Copyright 2025 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.
+-->
+
+# BriaFiboTransformer2DModel
+
+A modified flux Transformer model from [Bria](https://huggingface.co/briaai/FIBO)
+
+## BriaFiboTransformer2DModel
+
+[[autodoc]] BriaFiboTransformer2DModel

docs/source/en/api/models/wan_animate_transformer_3d.md

@@ -0,0 +1,30 @@
+<!-- Copyright 2025 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. -->
+
+# WanAnimateTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [Wan Animate](https://github.com/Wan-Video/Wan2.2) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import WanAnimateTransformer3DModel
+
+transformer = WanAnimateTransformer3DModel.from_pretrained("Wan-AI/Wan2.2-Animate-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## WanAnimateTransformer3DModel
+
+[[autodoc]] WanAnimateTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

@@ -0,0 +1,45 @@
+<!--Copyright 2025 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.
+-->
+
+# Bria Fibo
+
+Text-to-image models have mastered imagination - but not control. FIBO changes that.
+
+FIBO is trained on structured JSON captions up to 1,000+ words and designed to understand and control different visual parameters such as lighting, composition, color, and camera settings, enabling precise and reproducible outputs.
+
+With only 8 billion parameters, FIBO provides a new level of image quality, prompt adherence and proffesional control.
+
+FIBO is trained exclusively on a structured prompt and will not work with freeform text prompts.
+you can use the [FIBO-VLM-prompt-to-JSON](https://huggingface.co/briaai/FIBO-VLM-prompt-to-JSON) model or the [FIBO-gemini-prompt-to-JSON](https://huggingface.co/briaai/FIBO-gemini-prompt-to-JSON)  to convert your freeform text prompt to a structured JSON prompt.
+
+its not recommended to use freeform text prompts directly with FIBO, as it will not produce the best results.
+
+you can learn more about FIBO in  [Bria Fibo Hugging Face page](https://huggingface.co/briaai/FIBO).
+
+
+## Usage
+
+_As the model is gated, before using it with diffusers you first need to go to the [Bria Fibo Hugging Face page](https://huggingface.co/briaai/FIBO), fill in the form and accept the gate. Once you are in, you need to login so that your system knows you’ve accepted the gate._
+
+Use the command below to log in:
+
+```bash
+hf auth login
+```
+
+
+## BriaPipeline
+
+[[autodoc]] BriaPipeline
+	- all
+	- __call__
+

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

@@ -0,0 +1,156 @@
+<!-- Copyright 2025 The ChronoEdit Team and 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. -->
+
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
+</div>
+
+# ChronoEdit
+
+[ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation](https://huggingface.co/papers/2510.04290) from NVIDIA and University of Toronto, by Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M. Alvarez, Jun Gao, Sanja Fidler, Zian Wang, Huan Ling.
+
+> **TL;DR:** ChronoEdit reframes image editing as a video generation task, using input and edited images as start/end frames to leverage pretrained video models with temporal consistency. A temporal reasoning stage introduces reasoning tokens to ensure physically plausible edits and visualize the editing trajectory.
+
+*Recent advances in large generative models have greatly enhanced both image editing and in-context image generation, yet a critical gap remains in ensuring physical consistency, where edited objects must remain coherent. This capability is especially vital for world simulation related tasks. In this paper, we present ChronoEdit, a framework that reframes image editing as a video generation problem. First, ChronoEdit treats the input and edited images as the first and last frames of a video, allowing it to leverage large pretrained video generative models that capture not only object appearance but also the implicit physics of motion and interaction through learned temporal consistency. Second, ChronoEdit introduces a temporal reasoning stage that explicitly performs editing at inference time. Under this setting, target frame is jointly denoised with reasoning tokens to imagine a plausible editing trajectory that constrains the solution space to physically viable transformations. The reasoning tokens are then dropped after a few steps to avoid the high computational cost of rendering a full video. To validate ChronoEdit, we introduce PBench-Edit, a new benchmark of image-prompt pairs for contexts that require physical consistency, and demonstrate that ChronoEdit surpasses state-of-the-art baselines in both visual fidelity and physical plausibility. Project page for code and models: [this https URL](https://research.nvidia.com/labs/toronto-ai/chronoedit).*
+
+The ChronoEdit pipeline is developed by the ChronoEdit Team. The original code is available on [GitHub](https://github.com/nv-tlabs/ChronoEdit), and pretrained models can be found in the [nvidia/ChronoEdit](https://huggingface.co/collections/nvidia/chronoedit) collection on Hugging Face.
+
+
+### Image Editing
+
+```py
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, ChronoEditTransformer3DModel, ChronoEditPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+from PIL import Image
+
+model_id = "nvidia/ChronoEdit-14B-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = ChronoEditTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+pipe = ChronoEditPipeline.from_pretrained(model_id, image_encoder=image_encoder, transformer=transformer, vae=vae, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/spaces/nvidia/ChronoEdit/resolve/main/examples/3.png"
+)
+max_area = 720 * 1280
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+print("width", width, "height", height)
+image = image.resize((width, height))
+prompt = (
+    "The user wants to transform the image by adding a small, cute mouse sitting inside the floral teacup, enjoying a spa bath. The mouse should appear relaxed and cheerful, with a tiny white bath towel draped over its head like a turban. It should be positioned comfortably in the cup’s liquid, with gentle steam rising around it to blend with the cozy atmosphere. "
+    "The mouse’s pose should be natural—perhaps sitting upright with paws resting lightly on the rim or submerged in the tea. The teacup’s floral design, gold trim, and warm lighting must remain unchanged to preserve the original aesthetic. The steam should softly swirl around the mouse, enhancing the spa-like, whimsical mood."
+)
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=5,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+    enable_temporal_reasoning=False,
+    num_temporal_reasoning_steps=0,
+).frames[0]
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+Optionally, enable **temporal reasoning** for improved physical consistency:
+```py
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=29,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+    enable_temporal_reasoning=True,
+    num_temporal_reasoning_steps=50,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+### Inference with 8-Step Distillation Lora
+
+```py
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, ChronoEditTransformer3DModel, ChronoEditPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+from PIL import Image
+
+model_id = "nvidia/ChronoEdit-14B-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = ChronoEditTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+pipe = ChronoEditPipeline.from_pretrained(model_id, image_encoder=image_encoder, transformer=transformer, vae=vae, torch_dtype=torch.bfloat16)
+lora_path = hf_hub_download(repo_id=model_id, filename="lora/chronoedit_distill_lora.safetensors")
+pipe.load_lora_weights(lora_path)
+pipe.fuse_lora(lora_scale=1.0)
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=2.0)
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/spaces/nvidia/ChronoEdit/resolve/main/examples/3.png"
+)
+max_area = 720 * 1280
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+print("width", width, "height", height)
+image = image.resize((width, height))
+prompt = (
+    "The user wants to transform the image by adding a small, cute mouse sitting inside the floral teacup, enjoying a spa bath. The mouse should appear relaxed and cheerful, with a tiny white bath towel draped over its head like a turban. It should be positioned comfortably in the cup’s liquid, with gentle steam rising around it to blend with the cozy atmosphere. "
+    "The mouse’s pose should be natural—perhaps sitting upright with paws resting lightly on the rim or submerged in the tea. The teacup’s floral design, gold trim, and warm lighting must remain unchanged to preserve the original aesthetic. The steam should softly swirl around the mouse, enhancing the spa-like, whimsical mood."
+)
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=5,
+    num_inference_steps=8,
+    guidance_scale=1.0,
+    enable_temporal_reasoning=False,
+    num_temporal_reasoning_steps=0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+## ChronoEditPipeline
+
+[[autodoc]] ChronoEditPipeline
+  - all
+  - __call__
+
+## ChronoEditPipelineOutput
+
+[[autodoc]] pipelines.chronoedit.pipeline_output.ChronoEditPipelineOutput

docs/source/en/api/pipelines/flux2.md

@@ -0,0 +1,33 @@
+<!--Copyright 2025 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.
+-->
+
+# Flux2
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
+</div>
+
+Flux.2 is the recent series of image generation models from Black Forest Labs, preceded by the [Flux.1](./flux.md) series. It is an entirely new model with a new architecture and pre-training done from scratch!
+
+Original model checkpoints for Flux can be found [here](https://huggingface.co/black-forest-labs). Original inference code can be found [here](https://github.com/black-forest-labs/flux2).
+
+> [!TIP]
+> Flux2 can be quite expensive to run on consumer hardware devices. However, you can perform a suite of optimizations to run it faster and in a more memory-friendly manner. Check out [this section](https://huggingface.co/blog/sd3#memory-optimizations-for-sd3) for more details. Additionally, Flux can benefit from quantization for memory efficiency with a trade-off in inference latency. Refer to [this blog post](https://huggingface.co/blog/quanto-diffusers) to learn more.
+>
+> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
+
+## Flux2Pipeline
+
+[[autodoc]] Flux2Pipeline
+	- all
+	- __call__

docs/source/en/api/pipelines/kandinsky5_video.md

@@ -7,9 +7,9 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Kandinsky 5.0
+# Kandinsky 5.0 Video
 
-Kandinsky 5.0 is created by the Kandinsky team: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev, Anna Dmitrienko,  Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova, Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Denis Dimitrov
+Kandinsky 5.0 Video is created by the Kandinsky team: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev, Anna Dmitrienko,  Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova, Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Denis Dimitrov
 
 
 Kandinsky 5.0 is a family of diffusion models for Video & Image generation. Kandinsky 5.0 T2V Lite is a lightweight video generation model (2B parameters) that ranks #1 among open-source models in its class. It outperforms larger models and offers the best understanding of Russian concepts in the open-source ecosystem.
@@ -92,7 +92,7 @@ pipe = pipe.to("cuda")
 
 pipe.transformer.set_attention_backend(
     "flex"
-)                                       # <--- Set attention backend to Flex
+)                                       # <--- Sett attention bakend to Flex
 pipe.transformer.compile(
     mode="max-autotune-no-cudagraphs", 
     dynamic=True
@@ -115,7 +115,7 @@ export_to_video(output, "output.mp4", fps=24, quality=9)
 ```
 
 ### Diffusion Distilled model
-**⚠️ Warning!** all nocfg and diffusion distilled models should be inferred without CFG (```guidance_scale=1.0```):
+**⚠️ Warning!** all nocfg and diffusion distilled models should be infered wothout CFG (```guidance_scale=1.0```):
 
 ```python
 model_id = "ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers"

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

@@ -24,9 +24,6 @@ The abstract from the paper is:
 
 *This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.*
 
-> [!TIP]
-> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
-
 This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj), [shuchen Xue](https://github.com/scxue) and [Enze Xie](https://github.com/xieenze). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model/).
 
 Available models:

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

@@ -0,0 +1,189 @@
+<!-- Copyright 2025 The SANA-Video Authors and 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. -->
+
+# Sana-Video
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
+</div>
+
+[SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer](https://huggingface.co/papers/2509.24695) from NVIDIA and MIT HAN Lab, by Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie.
+
+The abstract from the paper is:
+
+*We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation. [this https URL](https://github.com/NVlabs/SANA).*
+
+This pipeline was contributed by SANA Team. The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://hf.co/collections/Efficient-Large-Model/sana-video).
+
+Available models:
+
+| Model | Recommended dtype |
+|:-----:|:-----------------:|
+| [`Efficient-Large-Model/SANA-Video_2B_480p_diffusers`](https://huggingface.co/Efficient-Large-Model/ANA-Video_2B_480p_diffusers) | `torch.bfloat16` |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-video) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+
+## Generation Pipelines
+
+<hfoptions id="generation pipelines">`
+<hfoption id="Text-to-Video">
+
+The example below demonstrates how to use the text-to-video pipeline to generate a video using a text description.
+
+```python
+pipe = SanaVideoPipeline.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers", 
+    torch_dtype=torch.bfloat16,
+)
+pipe.text_encoder.to(torch.bfloat16)
+pipe.vae.to(torch.float32)
+pipe.to("cuda")
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
+motion_scale = 30
+motion_prompt = f" motion score: {motion_scale}."
+prompt = prompt + motion_prompt
+
+video = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=832,
+    frames=81,
+    guidance_scale=6,
+    num_inference_steps=50,
+    generator=torch.Generator(device="cuda").manual_seed(0),
+).frames[0]
+
+export_to_video(video, "sana_video.mp4", fps=16)
+```
+
+</hfoption>
+<hfoption id="Image-to-Video">
+
+The example below demonstrates how to use the image-to-video pipeline to generate a video using a text description and a starting frame.
+
+```python
+pipe = SanaImageToVideoPipeline.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    torch_dtype=torch.bfloat16,
+)
+pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config, flow_shift=8.0)
+pipe.vae.to(torch.float32)
+pipe.text_encoder.to(torch.bfloat16)
+pipe.to("cuda")
+
+image = load_image("https://raw.githubusercontent.com/NVlabs/Sana/refs/heads/main/asset/samples/i2v-1.png")
+prompt = "A woman stands against a stunning sunset backdrop, her long, wavy brown hair gently blowing in the breeze. She wears a sleeveless, light-colored blouse with a deep V-neckline, which accentuates her graceful posture. The warm hues of the setting sun cast a golden glow across her face and hair, creating a serene and ethereal atmosphere. The background features a blurred landscape with soft, rolling hills and scattered clouds, adding depth to the scene. The camera remains steady, capturing the tranquil moment from a medium close-up angle."
+negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
+motion_scale = 30
+motion_prompt = f" motion score: {motion_scale}."
+prompt = prompt + motion_prompt
+
+motion_scale = 30.0
+
+video = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=832,
+    frames=81,
+    guidance_scale=6,
+    num_inference_steps=50,
+    generator=torch.Generator(device="cuda").manual_seed(0),
+).frames[0]
+
+export_to_video(video, "sana-i2v.mp4", fps=16)
+```
+
+</hfoption>
+</hfoptions>
+
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaVideoPipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaVideoTransformer3DModel, SanaVideoPipeline
+from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+text_encoder_8bit = AutoModel.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    subfolder="text_encoder",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = SanaVideoTransformer3DModel.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+pipeline = SanaVideoPipeline.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    text_encoder=text_encoder_8bit,
+    transformer=transformer_8bit,
+    torch_dtype=torch.float16,
+    device_map="balanced",
+)
+
+model_score = 30
+prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest, golden light glimmers on his hair as sunlight filters through the leaves. He wears a light shirt, wind gently blowing his hair and collar, light dances across his face with his movements. The background is blurred, with dappled light and soft tree shadows in the distance. The camera focuses on his lifted gaze, clear and emotional."
+negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
+motion_prompt = f" motion score: {model_score}."
+prompt = prompt + motion_prompt
+
+output = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=832,
+    num_frames=81,
+    guidance_scale=6.0,
+    num_inference_steps=50
+).frames[0]
+export_to_video(output, "sana-video-output.mp4", fps=16)
+```
+
+## SanaVideoPipeline
+
+[[autodoc]] SanaVideoPipeline
+  - all
+  - __call__
+
+
+## SanaImageToVideoPipeline
+
+[[autodoc]] SanaImageToVideoPipeline
+  - all
+  - __call__
+
+
+## SanaVideoPipelineOutput
+
+[[autodoc]] pipelines.sana_video.pipeline_sana_video.SanaVideoPipelineOutput

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

@@ -40,6 +40,7 @@ The following Wan models are supported in Diffusers:
 - [Wan 2.2 T2V 14B](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B-Diffusers)
 - [Wan 2.2 I2V 14B](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B-Diffusers)
 - [Wan 2.2 TI2V 5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B-Diffusers)
+- [Wan 2.2 Animate 14B](https://huggingface.co/Wan-AI/Wan2.2-Animate-14B-Diffusers)
 
 > [!TIP]
 > Click on the Wan models in the right sidebar for more examples of video generation.
@@ -95,15 +96,15 @@ pipeline = WanPipeline.from_pretrained(
 pipeline.to("cuda")
 
 prompt = """
-The camera rushes from far to near in a low-angle shot, 
-revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
-for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
-Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+The camera rushes from far to near in a low-angle shot,
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground.
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic
 shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
 """
 negative_prompt = """
-Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, 
-low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, 
+Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality,
+low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured,
 misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards
 """
 
@@ -150,15 +151,15 @@ pipeline.transformer = torch.compile(
 )
 
 prompt = """
-The camera rushes from far to near in a low-angle shot, 
-revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
-for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
-Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+The camera rushes from far to near in a low-angle shot,
+revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in
+for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground.
+Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic
 shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
 """
 negative_prompt = """
-Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, 
-low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, 
+Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality,
+low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured,
 misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards
 """
 
@@ -249,6 +250,208 @@ The code snippets available in [this](https://github.com/huggingface/diffusers/p
 
 The general rule of thumb to keep in mind when preparing inputs for the VACE pipeline is that the input images, or frames of a video that you want to use for conditioning, should have a corresponding mask that is black in color. The black mask signifies that the model will not generate new content for that area, and only use those parts for conditioning the generation process. For parts/frames that should be generated by the model, the mask should be white in color.
 
+</hfoption>
+</hfoptions>
+
+### Wan-Animate: Unified Character Animation and Replacement with Holistic Replication
+
+[Wan-Animate](https://huggingface.co/papers/2509.14055) by the Wan Team.
+
+*We introduce Wan-Animate, a unified framework for character animation and replacement. Given a character image and a reference video, Wan-Animate can animate the character by precisely replicating the expressions and movements of the character in the video to generate high-fidelity character videos. Alternatively, it can integrate the animated character into the reference video to replace the original character, replicating the scene's lighting and color tone to achieve seamless environmental integration. Wan-Animate is built upon the Wan model. To adapt it for character animation tasks, we employ a modified input paradigm to differentiate between reference conditions and regions for generation. This design unifies multiple tasks into a common symbolic representation. We use spatially-aligned skeleton signals to replicate body motion and implicit facial features extracted from source images to reenact expressions, enabling the generation of character videos with high controllability and expressiveness. Furthermore, to enhance environmental integration during character replacement, we develop an auxiliary Relighting LoRA. This module preserves the character's appearance consistency while applying the appropriate environmental lighting and color tone. Experimental results demonstrate that Wan-Animate achieves state-of-the-art performance. We are committed to open-sourcing the model weights and its source code.*
+
+The project page: https://humanaigc.github.io/wan-animate
+
+This model was mostly contributed by [M. Tolga Cangöz](https://github.com/tolgacangoz).
+
+#### Usage
+
+The Wan-Animate pipeline supports two modes of operation:
+
+1. **Animation Mode** (default): Animates a character image based on motion and expression from reference videos
+2. **Replacement Mode**: Replaces a character in a background video with a new character while preserving the scene
+
+##### Prerequisites
+
+Before using the pipeline, you need to preprocess your reference video to extract:
+- **Pose video**: Contains skeletal keypoints representing body motion
+- **Face video**: Contains facial feature representations for expression control
+
+For replacement mode, you additionally need:
+- **Background video**: The original video containing the scene
+- **Mask video**: A mask indicating where to generate content (white) vs. preserve original (black)
+
+> [!NOTE]
+> Raw videos should not be used for inputs such as `pose_video`, which the pipeline expects to be preprocessed to extract the proper information. Preprocessing scripts to prepare these inputs are available in the [original Wan-Animate repository](https://github.com/Wan-Video/Wan2.2?tab=readme-ov-file#1-preprocessing). Integration of these preprocessing steps into Diffusers is planned for a future release.
+
+The example below demonstrates how to use the Wan-Animate pipeline:
+
+<hfoptions id="Animate usage">
+<hfoption id="Animation mode">
+
+```python
+import numpy as np
+import torch
+from diffusers import AutoencoderKLWan, WanAnimatePipeline
+from diffusers.utils import export_to_video, load_image, load_video
+
+model_id = "Wan-AI/Wan2.2-Animate-14B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanAnimatePipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+# Load character image and preprocessed videos
+image = load_image("path/to/character.jpg")
+pose_video = load_video("path/to/pose_video.mp4")  # Preprocessed skeletal keypoints
+face_video = load_video("path/to/face_video.mp4")  # Preprocessed facial features
+
+# Resize image to match VAE constraints
+def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+
+image, height, width = aspect_ratio_resize(image, pipe)
+
+prompt = "A person dancing energetically in a studio with dynamic lighting and professional camera work"
+negative_prompt = "blurry, low quality, distorted, deformed, static, poorly drawn"
+
+# Generate animated video
+output = pipe(
+    image=image,
+    pose_video=pose_video,
+    face_video=face_video,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    segment_frame_length=77,
+    guidance_scale=1.0,
+    mode="animate",  # Animation mode (default)
+).frames[0]
+export_to_video(output, "animated_character.mp4", fps=30)
+```
+
+</hfoption>
+<hfoption id="Replacement mode">
+
+```python
+import numpy as np
+import torch
+from diffusers import AutoencoderKLWan, WanAnimatePipeline
+from diffusers.utils import export_to_video, load_image, load_video
+
+model_id = "Wan-AI/Wan2.2-Animate-14B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanAnimatePipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+# Load all required inputs for replacement mode
+image = load_image("path/to/new_character.jpg")
+pose_video = load_video("path/to/pose_video.mp4")  # Preprocessed skeletal keypoints
+face_video = load_video("path/to/face_video.mp4")  # Preprocessed facial features
+background_video = load_video("path/to/background_video.mp4")  # Original scene
+mask_video = load_video("path/to/mask_video.mp4")  # Black: preserve, White: generate
+
+# Resize image to match video dimensions
+def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+
+image, height, width = aspect_ratio_resize(image, pipe)
+
+prompt = "A person seamlessly integrated into the scene with consistent lighting and environment"
+negative_prompt = "blurry, low quality, inconsistent lighting, floating, disconnected from scene"
+
+# Replace character in background video
+output = pipe(
+    image=image,
+    pose_video=pose_video,
+    face_video=face_video,
+    background_video=background_video,
+    mask_video=mask_video,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    segment_frame_lengths=77,
+    guidance_scale=1.0,
+    mode="replace",  # Replacement mode
+).frames[0]
+export_to_video(output, "character_replaced.mp4", fps=30)
+```
+
+</hfoption>
+<hfoption id="Advanced options">
+
+```python
+import numpy as np
+import torch
+from diffusers import AutoencoderKLWan, WanAnimatePipeline
+from diffusers.utils import export_to_video, load_image, load_video
+
+model_id = "Wan-AI/Wan2.2-Animate-14B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanAnimatePipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+image = load_image("path/to/character.jpg")
+pose_video = load_video("path/to/pose_video.mp4")
+face_video = load_video("path/to/face_video.mp4")
+
+def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+
+image, height, width = aspect_ratio_resize(image, pipe)
+
+prompt = "A person dancing energetically in a studio"
+negative_prompt = "blurry, low quality"
+
+# Advanced: Use temporal guidance and custom callback
+def callback_fn(pipe, step_index, timestep, callback_kwargs):
+    # You can modify latents or other tensors here
+    print(f"Step {step_index}, Timestep {timestep}")
+    return callback_kwargs
+
+output = pipe(
+    image=image,
+    pose_video=pose_video,
+    face_video=face_video,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=height,
+    width=width,
+    segment_frame_length=77,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+    prev_segment_conditioning_frames=5,  # Use 5 frames for temporal guidance (1 or 5 recommended)
+    callback_on_step_end=callback_fn,
+    callback_on_step_end_tensor_inputs=["latents"],
+).frames[0]
+export_to_video(output, "animated_advanced.mp4", fps=30)
+```
+
+</hfoption>
+</hfoptions>
+
+#### Key Parameters
+
+- **mode**: Choose between `"animate"` (default) or `"replace"`
+- **prev_segment_conditioning_frames**: Number of frames for temporal guidance (1 or 5 recommended). Using 5 provides better temporal consistency but requires more memory
+- **guidance_scale**: Controls how closely the output follows the text prompt. Higher values (5-7) produce results more aligned with the prompt. For Wan-Animate, CFG is disabled by default (`guidance_scale=1.0`) but can be enabled to support negative prompts and finer control over facial expressions. (Note that CFG will only target the text prompt and face conditioning.)
+
+
 ## Notes
 
 - Wan2.1 supports LoRAs with [`~loaders.WanLoraLoaderMixin.load_lora_weights`].
@@ -281,10 +484,10 @@ The general rule of thumb to keep in mind when preparing inputs for the VACE pip
 
   # use "steamboat willie style" to trigger the LoRA
   prompt = """
-  steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot, 
-  revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
-  for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
-  Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
+  steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot,
+  revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in
+  for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground.
+  Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic
   shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
   """
 
@@ -359,6 +562,12 @@ The general rule of thumb to keep in mind when preparing inputs for the VACE pip
   - all
   - __call__
 
+## WanAnimatePipeline
+
+[[autodoc]] WanAnimatePipeline
+  - all
+  - __call__
+
 ## WanPipelineOutput
 
-[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput
+[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput

docs/source/en/modular_diffusers/custom_blocks.md

@@ -0,0 +1,492 @@
+<!--Copyright 2025 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.
+-->
+
+
+# Building Custom Blocks
+
+[ModularPipelineBlocks](./pipeline_block) are the fundamental building blocks of a [`ModularPipeline`]. You can create custom blocks by defining their inputs, outputs, and computation logic. This guide demonstrates how to create and use a custom block.
+
+> [!TIP]
+> Explore the [Modular Diffusers Custom Blocks](https://huggingface.co/collections/diffusers/modular-diffusers-custom-blocks) collection for official custom modular blocks like Nano Banana.
+
+## Project Structure
+
+Your custom block project should use the following structure:
+
+```shell
+.
+├── block.py
+└── modular_config.json
+```
+
+- `block.py` contains the custom block implementation
+- `modular_config.json` contains the metadata needed to load the block
+
+## Example: Florence 2 Inpainting Block
+
+In this example we will create a custom block that uses the [Florence 2](https://huggingface.co/docs/transformers/model_doc/florence2) model to process an input image and generate a mask for inpainting.
+
+The first step is to define the components that the block will use. In this case, we will need to use the `Florence2ForConditionalGeneration` model and its corresponding processor `AutoProcessor`. When defining components, we must specify the name of the component within our pipeline, model class via `type_hint`, and provide a `pretrained_model_name_or_path` for the component if we intend to load the model weights from a specific repository on the Hub.
+
+```py
+# Inside block.py
+from diffusers.modular_pipelines import (
+    ModularPipelineBlocks,
+    ComponentSpec,
+)
+from transformers import AutoProcessor, Florence2ForConditionalGeneration
+
+
+class Florence2ImageAnnotatorBlock(ModularPipelineBlocks):
+
+    @property
+    def expected_components(self):
+        return [
+            ComponentSpec(
+                name="image_annotator",
+                type_hint=Florence2ForConditionalGeneration,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+            ComponentSpec(
+                name="image_annotator_processor",
+                type_hint=AutoProcessor,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+        ]
+```
+
+Next, we define the inputs and outputs of the block. The inputs include the image to be annotated, the annotation task, and the annotation prompt. The outputs include the generated mask image and annotations.
+
+```py
+from typing import List, Union
+from PIL import Image, ImageDraw
+import torch
+import numpy as np
+
+from diffusers.modular_pipelines import (
+    PipelineState,
+    ModularPipelineBlocks,
+    InputParam,
+    ComponentSpec,
+    OutputParam,
+)
+from transformers import AutoProcessor, Florence2ForConditionalGeneration
+
+
+class Florence2ImageAnnotatorBlock(ModularPipelineBlocks):
+
+    @property
+    def expected_components(self):
+        return [
+            ComponentSpec(
+                name="image_annotator",
+                type_hint=Florence2ForConditionalGeneration,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+            ComponentSpec(
+                name="image_annotator_processor",
+                type_hint=AutoProcessor,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+        ]
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam(
+                "image",
+                type_hint=Union[Image.Image, List[Image.Image]],
+                required=True,
+                description="Image(s) to annotate",
+            ),
+            InputParam(
+                "annotation_task",
+                type_hint=Union[str, List[str]],
+                required=True,
+                default="<REFERRING_EXPRESSION_SEGMENTATION>",
+                description="""Annotation Task to perform on the image.
+                Supported Tasks:
+
+                <OD>
+                <REFERRING_EXPRESSION_SEGMENTATION>
+                <CAPTION>
+                <DETAILED_CAPTION>
+                <MORE_DETAILED_CAPTION>
+                <DENSE_REGION_CAPTION>
+                <CAPTION_TO_PHRASE_GROUNDING>
+                <OPEN_VOCABULARY_DETECTION>
+
+                """,
+            ),
+            InputParam(
+                "annotation_prompt",
+                type_hint=Union[str, List[str]],
+                required=True,
+                description="""Annotation Prompt to provide more context to the task.
+                Can be used to detect or segment out specific elements in the image
+                """,
+            ),
+            InputParam(
+                "annotation_output_type",
+                type_hint=str,
+                required=True,
+                default="mask_image",
+                description="""Output type from annotation predictions. Availabe options are
+                mask_image:
+                    -black and white mask image for the given image based on the task type
+                mask_overlay:
+                    - mask overlayed on the original image
+                bounding_box:
+                    - bounding boxes drawn on the original image
+                """,
+            ),
+            InputParam(
+                "annotation_overlay",
+                type_hint=bool,
+                required=True,
+                default=False,
+                description="",
+            ),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam(
+                "mask_image",
+                type_hint=Image,
+                description="Inpainting Mask for input Image(s)",
+            ),
+            OutputParam(
+                "annotations",
+                type_hint=dict,
+                description="Annotations Predictions for input Image(s)",
+            ),
+            OutputParam(
+                "image",
+                type_hint=Image,
+                description="Annotated input Image(s)",
+            ),
+        ]
+
+```
+
+Now we implement the `__call__` method, which contains the logic for processing the input image and generating the mask.
+
+```py
+from typing import List, Union
+from PIL import Image, ImageDraw
+import torch
+import numpy as np
+
+from diffusers.modular_pipelines import (
+    PipelineState,
+    ModularPipelineBlocks,
+    InputParam,
+    ComponentSpec,
+    OutputParam,
+)
+from transformers import AutoProcessor, Florence2ForConditionalGeneration
+
+
+class Florence2ImageAnnotatorBlock(ModularPipelineBlocks):
+
+    @property
+    def expected_components(self):
+        return [
+            ComponentSpec(
+                name="image_annotator",
+                type_hint=Florence2ForConditionalGeneration,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+            ComponentSpec(
+                name="image_annotator_processor",
+                type_hint=AutoProcessor,
+                pretrained_model_name_or_path="florence-community/Florence-2-base-ft",
+            ),
+        ]
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam(
+                "image",
+                type_hint=Union[Image.Image, List[Image.Image]],
+                required=True,
+                description="Image(s) to annotate",
+            ),
+            InputParam(
+                "annotation_task",
+                type_hint=Union[str, List[str]],
+                required=True,
+                default="<REFERRING_EXPRESSION_SEGMENTATION>",
+                description="""Annotation Task to perform on the image.
+                Supported Tasks:
+
+                <OD>
+                <REFERRING_EXPRESSION_SEGMENTATION>
+                <CAPTION>
+                <DETAILED_CAPTION>
+                <MORE_DETAILED_CAPTION>
+                <DENSE_REGION_CAPTION>
+                <CAPTION_TO_PHRASE_GROUNDING>
+                <OPEN_VOCABULARY_DETECTION>
+
+                """,
+            ),
+            InputParam(
+                "annotation_prompt",
+                type_hint=Union[str, List[str]],
+                required=True,
+                description="""Annotation Prompt to provide more context to the task.
+                Can be used to detect or segment out specific elements in the image
+                """,
+            ),
+            InputParam(
+                "annotation_output_type",
+                type_hint=str,
+                required=True,
+                default="mask_image",
+                description="""Output type from annotation predictions. Availabe options are
+                mask_image:
+                    -black and white mask image for the given image based on the task type
+                mask_overlay:
+                    - mask overlayed on the original image
+                bounding_box:
+                    - bounding boxes drawn on the original image
+                """,
+            ),
+            InputParam(
+                "annotation_overlay",
+                type_hint=bool,
+                required=True,
+                default=False,
+                description="",
+            ),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam(
+                "mask_image",
+                type_hint=Image,
+                description="Inpainting Mask for input Image(s)",
+            ),
+            OutputParam(
+                "annotations",
+                type_hint=dict,
+                description="Annotations Predictions for input Image(s)",
+            ),
+            OutputParam(
+                "image",
+                type_hint=Image,
+                description="Annotated input Image(s)",
+            ),
+        ]
+
+    def get_annotations(self, components, images, prompts, task):
+        task_prompts = [task + prompt for prompt in prompts]
+
+        inputs = components.image_annotator_processor(
+            text=task_prompts, images=images, return_tensors="pt"
+        ).to(components.image_annotator.device, components.image_annotator.dtype)
+
+        generated_ids = components.image_annotator.generate(
+            input_ids=inputs["input_ids"],
+            pixel_values=inputs["pixel_values"],
+            max_new_tokens=1024,
+            early_stopping=False,
+            do_sample=False,
+            num_beams=3,
+        )
+        annotations = components.image_annotator_processor.batch_decode(
+            generated_ids, skip_special_tokens=False
+        )
+        outputs = []
+        for image, annotation in zip(images, annotations):
+            outputs.append(
+                components.image_annotator_processor.post_process_generation(
+                    annotation, task=task, image_size=(image.width, image.height)
+                )
+            )
+        return outputs
+
+    def prepare_mask(self, images, annotations, overlay=False, fill="white"):
+        masks = []
+        for image, annotation in zip(images, annotations):
+            mask_image = image.copy() if overlay else Image.new("L", image.size, 0)
+            draw = ImageDraw.Draw(mask_image)
+
+            for _, _annotation in annotation.items():
+                if "polygons" in _annotation:
+                    for polygon in _annotation["polygons"]:
+                        polygon = np.array(polygon).reshape(-1, 2)
+                        if len(polygon) < 3:
+                            continue
+                        polygon = polygon.reshape(-1).tolist()
+                        draw.polygon(polygon, fill=fill)
+
+                elif "bbox" in _annotation:
+                    bbox = _annotation["bbox"]
+                    draw.rectangle(bbox, fill="white")
+
+            masks.append(mask_image)
+
+        return masks
+
+    def prepare_bounding_boxes(self, images, annotations):
+        outputs = []
+        for image, annotation in zip(images, annotations):
+            image_copy = image.copy()
+            draw = ImageDraw.Draw(image_copy)
+            for _, _annotation in annotation.items():
+                bbox = _annotation["bbox"]
+                label = _annotation["label"]
+
+                draw.rectangle(bbox, outline="red", width=3)
+                draw.text((bbox[0], bbox[1] - 20), label, fill="red")
+
+            outputs.append(image_copy)
+
+        return outputs
+
+    def prepare_inputs(self, images, prompts):
+        prompts = prompts or ""
+
+        if isinstance(images, Image.Image):
+            images = [images]
+        if isinstance(prompts, str):
+            prompts = [prompts]
+
+        if len(images) != len(prompts):
+            raise ValueError("Number of images and annotation prompts must match.")
+
+        return images, prompts
+
+    @torch.no_grad()
+    def __call__(self, components, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+        images, annotation_task_prompt = self.prepare_inputs(
+            block_state.image, block_state.annotation_prompt
+        )
+        task = block_state.annotation_task
+        fill = block_state.fill
+
+        annotations = self.get_annotations(
+            components, images, annotation_task_prompt, task
+        )
+        block_state.annotations = annotations
+        if block_state.annotation_output_type == "mask_image":
+            block_state.mask_image = self.prepare_mask(images, annotations)
+        else:
+            block_state.mask_image = None
+
+        if block_state.annotation_output_type == "mask_overlay":
+            block_state.image = self.prepare_mask(images, annotations, overlay=True, fill=fill)
+
+        elif block_state.annotation_output_type == "bounding_box":
+            block_state.image = self.prepare_bounding_boxes(images, annotations)
+
+        self.set_block_state(state, block_state)
+
+        return components, state
+
+```
+
+Once we have defined our custom block, we can save it to the Hub, using either the CLI or the [`push_to_hub`] method. This will make it easy to share and reuse our custom block with other pipelines.
+
+<hfoptions id="share">
+<hfoption id="hf CLI">
+
+```shell
+# In the folder with the `block.py` file, run:
+diffusers-cli custom_block
+```
+
+Then upload the block to the Hub:
+
+```shell
+hf upload <your repo id> . .
+```
+</hfoption>
+<hfoption id="push_to_hub">
+
+```py
+from block import Florence2ImageAnnotatorBlock
+block = Florence2ImageAnnotatorBlock()
+block.push_to_hub("<your repo id>")
+```
+
+</hfoption>
+</hfoptions>
+
+## Using Custom Blocks
+
+Load the custom block with [`~ModularPipelineBlocks.from_pretrained`] and set `trust_remote_code=True`.
+
+```py
+import torch
+from diffusers.modular_pipelines import ModularPipelineBlocks, SequentialPipelineBlocks
+from diffusers.modular_pipelines.stable_diffusion_xl import INPAINT_BLOCKS
+from diffusers.utils import load_image
+
+# Fetch the Florence2 image annotator block that will create our mask
+image_annotator_block = ModularPipelineBlocks.from_pretrained("diffusers/florence-2-custom-block", trust_remote_code=True)
+
+my_blocks = INPAINT_BLOCKS.copy()
+# insert the annotation block before the image encoding step
+my_blocks.insert("image_annotator", image_annotator_block, 1)
+
+# Create our initial set of inpainting blocks
+blocks = SequentialPipelineBlocks.from_blocks_dict(my_blocks)
+
+repo_id = "diffusers/modular-stable-diffusion-xl-base-1.0"
+pipe = blocks.init_pipeline(repo_id)
+pipe.load_components(torch_dtype=torch.float16, device_map="cuda", trust_remote_code=True)
+
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true")
+image = image.resize((1024, 1024))
+
+prompt = ["A red car"]
+annotation_task = "<REFERRING_EXPRESSION_SEGMENTATION>"
+annotation_prompt = ["the car"]
+
+output = pipe(
+    prompt=prompt,
+    image=image,
+    annotation_task=annotation_task,
+    annotation_prompt=annotation_prompt,
+    annotation_output_type="mask_image",
+    num_inference_steps=35,
+    guidance_scale=7.5,
+    strength=0.95,
+    output="images"
+)
+output[0].save("florence-inpainting.png")
+```
+
+## Editing Custom Blocks
+
+By default, custom blocks are saved in your cache directory. Use the `local_dir` argument to download and edit a custom block in a specific folder.
+
+```py
+import torch
+from diffusers.modular_pipelines import ModularPipelineBlocks, SequentialPipelineBlocks
+from diffusers.modular_pipelines.stable_diffusion_xl import INPAINT_BLOCKS
+from diffusers.utils import load_image
+
+# Fetch the Florence2 image annotator block that will create our mask
+image_annotator_block = ModularPipelineBlocks.from_pretrained("diffusers/florence-2-custom-block", trust_remote_code=True, local_dir="/my-local-folder")
+```
+
+Any changes made to the block files in this folder will be reflected when you load the block again.

docs/source/en/modular_diffusers/loop_sequential_pipeline_blocks.md

@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # LoopSequentialPipelineBlocks
 
-[`~modular_pipelines.LoopSequentialPipelineBlocks`] are a multi-block type that composes other [`~modular_pipelines.ModularPipelineBlocks`] together in a loop. Data flows circularly, using `intermediate_inputs` and `intermediate_outputs`, and each block is run iteratively. This is typically used to create a denoising loop which is iterative by default.
+[`~modular_pipelines.LoopSequentialPipelineBlocks`] are a multi-block type that composes other [`~modular_pipelines.ModularPipelineBlocks`] together in a loop. Data flows circularly, using `inputs` and `intermediate_outputs`, and each block is run iteratively. This is typically used to create a denoising loop which is iterative by default.
 
 This guide shows you how to create [`~modular_pipelines.LoopSequentialPipelineBlocks`].
 
@@ -21,7 +21,6 @@ This guide shows you how to create [`~modular_pipelines.LoopSequentialPipelineBl
 [`~modular_pipelines.LoopSequentialPipelineBlocks`], is also known as the *loop wrapper* because it defines the loop structure, iteration variables, and configuration. Within the loop wrapper, you need the following variables.
 
 - `loop_inputs` are user provided values and equivalent to [`~modular_pipelines.ModularPipelineBlocks.inputs`].
-- `loop_intermediate_inputs` are intermediate variables from the [`~modular_pipelines.PipelineState`] and equivalent to [`~modular_pipelines.ModularPipelineBlocks.intermediate_inputs`].
 - `loop_intermediate_outputs` are new intermediate variables created by the block and added to the [`~modular_pipelines.PipelineState`]. It is equivalent to [`~modular_pipelines.ModularPipelineBlocks.intermediate_outputs`].
 - `__call__` method defines the loop structure and iteration logic.
 
@@ -90,4 +89,4 @@ Add more loop blocks to run within each iteration with [`~modular_pipelines.Loop
 
 ```py
 loop = LoopWrapper.from_blocks_dict({"block1": LoopBlock(), "block2": LoopBlock})
-```
+```

docs/source/en/modular_diffusers/pipeline_block.md

@@ -37,17 +37,7 @@ A [`~modular_pipelines.ModularPipelineBlocks`] requires `inputs`, and `intermedi
     ]
     ```
 
-- `intermediate_inputs` are values typically created from a previous block but it can also be directly provided if no preceding block generates them. Unlike `inputs`, `intermediate_inputs` can be modified.
-
-    Use `InputParam` to define `intermediate_inputs`.
-
-    ```py
-    user_intermediate_inputs = [
-        InputParam(name="processed_image", type_hint="torch.Tensor", description="image that has been preprocessed and normalized"),
-    ]
-    ```
-
-- `intermediate_outputs` are new values created by a block and added to the [`~modular_pipelines.PipelineState`]. The `intermediate_outputs` are available as `intermediate_inputs` for subsequent blocks or available as the final output from running the pipeline.
+- `intermediate_outputs` are new values created by a block and added to the [`~modular_pipelines.PipelineState`]. The `intermediate_outputs` are available as `inputs` for subsequent blocks or available as the final output from running the pipeline.
 
     Use `OutputParam` to define `intermediate_outputs`.
 
@@ -65,8 +55,8 @@ The intermediate inputs and outputs share data to connect blocks. They are acces
 
 The computation a block performs is defined in the `__call__` method and it follows a specific structure.
 
-1. Retrieve the [`~modular_pipelines.BlockState`] to get a local view of the `inputs` and `intermediate_inputs`.
-2. Implement the computation logic on the `inputs` and `intermediate_inputs`.
+1. Retrieve the [`~modular_pipelines.BlockState`] to get a local view of the `inputs`
+2. Implement the computation logic on the `inputs`.
 3. Update [`~modular_pipelines.PipelineState`] to push changes from the local [`~modular_pipelines.BlockState`] back to the global [`~modular_pipelines.PipelineState`].
 4. Return the components and state which becomes available to the next block.
 
@@ -76,7 +66,7 @@ def __call__(self, components, state):
     block_state = self.get_block_state(state)
 
     # Your computation logic here
-    # block_state contains all your inputs and intermediate_inputs
+    # block_state contains all your inputs
     # Access them like: block_state.image, block_state.processed_image
 
     # Update the pipeline state with your updated block_states
@@ -112,4 +102,4 @@ def __call__(self, components, state):
     unet = components.unet
     vae = components.vae
     scheduler = components.scheduler
-```
+```

docs/source/en/modular_diffusers/quickstart.md

@@ -183,7 +183,7 @@ from diffusers.modular_pipelines import ComponentsManager
 components = ComponentManager()
 
 dd_pipeline = dd_blocks.init_pipeline("YiYiXu/modular-demo-auto", components_manager=components, collection="diffdiff")
-dd_pipeline.load_default_componenets(torch_dtype=torch.float16)
+dd_pipeline.load_componenets(torch_dtype=torch.float16)
 dd_pipeline.to("cuda")
 ```
 

docs/source/en/modular_diffusers/sequential_pipeline_blocks.md

@@ -12,11 +12,11 @@ specific language governing permissions and limitations under the License.
 
 # SequentialPipelineBlocks
 
-[`~modular_pipelines.SequentialPipelineBlocks`] are a multi-block type that composes other [`~modular_pipelines.ModularPipelineBlocks`] together in a sequence. Data flows linearly from one block to the next using `intermediate_inputs` and `intermediate_outputs`. Each block in [`~modular_pipelines.SequentialPipelineBlocks`] usually represents a step in the pipeline, and by combining them, you gradually build a pipeline.
+[`~modular_pipelines.SequentialPipelineBlocks`] are a multi-block type that composes other [`~modular_pipelines.ModularPipelineBlocks`] together in a sequence. Data flows linearly from one block to the next using `inputs` and `intermediate_outputs`. Each block in [`~modular_pipelines.SequentialPipelineBlocks`] usually represents a step in the pipeline, and by combining them, you gradually build a pipeline.
 
 This guide shows you how to connect two blocks into a [`~modular_pipelines.SequentialPipelineBlocks`].
 
-Create two [`~modular_pipelines.ModularPipelineBlocks`]. The first block, `InputBlock`, outputs a `batch_size` value and the second block, `ImageEncoderBlock` uses `batch_size` as `intermediate_inputs`.
+Create two [`~modular_pipelines.ModularPipelineBlocks`]. The first block, `InputBlock`, outputs a `batch_size` value and the second block, `ImageEncoderBlock` uses `batch_size` as `inputs`.
 
 <hfoptions id="sequential">
 <hfoption id="InputBlock">
@@ -110,4 +110,4 @@ Inspect the sub-blocks in [`~modular_pipelines.SequentialPipelineBlocks`] by cal
 ```py
 print(blocks)
 print(blocks.doc)
-```
+```

docs/source/en/optimization/attention_backends.md

@@ -139,12 +139,14 @@ Refer to the table below for a complete list of available attention backends and
 | `_native_npu` | [PyTorch native](https://docs.pytorch.org/docs/stable/generated/torch.nn.attention.SDPBackend.html#torch.nn.attention.SDPBackend) | NPU-optimized attention |
 | `_native_xla` | [PyTorch native](https://docs.pytorch.org/docs/stable/generated/torch.nn.attention.SDPBackend.html#torch.nn.attention.SDPBackend) | XLA-optimized attention |
 | `flash` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-2 |
+| `flash_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-2 from kernels |
 | `flash_varlen` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention |
 | `aiter` | [AI Tensor Engine for ROCm](https://github.com/ROCm/aiter) | FlashAttention for AMD ROCm |
 | `_flash_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 |
 | `_flash_varlen_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention-3 |
 | `_flash_3_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 from kernels |
 | `sage` | [SageAttention](https://github.com/thu-ml/SageAttention) | Quantized attention (INT8 QK) |
+| `sage_hub` | [SageAttention](https://github.com/thu-ml/SageAttention) | Quantized attention (INT8 QK) from kernels |
 | `sage_varlen` | [SageAttention](https://github.com/thu-ml/SageAttention) | Variable length SageAttention |
 | `_sage_qk_int8_pv_fp8_cuda` | [SageAttention](https://github.com/thu-ml/SageAttention) | INT8 QK + FP8 PV (CUDA) |
 | `_sage_qk_int8_pv_fp8_cuda_sm90` | [SageAttention](https://github.com/thu-ml/SageAttention) | INT8 QK + FP8 PV (SM90) |

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

@@ -0,0 +1,46 @@
+<!--Copyright 2025 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.
+-->
+
+# AutoModel
+
+The [`AutoModel`] class automatically detects and loads the correct model class (UNet, transformer, VAE) from a `config.json` file. You don't need to know the specific model class name ahead of time. It supports data types and device placement, and works across model types and libraries.
+
+The example below loads a transformer from Diffusers and a text encoder from Transformers. Use the `subfolder` parameter to specify where to load the `config.json` file from.
+
+```py
+import torch
+from diffusers import AutoModel, DiffusionPipeline
+
+transformer = AutoModel.from_pretrained(
+    "Qwen/Qwen-Image", subfolder="transformer", torch_dtype=torch.bfloat16, device_map="cuda"
+)
+
+text_encoder = AutoModel.from_pretrained(
+    "Qwen/Qwen-Image", subfolder="text_encoder", torch_dtype=torch.bfloat16, device_map="cuda"
+)
+```
+
+[`AutoModel`] also loads models from the [Hub](https://huggingface.co/models) that aren't included in Diffusers. Set `trust_remote_code=True` in [`AutoModel.from_pretrained`] to load custom models.
+
+```py
+import torch
+from diffusers import AutoModel
+
+transformer = AutoModel.from_pretrained(
+    "custom/custom-transformer-model", trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="cuda"
+)
+```
+
+If the custom model inherits from the [`ModelMixin`] class, it gets access to the same features as Diffusers model classes, like [regional compilation](../optimization/fp16#regional-compilation) and [group offloading](../optimization/memory#group-offloading).
+
+> [!NOTE]
+> Learn more about implementing custom models in the [Community components](../using-diffusers/custom_pipeline_overview#community-components) guide.

docs/source/pt/_toctree.yml

@@ -1,8 +1,10 @@
 - sections:
-    - local: index
-      title: 🧨 Diffusers
-    - local: quicktour
-      title: Tour rápido
-    - local: installation
-      title: Instalação
+  - local: index
+    title: Diffusers
+  - local: installation
+    title: Instalação
+  - local: quicktour
+    title: Tour rápido
+  - local: stable_diffusion
+    title: Desempenho básico
   title: Primeiros passos

docs/source/pt/index.md

@@ -18,11 +18,11 @@ specific language governing permissions and limitations under the License.
 
 # Diffusers
 
-🤗 Diffusers é uma biblioteca de modelos de difusão de última geração para geração de imagens, áudio e até mesmo estruturas 3D de moléculas. Se você está procurando uma solução de geração simples ou queira treinar seu próprio modelo de difusão, 🤗 Diffusers é uma modular caixa de ferramentas que suporta ambos. Nossa biblioteca é desenhada com foco em [usabilidade em vez de desempenho](conceptual/philosophy#usability-over-performance), [simples em vez de fácil](conceptual/philosophy#simple-over-easy) e [customizável em vez de abstrações](conceptual/philosophy#tweakable-contributorfriendly-over-abstraction).
+🤗 Diffusers é uma biblioteca de modelos de difusão de última geração para geração de imagens, áudio e até mesmo estruturas 3D de moléculas. Se você está procurando uma solução de geração simples ou quer treinar seu próprio modelo de difusão, 🤗 Diffusers é uma caixa de ferramentas modular que suporta ambos. Nossa biblioteca é desenhada com foco em [usabilidade em vez de desempenho](conceptual/philosophy#usability-over-performance), [simples em vez de fácil](conceptual/philosophy#simple-over-easy) e [customizável em vez de abstrações](conceptual/philosophy#tweakable-contributorfriendly-over-abstraction).
 
 A Biblioteca tem três componentes principais:
 
-- Pipelines de última geração para a geração em poucas linhas de código. Têm muitos pipelines no 🤗 Diffusers, veja a tabela no pipeline [Visão geral](api/pipelines/overview) para uma lista completa de pipelines disponíveis e as tarefas que eles resolvem.
+- Pipelines de última geração para a geração em poucas linhas de código. Há muitos pipelines no 🤗 Diffusers, veja a tabela no pipeline [Visão geral](api/pipelines/overview) para uma lista completa de pipelines disponíveis e as tarefas que eles resolvem.
 - Intercambiáveis [agendadores de ruído](api/schedulers/overview) para balancear as compensações entre velocidade e qualidade de geração.
 - [Modelos](api/models) pré-treinados que podem ser usados como se fossem blocos de construção, e combinados com agendadores, para criar seu próprio sistema de difusão de ponta a ponta.
 

docs/source/pt/installation.md

@@ -21,7 +21,7 @@ specific language governing permissions and limitations under the License.
 
 Recomenda-se instalar 🤗 Diffusers em um [ambiente virtual](https://docs.python.org/3/library/venv.html).
 Se você não está familiarizado com ambiente virtuals, veja o [guia](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/).
-Um ambiente virtual deixa mais fácil gerenciar diferentes projetos e evitar problemas de compatibilidade entre dependências.
+Um ambiente virtual facilita gerenciar diferentes projetos e evitar problemas de compatibilidade entre dependências.
 
 Comece criando um ambiente virtual no diretório do projeto:
 
@@ -100,12 +100,12 @@ pip install -e ".[flax]"
 </jax>
 </frameworkcontent>
 
-Esses comandos irá linkar a pasta que você clonou o repositório e os caminhos das suas bibliotecas Python.
+Esses comandos irão vincular a pasta que você clonou o repositório e os caminhos das suas bibliotecas Python.
 Python então irá procurar dentro da pasta que você clonou além dos caminhos normais das bibliotecas.
 Por exemplo, se o pacote python for tipicamente instalado no `~/anaconda3/envs/main/lib/python3.10/site-packages/`, o Python também irá procurar na pasta `~/diffusers/` que você clonou.
 
 > [!WARNING]
-> Você deve deixar a pasta `diffusers` se você quiser continuar usando a biblioteca.
+> Você deve manter a pasta `diffusers` se quiser continuar usando a biblioteca.
 
 Agora você pode facilmente atualizar seu clone para a última versão do 🤗 Diffusers com o seguinte comando:
 

docs/source/pt/stable_diffusion.md

@@ -0,0 +1,132 @@
+<!--Copyright 2025 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.
+-->
+
+[[open-in-colab]]
+
+# Desempenho básico
+
+Difusão é um processo aleatório que demanda muito processamento. Você pode precisar executar o [`DiffusionPipeline`] várias vezes antes de obter o resultado desejado. Por isso é importante equilibrar cuidadosamente a velocidade de geração e o uso de memória para iterar mais rápido.
+
+Este guia recomenda algumas dicas básicas de desempenho para usar o [`DiffusionPipeline`]. Consulte a seção de documentação sobre Otimização de Inferência, como [Acelerar inferência](./optimization/fp16) ou [Reduzir uso de memória](./optimization/memory) para guias de desempenho mais detalhados.
+
+## Uso de memória
+
+Reduzir a quantidade de memória usada indiretamente acelera a geração e pode ajudar um modelo a caber no dispositivo.
+
+O método [`~DiffusionPipeline.enable_model_cpu_offload`] move um modelo para a CPU quando não está em uso para economizar memória da GPU.
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+
+pipeline = DiffusionPipeline.from_pretrained(
+  "stabilityai/stable-diffusion-xl-base-1.0",
+  torch_dtype=torch.bfloat16,
+  device_map="cuda"
+)
+pipeline.enable_model_cpu_offload()
+
+prompt = """
+cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
+highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
+"""
+pipeline(prompt).images[0]
+print(f"Memória máxima reservada: {torch.cuda.max_memory_allocated() / 1024**3:.2f} GB")
+```
+
+## Velocidade de inferência
+
+O processo de remoção de ruído é o mais exigente computacionalmente durante a difusão. Métodos que otimizam este processo aceleram a velocidade de inferência. Experimente os seguintes métodos para acelerar.
+
+- Adicione `device_map="cuda"` para colocar o pipeline em uma GPU. Colocar um modelo em um acelerador, como uma GPU, aumenta a velocidade porque realiza computações em paralelo.
+- Defina `torch_dtype=torch.bfloat16` para executar o pipeline em meia-precisão. Reduzir a precisão do tipo de dado aumenta a velocidade porque leva menos tempo para realizar computações em precisão mais baixa.
+
+```py
+import torch
+import time
+from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
+
+pipeline = DiffusionPipeline.from_pretrained(
+  "stabilityai/stable-diffusion-xl-base-1.0",
+  torch_dtype=torch.bfloat16,
+  device_map="cuda"
+)
+```
+
+- Use um agendador mais rápido, como [`DPMSolverMultistepScheduler`], que requer apenas ~20-25 passos.
+- Defina `num_inference_steps` para um valor menor. Reduzir o número de passos de inferência reduz o número total de computações. No entanto, isso pode resultar em menor qualidade de geração.
+
+```py
+pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
+
+prompt = """
+cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
+highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
+"""
+
+start_time = time.perf_counter()
+image = pipeline(prompt).images[0]
+end_time = time.perf_counter()
+
+print(f"Geração de imagem levou {end_time - start_time:.3f} segundos")
+```
+
+## Qualidade de geração
+
+Muitos modelos de difusão modernos entregam imagens de alta qualidade imediatamente. No entanto, você ainda pode melhorar a qualidade de geração experimentando o seguinte.
+
+- Experimente um prompt mais detalhado e descritivo. Inclua detalhes como o meio da imagem, assunto, estilo e estética. Um prompt negativo também pode ajudar, guiando um modelo para longe de características indesejáveis usando palavras como baixa qualidade ou desfocado.
+
+    ```py
+    import torch
+    from diffusers import DiffusionPipeline
+
+    pipeline = DiffusionPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda"
+    )
+
+    prompt = """
+    cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
+    highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
+    """
+    negative_prompt = "low quality, blurry, ugly, poor details"
+    pipeline(prompt, negative_prompt=negative_prompt).images[0]
+    ```
+
+    Para mais detalhes sobre como criar prompts melhores, consulte a documentação sobre [Técnicas de prompt](./using-diffusers/weighted_prompts).
+
+- Experimente um agendador diferente, como [`HeunDiscreteScheduler`] ou [`LMSDiscreteScheduler`], que sacrifica velocidade de geração por qualidade.
+
+    ```py
+    import torch
+    from diffusers import DiffusionPipeline, HeunDiscreteScheduler
+
+    pipeline = DiffusionPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        torch_dtype=torch.bfloat16,
+        device_map="cuda"
+    )
+    pipeline.scheduler = HeunDiscreteScheduler.from_config(pipeline.scheduler.config)
+
+    prompt = """
+    cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
+    highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
+    """
+    negative_prompt = "low quality, blurry, ugly, poor details"
+    pipeline(prompt, negative_prompt=negative_prompt).images[0]
+    ```
+
+## Próximos passos
+
+Diffusers oferece otimizações mais avançadas e poderosas, como [group-offloading](./optimization/memory#group-offloading) e [compilação regional](./optimization/fp16#regional-compilation). Para saber mais sobre como maximizar o desempenho, consulte a seção sobre Otimização de Inferência.

examples/community/README.md

@@ -88,7 +88,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | FaithDiff Stable Diffusion XL Pipeline | Implementation of [(CVPR 2025) FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolutionUnleashing Diffusion Priors for Faithful Image Super-resolution](https://huggingface.co/papers/2411.18824) - FaithDiff is a faithful image super-resolution method that leverages latent diffusion models by actively adapting the diffusion prior and jointly fine-tuning its components (encoder and diffusion model) with an alignment module to ensure high fidelity and structural consistency. | [FaithDiff Stable Diffusion XL Pipeline](#faithdiff-stable-diffusion-xl-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/jychen9811/FaithDiff) | [Junyang Chen, Jinshan Pan, Jiangxin Dong, IMAG Lab, (Adapted by Eliseu Silva)](https://github.com/JyChen9811/FaithDiff) |
 | Stable Diffusion 3 InstructPix2Pix Pipeline | Implementation of Stable Diffusion 3 InstructPix2Pix Pipeline | [Stable Diffusion 3 InstructPix2Pix Pipeline](#stable-diffusion-3-instructpix2pix-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/BleachNick/SD3_UltraEdit_freeform) [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/CaptainZZZ/sd3-instructpix2pix) | [Jiayu Zhang](https://github.com/xduzhangjiayu) and [Haozhe Zhao](https://github.com/HaozheZhao)|
 | Flux Kontext multiple images | A modified version of the `FluxKontextPipeline` that supports calling Flux Kontext with multiple reference images.| [Flux Kontext multiple input Pipeline](#flux-kontext-multiple-images) | - |  [Net-Mist](https://github.com/Net-Mist) |
-
+| Flux Fill ControlNet Pipeline | A modified version of the `FluxFillPipeline` and `FluxControlNetInpaintPipeline` that supports Controlnet with Flux Fill model.| [Flux Fill ControlNet Pipeline](#Flux-Fill-ControlNet-Pipeline) | - |  [pratim4dasude](https://github.com/pratim4dasude) |
 
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
 
@@ -5488,7 +5488,7 @@ Editing at Scale", many thanks to their contribution!
 
 This implementation of Flux Kontext allows users to pass multiple reference images. Each image is encoded separately, and the resulting latent vectors are concatenated.
 
-As explained in Section 3 of [the paper](https://arxiv.org/pdf/2506.15742), the model's sequence concatenation mechanism can extend its capabilities to handle multiple reference images. However, note that the current version of Flux Kontext was not trained for this use case. In practice, stacking along the first axis does not yield correct results, while stacking along the other two axes appears to work.
+As explained in Section 3 of [the paper](https://huggingface.co/papers/2506.15742), the model's sequence concatenation mechanism can extend its capabilities to handle multiple reference images. However, note that the current version of Flux Kontext was not trained for this use case. In practice, stacking along the first axis does not yield correct results, while stacking along the other two axes appears to work.
 
 ## Example Usage
 
@@ -5527,3 +5527,106 @@ images = pipe(
 ).images
 images[0].save("pizzeria.png")
 ```
+
+# Flux Fill ControlNet Pipeline
+
+This implementation of Flux Fill + ControlNet Inpaint combines the fill-style masked editing of FLUX.1-Fill-dev with full ControlNet conditioning. The base image is processed through the Fill model while the ControlNet receives the corresponding conditioning input (depth, canny, pose, etc.), and both outputs are fused during denoising to guide structure and composition.
+
+While FLUX.1-Fill-dev is designed for mask-based edits, it was not originally trained to operate jointly with ControlNet. In practice, this combined setup works well for structured inpainting tasks, though results may vary depending on the conditioning strength and the alignment between the mask and the control input.
+
+## Example Usage
+
+
+```python
+import torch
+from diffusers import (
+    FluxControlNetModel,
+    FluxPriorReduxPipeline,
+)
+from diffusers.utils import load_image
+
+# NEW PIPELINE (updated name)
+from pipline_flux_fill_controlnet_Inpaint import  FluxControlNetFillInpaintPipeline
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.bfloat16
+
+# Models
+base_model = "black-forest-labs/FLUX.1-Fill-dev"
+controlnet_model = "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0"
+prior_model = "black-forest-labs/FLUX.1-Redux-dev"
+
+# Load ControlNet
+controlnet = FluxControlNetModel.from_pretrained(
+    controlnet_model,
+    torch_dtype=dtype,
+)
+
+# Load Fill + ControlNet Pipeline
+fill_pipe = FluxControlNetFillInpaintPipeline.from_pretrained(
+    base_model,
+    controlnet=controlnet,
+    torch_dtype=dtype,
+).to(device)
+
+# OPTIONAL FP8
+# fill_pipe.transformer.enable_layerwise_casting(
+#     storage_dtype=torch.float8_e4m3fn,
+#     compute_dtype=torch.bfloat16
+# )
+
+#  OPTIONAL Prior Redux
+#pipe_prior_redux = FluxPriorReduxPipeline.from_pretrained(
+#    prior_model,
+#    torch_dtype=dtype,
+#).to(device)
+
+# Inputs
+
+# combined_image = load_image("person_input.png")
+
+
+# 1. Prior conditioning
+#prior_out = pipe_prior_redux(
+#    image=cloth_image,
+#    prompt=cloth_prompt,
+#)
+
+# 2. Fill Inpaint with ControlNet
+
+# canny (0), tile (1), depth (2), blur (3), pose (4), gray (5), low quality (6).
+
+img = load_image(r"imgs/background.jpg")
+mask = load_image(r"imgs/mask.png")
+
+control_image_depth = load_image(r"imgs/dog_depth _2.png")
+
+result = fill_pipe(
+    prompt="a dog on a bench",
+    image=img,
+    mask_image=mask,
+
+    control_image=control_image_depth,
+    control_mode=[2],  # union mode
+    control_guidance_start=0.0,
+    control_guidance_end=0.8,
+    controlnet_conditioning_scale=0.9,
+
+    height=1024,
+    width=1024,
+
+    strength=1.0,
+    guidance_scale=50.0,
+    num_inference_steps=60,
+    max_sequence_length=512,
+
+#    **prior_out,
+)
+
+# result.images[0].save("flux_fill_controlnet_inpaint.png")
+
+from datetime import datetime
+timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+result.images[0].save(f"flux_fill_controlnet_inpaint_depth{timestamp}.jpg")
+```
+

examples/community/img2img_inpainting.py

@@ -45,7 +45,7 @@ def check_size(image, height, width):
         raise ValueError(f"Image size should be {height}x{width}, but got {h}x{w}")
 
 
-def overlay_inner_image(image, inner_image, paste_offset: Tuple[int] = (0, 0)):
+def overlay_inner_image(image, inner_image, paste_offset: Tuple[int, ...] = (0, 0)):
     inner_image = inner_image.convert("RGBA")
     image = image.convert("RGB")
 

examples/community/matryoshka.py

@@ -1966,16 +1966,21 @@ class MatryoshkaUNet2DConditionModel(
         center_input_sample: bool = False,
         flip_sin_to_cos: bool = True,
         freq_shift: int = 0,
-        down_block_types: Tuple[str] = (
+        down_block_types: Tuple[str, ...] = (
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "DownBlock2D",
         ),
         mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
-        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        up_block_types: Tuple[str, ...] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
         only_cross_attention: Union[bool, Tuple[bool]] = False,
-        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
         layers_per_block: Union[int, Tuple[int]] = 2,
         downsample_padding: int = 1,
         mid_block_scale_factor: float = 1,
@@ -2294,10 +2299,10 @@ class MatryoshkaUNet2DConditionModel(
 
     def _check_config(
         self,
-        down_block_types: Tuple[str],
-        up_block_types: Tuple[str],
+        down_block_types: Tuple[str, ...],
+        up_block_types: Tuple[str, ...],
         only_cross_attention: Union[bool, Tuple[bool]],
-        block_out_channels: Tuple[int],
+        block_out_channels: Tuple[int, ...],
         layers_per_block: Union[int, Tuple[int]],
         cross_attention_dim: Union[int, Tuple[int]],
         transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],

examples/community/pipeline_faithdiff_stable_diffusion_xl.py

@@ -438,16 +438,21 @@ class UNet2DConditionModel(OriginalUNet2DConditionModel, ConfigMixin, UNet2DCond
         center_input_sample: bool = False,
         flip_sin_to_cos: bool = True,
         freq_shift: int = 0,
-        down_block_types: Tuple[str] = (
+        down_block_types: Tuple[str, ...] = (
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "DownBlock2D",
         ),
         mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
-        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        up_block_types: Tuple[str, ...] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
         only_cross_attention: Union[bool, Tuple[bool]] = False,
-        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
         layers_per_block: Union[int, Tuple[int]] = 2,
         downsample_padding: int = 1,
         mid_block_scale_factor: float = 1,

examples/community/regional_prompting_stable_diffusion.py

@@ -490,7 +490,7 @@ class RegionalPromptingStableDiffusionPipeline(
     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
+        # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
         # and should be between [0, 1]
 
         accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
@@ -841,7 +841,7 @@ class RegionalPromptingStableDiffusionPipeline(
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only applies
                 to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
@@ -872,7 +872,7 @@ class RegionalPromptingStableDiffusionPipeline(
                 [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
             guidance_rescale (`float`, *optional*, defaults to 0.0):
                 Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
-                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+                Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when
                 using zero terminal SNR.
             clip_skip (`int`, *optional*):
                 Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
@@ -1062,7 +1062,7 @@ class RegionalPromptingStableDiffusionPipeline(
                     noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
 
                 if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
-                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    # Based on 3.4. in https://huggingface.co/papers/2305.08891
                     noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
 
                 # compute the previous noisy sample x_t -> x_t-1
@@ -1668,7 +1668,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     r"""
     Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
     Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
-    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+    Flawed](https://huggingface.co/papers/2305.08891).
 
     Args:
         noise_cfg (`torch.Tensor`):

examples/dreambooth/README_flux.md

@@ -268,12 +268,11 @@ provide a simple script for LoRA fine-tuning Kontext in [train_dreambooth_lora_f
 **important**
 
 > [!NOTE] 
-> To make sure you can successfully run the latest version of the kontext example script, we highly recommend installing from source, specifically from the commit mentioned below.
+> To make sure you can successfully run the latest version of the kontext example script, we highly recommend installing from source.
 > To do this, execute the following steps in a new virtual environment:
 > ```
 > git clone https://github.com/huggingface/diffusers
 > cd diffusers
-> git checkout 05e7a854d0a5661f5b433f6dd5954c224b104f0b
 > pip install -e .
 > ```
 

examples/dreambooth/README_flux2.md

@@ -0,0 +1,315 @@
+# DreamBooth training example for FLUX.2 [dev]
+
+[DreamBooth](https://huggingface.co/papers/2208.12242) is a method to personalize image generation models given just a few (3~5) images of a subject/concept.
+
+The `train_dreambooth_lora_flux2.py` script shows how to implement the training procedure for [LoRAs](https://huggingface.co/blog/lora) and adapt it for [FLUX.2 [dev]](https://github.com/black-forest-labs/flux2).
+
+> [!NOTE]
+> **Memory consumption**
+>
+> Flux can be quite expensive to run on consumer hardware devices and as a result finetuning it comes with high memory requirements -
+> a LoRA with a rank of 16 can exceed XXGB of VRAM for training. below we provide some tips and tricks to reduce memory consumption during training.
+
+> For more tips & guidance on training on a resource-constrained device and general good practices please check out these great guides and trainers for FLUX: 
+> 1) [`@bghira`'s guide](https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX2.md)
+> 2) [`ostris`'s guide](https://github.com/ostris/ai-toolkit?tab=readme-ov-file#flux2-training)
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [FLUX.2 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.2-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+hf auth login
+```
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+## Running locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the `examples/dreambooth` folder and run
+```bash
+pip install -r requirements_flux.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell (e.g., a notebook)
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
+
+
+### Dog toy example
+
+Now let's get our dataset. For this example we will use some dog images: https://huggingface.co/datasets/diffusers/dog-example.
+
+Let's first download it locally:
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = "./dog"
+snapshot_download(
+    "diffusers/dog-example",
+    local_dir=local_dir, repo_type="dataset",
+    ignore_patterns=".gitattributes",
+)
+```
+
+This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
+
+As mentioned, Flux2 LoRA training is *very* memory intensive. Here are memory optimizations we can use (some still experimental) for a more memory efficient training:
+
+## Memory Optimizations
+> [!NOTE] many of these techniques complement each other and can be used together to further reduce memory consumption. 
+> However some techniques may be mutually exclusive so be sure to check before launching a training run.
+### Remote Text Encoder 
+Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--remote_text_encoder` flag to enable remote computation of the prompt embeddings using the HuggingFace Inference API. 
+This way, the text encoder model is not loaded into memory during training.
+> [!NOTE] 
+> to enable remote text encoding you must either be logged in to your HuggingFace account (`hf auth login`) OR pass a token with `--hub_token`.
+### CPU Offloading 
+To offload parts of the model to CPU memory, you can use `--offload` flag. This will offload the vae and text encoder to CPU memory and only move them to GPU when needed.
+### Latent Caching 
+Pre-encode the training images with the vae, and then delete it to free up some memory. To enable `latent_caching` simply pass `--cache_latents`.
+### QLoRA: Low Precision Training with Quantization
+Perform low precision training using 8-bit or 4-bit quantization to reduce memory usage. You can use the following flags:
+- **FP8 training** with `torchao`: 
+enable FP8 training by passing `--do_fp8_training`. 
+> [!IMPORTANT] Since we are utilizing FP8 tensor cores we need CUDA GPUs with compute capability at least 8.9 or greater. 
+> If you're looking for memory-efficient training on relatively older cards, we encourage you to check out other trainers like SimpleTuner, ai-toolkit, etc.
+- **NF4 training** with `bitsandbytes`: 
+Alternatively, you can use 8-bit or 4-bit quantization with `bitsandbytes` by passing:
+`--bnb_quantization_config_path` to enable 4-bit NF4 quantization.
+### Gradient Checkpointing and Accumulation
+* `--gradient accumulation` refers to the number of updates steps to accumulate before performing a backward/update pass.
+by passing a value > 1 you can reduce the amount of backward/update passes and hence also memory reqs.
+* with `--gradient checkpointing` we can save memory by not storing all intermediate activations during the forward pass.
+Instead, only a subset of these activations (the checkpoints) are stored and the rest is recomputed as needed during the backward pass. Note that this comes at the expanse of a slower backward pass.
+### 8-bit-Adam Optimizer
+When training with `AdamW`(doesn't apply to `prodigy`) You can pass `--use_8bit_adam` to reduce the memory requirements of training. 
+Make sure to install `bitsandbytes` if you want to do so.
+### Image Resolution
+An easy way to mitigate some of the memory requirements is through `--resolution`. `--resolution` refers to the resolution for input images, all the images in the train/validation dataset are resized to this.
+Note that by default, images are resized to resolution of 512, but it's good to keep in mind in case you're accustomed to training on higher resolutions.
+### Precision of saved LoRA layers
+By default, trained transformer layers are saved in the precision dtype in which training was performed. E.g. when training in mixed precision is enabled with `--mixed_precision="bf16"`, final finetuned layers will be saved in `torch.bfloat16` as well. 
+This reduces memory requirements significantly w/o a significant quality loss. Note that if you do wish to save the final layers in float32 at the expanse of more memory usage, you can do so by passing `--upcast_before_saving`.
+
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.2-dev"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux2"
+
+accelerate launch train_dreambooth_lora_flux2.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --do_fp8_training \
+  --gradient_checkpointing \
+  --remote_text_encoder \
+  --cache_latents \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --guidance_scale=1 \
+  --use_8bit_adam \
+  --gradient_accumulation_steps=4 \
+  --optimizer="adamW" \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=100 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+To better track our training experiments, we're using the following flags in the command above:
+
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
+* `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
+
+> [!NOTE]
+> If you want to train using long prompts with the T5 text encoder, you can use `--max_sequence_length` to set the token limit. The default is 77, but it can be increased to as high as 512. Note that this will use more resources and may slow down the training in some cases.
+
+## LoRA + DreamBooth
+
+[LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.
+
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
+
+### Prodigy Optimizer
+Prodigy is an adaptive optimizer that dynamically adjusts the learning rate learned parameters based on past gradients, allowing for more efficient convergence. 
+By using prodigy we can "eliminate" the need for manual learning rate tuning. read more [here](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers).
+
+to use prodigy, first make sure to install the prodigyopt library: `pip install prodigyopt`, and then specify -
+```bash
+--optimizer="prodigy"
+```
+> [!TIP]
+> When using prodigy it's generally good practice to set- `--learning_rate=1.0`
+
+To perform DreamBooth with LoRA, run:
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.2-dev"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux2-lora"
+
+accelerate launch train_dreambooth_lora_flux2.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --do_fp8_training \
+  --gradient_checkpointing \
+  --remote_text_encoder \
+  --cache_latents \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --guidance_scale=1 \
+  --gradient_accumulation_steps=4 \
+  --optimizer="prodigy" \
+  --learning_rate=1. \
+  --report_to="wandb" \
+  --lr_scheduler="constant_with_warmup" \
+  --lr_warmup_steps=100 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+### LoRA Rank and Alpha
+Two key LoRA hyperparameters are LoRA rank and LoRA alpha. 
+- `--rank`: Defines the dimension of the trainable LoRA matrices. A higher rank means more expressiveness and capacity to learn (and more parameters).
+- `--lora_alpha`: A scaling factor for the LoRA's output. The LoRA update is scaled by lora_alpha / lora_rank.
+- lora_alpha vs. rank:
+This ratio dictates the LoRA's effective strength:
+lora_alpha == rank: Scaling factor is 1. The LoRA is applied with its learned strength. (e.g., alpha=16, rank=16)
+lora_alpha < rank: Scaling factor < 1. Reduces the LoRA's impact. Useful for subtle changes or to prevent overpowering the base model. (e.g., alpha=8, rank=16)
+lora_alpha > rank: Scaling factor > 1. Amplifies the LoRA's impact. Allows a lower rank LoRA to have a stronger effect. (e.g., alpha=32, rank=16)
+
+> [!TIP]
+> A common starting point is to set `lora_alpha` equal to `rank`. 
+> Some also set `lora_alpha` to be twice the `rank` (e.g., lora_alpha=32 for lora_rank=16) 
+> to give the LoRA updates more influence without increasing parameter count. 
+> If you find your LoRA is "overcooking" or learning too aggressively, consider setting `lora_alpha` to half of `rank` 
+> (e.g., lora_alpha=8 for rank=16). Experimentation is often key to finding the optimal balance for your use case.
+
+### Target Modules
+When LoRA was first adapted from language models to diffusion models, it was applied to the cross-attention layers in the Unet that relate the image representations with the prompts that describe them. 
+More recently, SOTA text-to-image diffusion models replaced the Unet with a diffusion Transformer(DiT). With this change, we may also want to explore 
+applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma separated string
+the exact modules for LoRA training. Here are some examples of target modules you can provide: 
+- for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0"`
+- to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2"`
+- to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2,norm1_context.linear, norm1.linear,norm.linear,proj_mlp,proj_out"`
+> [!NOTE]
+> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma separated string:
+> **single DiT blocks**: to target the ith single transformer block, add the prefix `single_transformer_blocks.i`, e.g. - `single_transformer_blocks.i.attn.to_k`
+> **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k` 
+> [!NOTE]
+> keep in mind that while training more layers can improve quality and expressiveness, it also increases the size of the output LoRA weights.
+
+
+
+## Training Image-to-Image
+
+Flux.2 lets us perform image editing as well as image generation. We provide a simple script for image-to-image(I2I) LoRA fine-tuning in [train_dreambooth_lora_flux2_img2img.py](./train_dreambooth_lora_flux2_img2img.py) for both T2I and I2I. The optimizations discussed above apply this script, too.
+
+**important**
+
+**Important**
+To make sure you can successfully run the latest version of the image-to-image example script, we highly recommend installing from source, specifically from the commit mentioned below. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+
+To start, you must have a dataset containing triplets:
+
+* Condition image - the input image to be transformed.
+* Target image - the desired output image after transformation.
+* Instruction - a text prompt describing the transformation from the condition image to the target image.
+
+[kontext-community/relighting](https://huggingface.co/datasets/kontext-community/relighting) is a good example of such a dataset. If you are using such a dataset, you can use the command below to launch training:
+
+```bash
+accelerate launch train_dreambooth_lora_flux2_img2img.py \
+  --pretrained_model_name_or_path=black-forest-labs/FLUX.2-dev  \
+  --output_dir="flux2-i2i" \
+  --dataset_name="kontext-community/relighting" \
+  --image_column="output" --cond_image_column="file_name" --caption_column="instruction" \
+  --do_fp8_training \
+  --gradient_checkpointing \
+  --remote_text_encoder \
+  --cache_latents \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --guidance_scale=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --optimizer="adamw" \
+  --use_8bit_adam \
+  --cache_latents \
+  --learning_rate=1e-4 \
+  --lr_scheduler="constant_with_warmup" \
+  --lr_warmup_steps=200 \
+  --max_train_steps=1000 \
+  --rank=16\
+  --seed="0" 
+```
+
+More generally, when performing I2I fine-tuning, we expect you to:
+
+* Have a dataset `kontext-community/relighting`
+* Supply `image_column`, `cond_image_column`, and `caption_column` values when launching training
+
+### Misc notes
+
+* By default, we use `mode` as the value of `--vae_encode_mode` argument. This is because Kontext uses `mode()` of the distribution predicted by the VAE instead of sampling from it.
+### Aspect Ratio Bucketing
+we've added aspect ratio bucketing support which allows training on images with different aspect ratios without cropping them to a single square resolution. This technique helps preserve the original composition of training images and can improve training efficiency.
+
+To enable aspect ratio bucketing, pass `--aspect_ratio_buckets` argument with a semicolon-separated list of height,width pairs, such as:
+
+`--aspect_ratio_buckets="672,1568;688,1504;720,1456;752,1392;800,1328;832,1248;880,1184;944,1104;1024,1024;1104,944;1184,880;1248,832;1328,800;1392,752;1456,720;1504,688;1568,672"
+`
+Since Flux.2 finetuning is still an experimental phase, we encourage you to explore different settings and share your insights! 🤗

examples/dreambooth/test_dreambooth_lora_flux2.py

@@ -0,0 +1,262 @@
+# coding=utf-8
+# Copyright 2025 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 json
+import logging
+import os
+import sys
+import tempfile
+
+import safetensors
+
+from diffusers.loaders.lora_base import LORA_ADAPTER_METADATA_KEY
+
+
+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 DreamBoothLoRAFlux2(ExamplesTestsAccelerate):
+    instance_data_dir = "docs/source/en/imgs"
+    instance_prompt = "dog"
+    pretrained_model_name_or_path = "hf-internal-testing/tiny-flux2"
+    script_path = "examples/dreambooth/train_dreambooth_lora_flux2.py"
+    transformer_layer_type = "single_transformer_blocks.0.attn.to_qkv_mlp_proj"
+
+    def test_dreambooth_lora_flux2(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --max_sequence_length 8
+                --text_encoder_out_layers 1
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_latent_caching(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --cache_latents
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --max_sequence_length 8
+                --text_encoder_out_layers 1
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_layers(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --cache_latents
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lora_layers {self.transformer_layer_type}
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --max_sequence_length 8
+                --text_encoder_out_layers 1
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names. In this test, we only params of
+            # transformer.single_transformer_blocks.0.attn.to_k should be in the state dict
+            starts_with_transformer = all(
+                key.startswith(f"transformer.{self.transformer_layer_type}") for key in lora_state_dict.keys()
+            )
+            self.assertTrue(starts_with_transformer)
+
+    def test_dreambooth_lora_flux2_checkpointing_checkpoints_total_limit(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=6
+            --checkpoints_total_limit=2
+            --max_sequence_length 8
+            --checkpointing_steps=2
+            --text_encoder_out_layers 1
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual(
+                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
+                {"checkpoint-4", "checkpoint-6"},
+            )
+
+    def test_dreambooth_lora_flux2_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --checkpointing_steps=2
+            --max_sequence_length 8
+            --text_encoder_out_layers 1
+            """.split()
+
+            run_command(self._launch_args + test_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-2", "checkpoint-4"})
+
+            resume_run_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --instance_prompt={self.instance_prompt}
+            --resolution=64
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=8
+            --checkpointing_steps=2
+            --resume_from_checkpoint=checkpoint-4
+            --checkpoints_total_limit=2
+            --max_sequence_length 8
+            --text_encoder_out_layers 1
+            """.split()
+
+            run_command(self._launch_args + resume_run_args)
+
+            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8"})
+
+    def test_dreambooth_lora_with_metadata(self):
+        # Use a `lora_alpha` that is different from `rank`.
+        lora_alpha = 8
+        rank = 4
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --lora_alpha={lora_alpha}
+                --rank={rank}
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --max_sequence_length 8
+                --text_encoder_out_layers 1
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            state_dict_file = os.path.join(tmpdir, "pytorch_lora_weights.safetensors")
+            self.assertTrue(os.path.isfile(state_dict_file))
+
+            # Check if the metadata was properly serialized.
+            with safetensors.torch.safe_open(state_dict_file, framework="pt", device="cpu") as f:
+                metadata = f.metadata() or {}
+
+            metadata.pop("format", None)
+            raw = metadata.get(LORA_ADAPTER_METADATA_KEY)
+            if raw:
+                raw = json.loads(raw)
+
+            loaded_lora_alpha = raw["transformer.lora_alpha"]
+            self.assertTrue(loaded_lora_alpha == lora_alpha)
+            loaded_lora_rank = raw["transformer.r"]
+            self.assertTrue(loaded_lora_rank == rank)

examples/unconditional_image_generation/README.md

@@ -104,6 +104,8 @@ To use your own dataset, there are 2 ways:
 - you can either provide your own folder as `--train_data_dir`
 - or you can upload your dataset to the hub (possibly as a private repo, if you prefer so), and simply pass the `--dataset_name` argument.
 
+If your dataset contains 16 or 32-bit channels (for example, medical TIFFs), add the `--preserve_input_precision` flag so the preprocessing keeps the original precision while still training a 3-channel model. Precision still depends on the decoder: Pillow keeps 16-bit grayscale and float inputs, but many 16-bit RGB files are decoded as 8-bit RGB, and the flag cannot recover precision lost at load time.
+
 Below, we explain both in more detail.
 
 #### Provide the dataset as a folder

examples/unconditional_image_generation/train_unconditional.py

@@ -52,6 +52,24 @@ def _extract_into_tensor(arr, timesteps, broadcast_shape):
     return res.expand(broadcast_shape)
 
 
+def _ensure_three_channels(tensor: torch.Tensor) -> torch.Tensor:
+    """
+    Ensure the tensor has exactly three channels (C, H, W) by repeating or truncating channels when needed.
+    """
+    if tensor.ndim == 2:
+        tensor = tensor.unsqueeze(0)
+    channels = tensor.shape[0]
+    if channels == 3:
+        return tensor
+    if channels == 1:
+        return tensor.repeat(3, 1, 1)
+    if channels == 2:
+        return torch.cat([tensor, tensor[:1]], dim=0)
+    if channels > 3:
+        return tensor[:3]
+    raise ValueError(f"Unsupported number of channels: {channels}")
+
+
 def parse_args():
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
     parser.add_argument(
@@ -260,6 +278,11 @@ def parse_args():
     parser.add_argument(
         "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
     )
+    parser.add_argument(
+        "--preserve_input_precision",
+        action="store_true",
+        help="Preserve 16/32-bit image precision by avoiding 8-bit RGB conversion while still producing 3-channel tensors.",
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -453,19 +476,41 @@ def main(args):
         # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
 
     # Preprocessing the datasets and DataLoaders creation.
+    spatial_augmentations = [
+        transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+        transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
+        transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
+    ]
+
     augmentations = transforms.Compose(
-        [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
-            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
-            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
+        spatial_augmentations
+        + [
             transforms.ToTensor(),
             transforms.Normalize([0.5], [0.5]),
         ]
     )
 
+    precision_augmentations = transforms.Compose(
+        [
+            transforms.PILToTensor(),
+            transforms.Lambda(_ensure_three_channels),
+            transforms.ConvertImageDtype(torch.float32),
+        ]
+        + spatial_augmentations
+        + [transforms.Normalize([0.5], [0.5])]
+    )
+
     def transform_images(examples):
-        images = [augmentations(image.convert("RGB")) for image in examples["image"]]
-        return {"input": images}
+        processed = []
+        for image in examples["image"]:
+            if not args.preserve_input_precision:
+                processed.append(augmentations(image.convert("RGB")))
+            else:
+                precise_image = image
+                if precise_image.mode == "P":
+                    precise_image = precise_image.convert("RGB")
+                processed.append(precision_augmentations(precise_image))
+        return {"input": processed}
 
     logger.info(f"Dataset size: {len(dataset)}")
 

scripts/convert_flux2_to_diffusers.py

@@ -0,0 +1,475 @@
+import argparse
+from contextlib import nullcontext
+from typing import Any, Dict, Tuple
+
+import safetensors.torch
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download
+from transformers import AutoProcessor, GenerationConfig, Mistral3ForConditionalGeneration
+
+from diffusers import AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler, Flux2Pipeline, Flux2Transformer2DModel
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+"""
+# VAE
+
+python scripts/convert_flux2_to_diffusers.py  \
+--original_state_dict_repo_id "diffusers-internal-dev/new-model-image" \
+--vae_filename "flux2-vae.sft" \
+--output_path "/raid/yiyi/dummy-flux2-diffusers" \
+--vae
+
+# DiT
+
+python scripts/convert_flux2_to_diffusers.py \
+  --original_state_dict_repo_id diffusers-internal-dev/new-model-image \
+  --dit_filename flux-dev-dummy.sft \
+  --dit \
+  --output_path .
+
+# Full pipe
+
+python scripts/convert_flux2_to_diffusers.py \
+  --original_state_dict_repo_id diffusers-internal-dev/new-model-image \
+  --dit_filename flux-dev-dummy.sft \
+  --vae_filename "flux2-vae.sft" \
+  --dit --vae --full_pipe \
+  --output_path .
+"""
+
+CTX = init_empty_weights if is_accelerate_available() else nullcontext
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--original_state_dict_repo_id", default=None, type=str)
+parser.add_argument("--vae_filename", default="flux2-vae.sft", type=str)
+parser.add_argument("--dit_filename", default="flux-dev-dummy.sft", type=str)
+parser.add_argument("--vae", action="store_true")
+parser.add_argument("--dit", action="store_true")
+parser.add_argument("--vae_dtype", type=str, default="fp32")
+parser.add_argument("--dit_dtype", type=str, default="bf16")
+parser.add_argument("--checkpoint_path", default=None, type=str)
+parser.add_argument("--full_pipe", action="store_true")
+parser.add_argument("--output_path", type=str)
+
+args = parser.parse_args()
+
+
+def load_original_checkpoint(args, filename):
+    if args.original_state_dict_repo_id is not None:
+        ckpt_path = hf_hub_download(repo_id=args.original_state_dict_repo_id, filename=filename)
+    elif args.checkpoint_path is not None:
+        ckpt_path = args.checkpoint_path
+    else:
+        raise ValueError(" please provide either `original_state_dict_repo_id` or a local `checkpoint_path`")
+
+    original_state_dict = safetensors.torch.load_file(ckpt_path)
+    return original_state_dict
+
+
+DIFFUSERS_VAE_TO_FLUX2_MAPPING = {
+    "encoder.conv_in.weight": "encoder.conv_in.weight",
+    "encoder.conv_in.bias": "encoder.conv_in.bias",
+    "encoder.conv_out.weight": "encoder.conv_out.weight",
+    "encoder.conv_out.bias": "encoder.conv_out.bias",
+    "encoder.conv_norm_out.weight": "encoder.norm_out.weight",
+    "encoder.conv_norm_out.bias": "encoder.norm_out.bias",
+    "decoder.conv_in.weight": "decoder.conv_in.weight",
+    "decoder.conv_in.bias": "decoder.conv_in.bias",
+    "decoder.conv_out.weight": "decoder.conv_out.weight",
+    "decoder.conv_out.bias": "decoder.conv_out.bias",
+    "decoder.conv_norm_out.weight": "decoder.norm_out.weight",
+    "decoder.conv_norm_out.bias": "decoder.norm_out.bias",
+    "quant_conv.weight": "encoder.quant_conv.weight",
+    "quant_conv.bias": "encoder.quant_conv.bias",
+    "post_quant_conv.weight": "decoder.post_quant_conv.weight",
+    "post_quant_conv.bias": "decoder.post_quant_conv.bias",
+    "bn.running_mean": "bn.running_mean",
+    "bn.running_var": "bn.running_var",
+}
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.conv_attn_to_linear
+def conv_attn_to_linear(checkpoint):
+    keys = list(checkpoint.keys())
+    attn_keys = ["query.weight", "key.weight", "value.weight"]
+    for key in keys:
+        if ".".join(key.split(".")[-2:]) in attn_keys:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0, 0]
+        elif "proj_attn.weight" in key:
+            if checkpoint[key].ndim > 2:
+                checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def update_vae_resnet_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
+    for ldm_key in keys:
+        diffusers_key = ldm_key.replace(mapping["old"], mapping["new"]).replace("nin_shortcut", "conv_shortcut")
+        new_checkpoint[diffusers_key] = checkpoint.get(ldm_key)
+
+
+def update_vae_attentions_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
+    for ldm_key in keys:
+        diffusers_key = (
+            ldm_key.replace(mapping["old"], mapping["new"])
+            .replace("norm.weight", "group_norm.weight")
+            .replace("norm.bias", "group_norm.bias")
+            .replace("q.weight", "to_q.weight")
+            .replace("q.bias", "to_q.bias")
+            .replace("k.weight", "to_k.weight")
+            .replace("k.bias", "to_k.bias")
+            .replace("v.weight", "to_v.weight")
+            .replace("v.bias", "to_v.bias")
+            .replace("proj_out.weight", "to_out.0.weight")
+            .replace("proj_out.bias", "to_out.0.bias")
+        )
+        new_checkpoint[diffusers_key] = checkpoint.get(ldm_key)
+
+        # proj_attn.weight has to be converted from conv 1D to linear
+        shape = new_checkpoint[diffusers_key].shape
+
+        if len(shape) == 3:
+            new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0]
+        elif len(shape) == 4:
+            new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0, 0]
+
+
+def convert_flux2_vae_checkpoint_to_diffusers(vae_state_dict, config):
+    new_checkpoint = {}
+    for diffusers_key, ldm_key in DIFFUSERS_VAE_TO_FLUX2_MAPPING.items():
+        if ldm_key not in vae_state_dict:
+            continue
+        new_checkpoint[diffusers_key] = vae_state_dict[ldm_key]
+
+    # Retrieves the keys for the encoder down blocks only
+    num_down_blocks = len(config["down_block_types"])
+    down_blocks = {
+        layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+    }
+
+    for i in range(num_down_blocks):
+        resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+        update_vae_resnet_ldm_to_diffusers(
+            resnets,
+            new_checkpoint,
+            vae_state_dict,
+            mapping={"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"},
+        )
+        if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.get(
+                f"encoder.down.{i}.downsample.conv.weight"
+            )
+            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.get(
+                f"encoder.down.{i}.downsample.conv.bias"
+            )
+
+    mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+        update_vae_resnet_ldm_to_diffusers(
+            resnets,
+            new_checkpoint,
+            vae_state_dict,
+            mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},
+        )
+
+    mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+    update_vae_attentions_ldm_to_diffusers(
+        mid_attentions, new_checkpoint, vae_state_dict, mapping={"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    )
+
+    # Retrieves the keys for the decoder up blocks only
+    num_up_blocks = len(config["up_block_types"])
+    up_blocks = {
+        layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+    }
+
+    for i in range(num_up_blocks):
+        block_id = num_up_blocks - 1 - i
+        resnets = [
+            key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+        ]
+        update_vae_resnet_ldm_to_diffusers(
+            resnets,
+            new_checkpoint,
+            vae_state_dict,
+            mapping={"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"},
+        )
+        if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.weight"
+            ]
+            new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+                f"decoder.up.{block_id}.upsample.conv.bias"
+            ]
+
+    mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+    num_mid_res_blocks = 2
+    for i in range(1, num_mid_res_blocks + 1):
+        resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+        update_vae_resnet_ldm_to_diffusers(
+            resnets,
+            new_checkpoint,
+            vae_state_dict,
+            mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},
+        )
+
+    mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+    update_vae_attentions_ldm_to_diffusers(
+        mid_attentions, new_checkpoint, vae_state_dict, mapping={"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+    )
+    conv_attn_to_linear(new_checkpoint)
+
+    return new_checkpoint
+
+
+FLUX2_TRANSFORMER_KEYS_RENAME_DICT = {
+    # Image and text input projections
+    "img_in": "x_embedder",
+    "txt_in": "context_embedder",
+    # Timestep and guidance embeddings
+    "time_in.in_layer": "time_guidance_embed.timestep_embedder.linear_1",
+    "time_in.out_layer": "time_guidance_embed.timestep_embedder.linear_2",
+    "guidance_in.in_layer": "time_guidance_embed.guidance_embedder.linear_1",
+    "guidance_in.out_layer": "time_guidance_embed.guidance_embedder.linear_2",
+    # Modulation parameters
+    "double_stream_modulation_img.lin": "double_stream_modulation_img.linear",
+    "double_stream_modulation_txt.lin": "double_stream_modulation_txt.linear",
+    "single_stream_modulation.lin": "single_stream_modulation.linear",
+    # Final output layer
+    # "final_layer.adaLN_modulation.1": "norm_out.linear",  # Handle separately since we need to swap mod params
+    "final_layer.linear": "proj_out",
+}
+
+
+FLUX2_TRANSFORMER_ADA_LAYER_NORM_KEY_MAP = {
+    "final_layer.adaLN_modulation.1": "norm_out.linear",
+}
+
+
+FLUX2_TRANSFORMER_DOUBLE_BLOCK_KEY_MAP = {
+    # Handle fused QKV projections separately as we need to break into Q, K, V projections
+    "img_attn.norm.query_norm": "attn.norm_q",
+    "img_attn.norm.key_norm": "attn.norm_k",
+    "img_attn.proj": "attn.to_out.0",
+    "img_mlp.0": "ff.linear_in",
+    "img_mlp.2": "ff.linear_out",
+    "txt_attn.norm.query_norm": "attn.norm_added_q",
+    "txt_attn.norm.key_norm": "attn.norm_added_k",
+    "txt_attn.proj": "attn.to_add_out",
+    "txt_mlp.0": "ff_context.linear_in",
+    "txt_mlp.2": "ff_context.linear_out",
+}
+
+
+FLUX2_TRANSFORMER_SINGLE_BLOCK_KEY_MAP = {
+    "linear1": "attn.to_qkv_mlp_proj",
+    "linear2": "attn.to_out",
+    "norm.query_norm": "attn.norm_q",
+    "norm.key_norm": "attn.norm_k",
+}
+
+
+# in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
+# while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use
+# diffusers implementation
+def swap_scale_shift(weight):
+    shift, scale = weight.chunk(2, dim=0)
+    new_weight = torch.cat([scale, shift], dim=0)
+    return new_weight
+
+
+def convert_ada_layer_norm_weights(key: str, state_dict: Dict[str, Any]) -> None:
+    # Skip if not a weight
+    if ".weight" not in key:
+        return
+
+    # If adaLN_modulation is in the key, swap scale and shift parameters
+    # Original implementation is (shift, scale); diffusers implementation is (scale, shift)
+    if "adaLN_modulation" in key:
+        key_without_param_type, param_type = key.rsplit(".", maxsplit=1)
+        # Assume all such keys are in the AdaLayerNorm key map
+        new_key_without_param_type = FLUX2_TRANSFORMER_ADA_LAYER_NORM_KEY_MAP[key_without_param_type]
+        new_key = ".".join([new_key_without_param_type, param_type])
+
+        swapped_weight = swap_scale_shift(state_dict.pop(key))
+        state_dict[new_key] = swapped_weight
+    return
+
+
+def convert_flux2_double_stream_blocks(key: str, state_dict: Dict[str, Any]) -> None:
+    # Skip if not a weight, bias, or scale
+    if ".weight" not in key and ".bias" not in key and ".scale" not in key:
+        return
+
+    new_prefix = "transformer_blocks"
+    if "double_blocks." in key:
+        parts = key.split(".")
+        block_idx = parts[1]
+        modality_block_name = parts[2]  # img_attn, img_mlp, txt_attn, txt_mlp
+        within_block_name = ".".join(parts[2:-1])
+        param_type = parts[-1]
+
+        if param_type == "scale":
+            param_type = "weight"
+
+        if "qkv" in within_block_name:
+            fused_qkv_weight = state_dict.pop(key)
+            to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+            if "img" in modality_block_name:
+                # double_blocks.{N}.img_attn.qkv --> transformer_blocks.{N}.attn.{to_q|to_k|to_v}
+                to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+                new_q_name = "attn.to_q"
+                new_k_name = "attn.to_k"
+                new_v_name = "attn.to_v"
+            elif "txt" in modality_block_name:
+                # double_blocks.{N}.txt_attn.qkv --> transformer_blocks.{N}.attn.{add_q_proj|add_k_proj|add_v_proj}
+                to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+                new_q_name = "attn.add_q_proj"
+                new_k_name = "attn.add_k_proj"
+                new_v_name = "attn.add_v_proj"
+            new_q_key = ".".join([new_prefix, block_idx, new_q_name, param_type])
+            new_k_key = ".".join([new_prefix, block_idx, new_k_name, param_type])
+            new_v_key = ".".join([new_prefix, block_idx, new_v_name, param_type])
+            state_dict[new_q_key] = to_q_weight
+            state_dict[new_k_key] = to_k_weight
+            state_dict[new_v_key] = to_v_weight
+        else:
+            new_within_block_name = FLUX2_TRANSFORMER_DOUBLE_BLOCK_KEY_MAP[within_block_name]
+            new_key = ".".join([new_prefix, block_idx, new_within_block_name, param_type])
+
+            param = state_dict.pop(key)
+            state_dict[new_key] = param
+    return
+
+
+def convert_flux2_single_stream_blocks(key: str, state_dict: Dict[str, Any]) -> None:
+    # Skip if not a weight, bias, or scale
+    if ".weight" not in key and ".bias" not in key and ".scale" not in key:
+        return
+
+    # Mapping:
+    #     - single_blocks.{N}.linear1               --> single_transformer_blocks.{N}.attn.to_qkv_mlp_proj
+    #     - single_blocks.{N}.linear2               --> single_transformer_blocks.{N}.attn.to_out
+    #     - single_blocks.{N}.norm.query_norm.scale --> single_transformer_blocks.{N}.attn.norm_q.weight
+    #     - single_blocks.{N}.norm.key_norm.scale   --> single_transformer_blocks.{N}.attn.norm_k.weight
+    new_prefix = "single_transformer_blocks"
+    if "single_blocks." in key:
+        parts = key.split(".")
+        block_idx = parts[1]
+        within_block_name = ".".join(parts[2:-1])
+        param_type = parts[-1]
+
+        if param_type == "scale":
+            param_type = "weight"
+
+        new_within_block_name = FLUX2_TRANSFORMER_SINGLE_BLOCK_KEY_MAP[within_block_name]
+        new_key = ".".join([new_prefix, block_idx, new_within_block_name, param_type])
+
+        param = state_dict.pop(key)
+        state_dict[new_key] = param
+    return
+
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {
+    "adaLN_modulation": convert_ada_layer_norm_weights,
+    "double_blocks": convert_flux2_double_stream_blocks,
+    "single_blocks": convert_flux2_single_stream_blocks,
+}
+
+
+def update_state_dict(state_dict: Dict[str, Any], old_key: str, new_key: str) -> None:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def get_flux2_transformer_config(model_type: str) -> Tuple[Dict[str, Any], ...]:
+    if model_type == "test" or model_type == "dummy-flux2":
+        config = {
+            "model_id": "diffusers-internal-dev/dummy-flux2",
+            "diffusers_config": {
+                "patch_size": 1,
+                "in_channels": 128,
+                "num_layers": 8,
+                "num_single_layers": 48,
+                "attention_head_dim": 128,
+                "num_attention_heads": 48,
+                "joint_attention_dim": 15360,
+                "timestep_guidance_channels": 256,
+                "mlp_ratio": 3.0,
+                "axes_dims_rope": (32, 32, 32, 32),
+                "rope_theta": 2000,
+                "eps": 1e-6,
+            },
+        }
+        rename_dict = FLUX2_TRANSFORMER_KEYS_RENAME_DICT
+        special_keys_remap = TRANSFORMER_SPECIAL_KEYS_REMAP
+    return config, rename_dict, special_keys_remap
+
+
+def convert_flux2_transformer_to_diffusers(original_state_dict: Dict[str, torch.Tensor], model_type: str):
+    config, rename_dict, special_keys_remap = get_flux2_transformer_config(model_type)
+
+    diffusers_config = config["diffusers_config"]
+
+    with init_empty_weights():
+        transformer = Flux2Transformer2DModel.from_config(diffusers_config)
+
+    # Handle official code --> diffusers key remapping via the remap dict
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in rename_dict.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict(original_state_dict, key, new_key)
+
+    # Handle any special logic which can't be expressed by a simple 1:1 remapping with the handlers in
+    # special_keys_remap
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in special_keys_remap.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True, assign=True)
+    return transformer
+
+
+def main(args):
+    if args.vae:
+        original_vae_ckpt = load_original_checkpoint(args, filename=args.vae_filename)
+        vae = AutoencoderKLFlux2()
+        converted_vae_state_dict = convert_flux2_vae_checkpoint_to_diffusers(original_vae_ckpt, vae.config)
+        vae.load_state_dict(converted_vae_state_dict, strict=True)
+        if not args.full_pipe:
+            vae_dtype = torch.bfloat16 if args.vae_dtype == "bf16" else torch.float32
+            vae.to(vae_dtype).save_pretrained(f"{args.output_path}/vae")
+
+    if args.dit:
+        original_dit_ckpt = load_original_checkpoint(args, filename=args.dit_filename)
+        transformer = convert_flux2_transformer_to_diffusers(original_dit_ckpt, "test")
+        if not args.full_pipe:
+            dit_dtype = torch.bfloat16 if args.dit_dtype == "bf16" else torch.float32
+            transformer.to(dit_dtype).save_pretrained(f"{args.output_path}/transformer")
+
+    if args.full_pipe:
+        tokenizer_id = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
+        text_encoder_id = "mistralai/Mistral-Small-3.2-24B-Instruct-2506"
+        generate_config = GenerationConfig.from_pretrained(text_encoder_id)
+        generate_config.do_sample = True
+        text_encoder = Mistral3ForConditionalGeneration.from_pretrained(
+            text_encoder_id, generation_config=generate_config, torch_dtype=torch.bfloat16
+        )
+        tokenizer = AutoProcessor.from_pretrained(tokenizer_id)
+        scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+            "black-forest-labs/FLUX.1-dev", subfolder="scheduler"
+        )
+
+        pipe = Flux2Pipeline(
+            vae=vae, transformer=transformer, text_encoder=text_encoder, tokenizer=tokenizer, scheduler=scheduler
+        )
+        pipe.save_pretrained(args.output_path)
+
+
+if __name__ == "__main__":
+    main(args)

scripts/convert_sana_controlnet_to_diffusers.py

@@ -10,7 +10,7 @@ from accelerate import init_empty_weights
 from diffusers import (
     SanaControlNetModel,
 )
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.utils.import_utils import is_accelerate_available
 
 

scripts/convert_sana_to_diffusers.py

@@ -20,7 +20,7 @@ from diffusers import (
     SanaTransformer2DModel,
     SCMScheduler,
 )
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.utils.import_utils import is_accelerate_available
 
 

scripts/convert_sana_video_to_diffusers.py

@@ -0,0 +1,327 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import os
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download, snapshot_download
+from termcolor import colored
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from diffusers import (
+    AutoencoderKLWan,
+    DPMSolverMultistepScheduler,
+    FlowMatchEulerDiscreteScheduler,
+    SanaVideoPipeline,
+    SanaVideoTransformer3DModel,
+    UniPCMultistepScheduler,
+)
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+ckpt_ids = ["Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth"]
+# https://github.com/NVlabs/Sana/blob/main/inference_video_scripts/inference_sana_video.py
+
+
+def main(args):
+    cache_dir_path = os.path.expanduser("~/.cache/huggingface/hub")
+
+    if args.orig_ckpt_path is None or args.orig_ckpt_path in ckpt_ids:
+        ckpt_id = args.orig_ckpt_path or ckpt_ids[0]
+        snapshot_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+        file_path = hf_hub_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            filename=f"{'/'.join(ckpt_id.split('/')[2:])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+    else:
+        file_path = args.orig_ckpt_path
+
+    print(colored(f"Loading checkpoint from {file_path}", "green", attrs=["bold"]))
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embedding.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embedding.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    # scheduler
+    flow_shift = 8.0
+    if args.task == "i2v":
+        assert args.scheduler_type == "flow-euler", "Scheduler type must be flow-euler for i2v task."
+
+    # model config
+    layer_num = 20
+    # Positional embedding interpolation scale.
+    qk_norm = True
+
+    # sample size
+    if args.video_size == 480:
+        sample_size = 30  # Wan-VAE: 8xp2 downsample factor
+        patch_size = (1, 2, 2)
+    elif args.video_size == 720:
+        sample_size = 22  # Wan-VAE: 32xp1 downsample factor
+        patch_size = (1, 1, 1)
+    else:
+        raise ValueError(f"Video size {args.video_size} is not supported.")
+
+    for depth in range(layer_num):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        if qk_norm is not None:
+            # Add Q/K normalization for self-attention (attn1) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.k_norm.weight"
+            )
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.point_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_temp.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.t_conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+        if qk_norm is not None:
+            # Add Q/K normalization for cross-attention (attn2) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.k_norm.weight"
+            )
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.bias"
+        )
+
+    # Final block.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
+    converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
+
+    # Transformer
+    with CTX():
+        transformer_kwargs = {
+            "in_channels": 16,
+            "out_channels": 16,
+            "num_attention_heads": 20,
+            "attention_head_dim": 112,
+            "num_layers": 20,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "caption_channels": 2304,
+            "mlp_ratio": 3.0,
+            "attention_bias": False,
+            "sample_size": sample_size,
+            "patch_size": patch_size,
+            "norm_elementwise_affine": False,
+            "norm_eps": 1e-6,
+            "qk_norm": "rms_norm_across_heads",
+            "rope_max_seq_len": 1024,
+        }
+
+        transformer = SanaVideoTransformer3DModel(**transformer_kwargs)
+
+    transformer.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    try:
+        state_dict.pop("y_embedder.y_embedding")
+        state_dict.pop("pos_embed")
+        state_dict.pop("logvar_linear.weight")
+        state_dict.pop("logvar_linear.bias")
+    except KeyError:
+        print("y_embedder.y_embedding or pos_embed not found in the state_dict")
+
+    assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    transformer = transformer.to(weight_dtype)
+
+    if not args.save_full_pipeline:
+        print(
+            colored(
+                f"Only saving transformer model of {args.model_type}. "
+                f"Set --save_full_pipeline to save the whole Pipeline",
+                "green",
+                attrs=["bold"],
+            )
+        )
+        transformer.save_pretrained(
+            os.path.join(args.dump_path, "transformer"), safe_serialization=True, max_shard_size="5GB"
+        )
+    else:
+        print(colored(f"Saving the whole Pipeline containing {args.model_type}", "green", attrs=["bold"]))
+        # VAE
+        vae = AutoencoderKLWan.from_pretrained(
+            "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32
+        )
+
+        # Text Encoder
+        text_encoder_model_path = "Efficient-Large-Model/gemma-2-2b-it"
+        tokenizer = AutoTokenizer.from_pretrained(text_encoder_model_path)
+        tokenizer.padding_side = "right"
+        text_encoder = AutoModelForCausalLM.from_pretrained(
+            text_encoder_model_path, torch_dtype=torch.bfloat16
+        ).get_decoder()
+
+        # Choose the appropriate pipeline and scheduler based on model type
+        # Original Sana scheduler
+        if args.scheduler_type == "flow-dpm_solver":
+            scheduler = DPMSolverMultistepScheduler(
+                flow_shift=flow_shift,
+                use_flow_sigmas=True,
+                prediction_type="flow_prediction",
+            )
+        elif args.scheduler_type == "flow-euler":
+            scheduler = FlowMatchEulerDiscreteScheduler(shift=flow_shift)
+        elif args.scheduler_type == "uni-pc":
+            scheduler = UniPCMultistepScheduler(
+                prediction_type="flow_prediction",
+                use_flow_sigmas=True,
+                num_train_timesteps=1000,
+                flow_shift=flow_shift,
+            )
+        else:
+            raise ValueError(f"Scheduler type {args.scheduler_type} is not supported")
+
+        pipe = SanaVideoPipeline(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+        pipe.save_pretrained(args.dump_path, safe_serialization=True, max_shard_size="5GB")
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--video_size",
+        default=480,
+        type=int,
+        choices=[480, 720],
+        required=False,
+        help="Video size of pretrained model, 480 or 720.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="SanaVideo",
+        type=str,
+        choices=[
+            "SanaVideo",
+        ],
+    )
+    parser.add_argument(
+        "--scheduler_type",
+        default="flow-dpm_solver",
+        type=str,
+        choices=["flow-dpm_solver", "flow-euler", "uni-pc"],
+        help="Scheduler type to use.",
+    )
+    parser.add_argument("--task", default="t2v", type=str, required=True, help="Task to convert, t2v or i2v.")
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipeline elements in one.")
+    parser.add_argument("--dtype", default="fp32", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+
+    main(args)

scripts/convert_sd3_to_diffusers.py

@@ -7,7 +7,7 @@ from accelerate import init_empty_weights
 
 from diffusers import AutoencoderKL, SD3Transformer2DModel
 from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.utils.import_utils import is_accelerate_available
 
 

scripts/convert_stable_audio.py

@@ -18,7 +18,7 @@ from diffusers import (
     StableAudioPipeline,
     StableAudioProjectionModel,
 )
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.utils import is_accelerate_available
 
 

scripts/convert_stable_cascade.py

@@ -20,7 +20,7 @@ from diffusers import (
 )
 from diffusers.loaders.single_file_utils import convert_stable_cascade_unet_single_file_to_diffusers
 from diffusers.models import StableCascadeUNet
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.pipelines.wuerstchen import PaellaVQModel
 from diffusers.utils import is_accelerate_available
 

scripts/convert_stable_cascade_lite.py

@@ -20,7 +20,7 @@ from diffusers import (
 )
 from diffusers.loaders.single_file_utils import convert_stable_cascade_unet_single_file_to_diffusers
 from diffusers.models import StableCascadeUNet
-from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.models.model_loading_utils import load_model_dict_into_meta
 from diffusers.pipelines.wuerstchen import PaellaVQModel
 from diffusers.utils import is_accelerate_available
 

scripts/convert_wan_to_diffusers.py

@@ -6,11 +6,20 @@ import torch
 from accelerate import init_empty_weights
 from huggingface_hub import hf_hub_download, snapshot_download
 from safetensors.torch import load_file
-from transformers import AutoProcessor, AutoTokenizer, CLIPVisionModelWithProjection, UMT5EncoderModel
+from transformers import (
+    AutoProcessor,
+    AutoTokenizer,
+    CLIPImageProcessor,
+    CLIPVisionModel,
+    CLIPVisionModelWithProjection,
+    UMT5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKLWan,
     UniPCMultistepScheduler,
+    WanAnimatePipeline,
+    WanAnimateTransformer3DModel,
     WanImageToVideoPipeline,
     WanPipeline,
     WanTransformer3DModel,
@@ -105,8 +114,203 @@ VACE_TRANSFORMER_KEYS_RENAME_DICT = {
     "after_proj": "proj_out",
 }
 
+ANIMATE_TRANSFORMER_KEYS_RENAME_DICT = {
+    "time_embedding.0": "condition_embedder.time_embedder.linear_1",
+    "time_embedding.2": "condition_embedder.time_embedder.linear_2",
+    "text_embedding.0": "condition_embedder.text_embedder.linear_1",
+    "text_embedding.2": "condition_embedder.text_embedder.linear_2",
+    "time_projection.1": "condition_embedder.time_proj",
+    "head.modulation": "scale_shift_table",
+    "head.head": "proj_out",
+    "modulation": "scale_shift_table",
+    "ffn.0": "ffn.net.0.proj",
+    "ffn.2": "ffn.net.2",
+    # Hack to swap the layer names
+    # The original model calls the norms in following order: norm1, norm3, norm2
+    # We convert it to: norm1, norm2, norm3
+    "norm2": "norm__placeholder",
+    "norm3": "norm2",
+    "norm__placeholder": "norm3",
+    "img_emb.proj.0": "condition_embedder.image_embedder.norm1",
+    "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
+    "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
+    "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+    # Add attention component mappings
+    "self_attn.q": "attn1.to_q",
+    "self_attn.k": "attn1.to_k",
+    "self_attn.v": "attn1.to_v",
+    "self_attn.o": "attn1.to_out.0",
+    "self_attn.norm_q": "attn1.norm_q",
+    "self_attn.norm_k": "attn1.norm_k",
+    "cross_attn.q": "attn2.to_q",
+    "cross_attn.k": "attn2.to_k",
+    "cross_attn.v": "attn2.to_v",
+    "cross_attn.o": "attn2.to_out.0",
+    "cross_attn.norm_q": "attn2.norm_q",
+    "cross_attn.norm_k": "attn2.norm_k",
+    "cross_attn.k_img": "attn2.to_k_img",
+    "cross_attn.v_img": "attn2.to_v_img",
+    "cross_attn.norm_k_img": "attn2.norm_k_img",
+    # After cross_attn -> attn2 rename, we need to rename the img keys
+    "attn2.to_k_img": "attn2.add_k_proj",
+    "attn2.to_v_img": "attn2.add_v_proj",
+    "attn2.norm_k_img": "attn2.norm_added_k",
+    # Wan Animate-specific mappings (motion encoder, face encoder, face adapter)
+    # Motion encoder mappings
+    # The name mapping is complicated for the convolutional part so we handle that in its own function
+    "motion_encoder.enc.fc": "motion_encoder.motion_network",
+    "motion_encoder.dec.direction.weight": "motion_encoder.motion_synthesis_weight",
+    # Face encoder mappings - CausalConv1d has a .conv submodule that we need to flatten
+    "face_encoder.conv1_local.conv": "face_encoder.conv1_local",
+    "face_encoder.conv2.conv": "face_encoder.conv2",
+    "face_encoder.conv3.conv": "face_encoder.conv3",
+    # Face adapter mappings are handled in a separate function
+}
+
+
+# TODO: Verify this and simplify if possible.
+def convert_animate_motion_encoder_weights(key: str, state_dict: Dict[str, Any], final_conv_idx: int = 8) -> None:
+    """
+    Convert all motion encoder weights for Animate model.
+
+    In the original model:
+    - All Linear layers in fc use EqualLinear
+    - All Conv2d layers in convs use EqualConv2d (except blur_conv which is initialized separately)
+    - Blur kernels are stored as buffers in Sequential modules
+    - ConvLayer is nn.Sequential with indices: [Blur (optional), EqualConv2d, FusedLeakyReLU (optional)]
+
+    Conversion strategy:
+    1. Drop .kernel buffers (blur kernels)
+    2. Rename sequential indices to named components (e.g., 0 -> conv2d, 1 -> bias_leaky_relu)
+    """
+    # Skip if not a weight, bias, or kernel
+    if ".weight" not in key and ".bias" not in key and ".kernel" not in key:
+        return
+
+    # Handle Blur kernel buffers from original implementation.
+    # After renaming, these appear under: motion_encoder.res_blocks.*.conv{2,skip}.blur_kernel
+    # Diffusers constructs blur kernels as a non-persistent buffer so we must drop these keys
+    if ".kernel" in key and "motion_encoder" in key:
+        # Remove unexpected blur kernel buffers to avoid strict load errors
+        state_dict.pop(key, None)
+        return
+
+    # Rename Sequential indices to named components in ConvLayer and ResBlock
+    if ".enc.net_app.convs." in key and (".weight" in key or ".bias" in key):
+        parts = key.split(".")
+
+        # Find the sequential index (digit) after convs or after conv1/conv2/skip
+        # Examples:
+        # - enc.net_app.convs.0.0.weight -> conv_in.weight (initial conv layer weight)
+        # - enc.net_app.convs.0.1.bias -> conv_in.act_fn.bias (initial conv layer bias)
+        # - enc.net_app.convs.{n:1-7}.conv1.0.weight -> res_blocks.{(n-1):0-6}.conv1.weight (conv1 weight)
+        #     - e.g. enc.net_app.convs.1.conv1.0.weight -> res_blocks.0.conv1.weight
+        # - enc.net_app.convs.{n:1-7}.conv1.1.bias -> res_blocks.{(n-1):0-6}.conv1.act_fn.bias (conv1 bias)
+        #     - e.g. enc.net_app.convs.1.conv1.1.bias -> res_blocks.0.conv1.act_fn.bias
+        # - enc.net_app.convs.{n:1-7}.conv2.1.weight -> res_blocks.{(n-1):0-6}.conv2.weight (conv2 weight)
+        # - enc.net_app.convs.1.conv2.2.bias -> res_blocks.0.conv2.act_fn.bias (conv2 bias)
+        # - enc.net_app.convs.{n:1-7}.skip.1.weight -> res_blocks.{(n-1):0-6}.conv_skip.weight (skip conv weight)
+        # - enc.net_app.convs.8 -> conv_out (final conv layer)
+
+        convs_idx = parts.index("convs") if "convs" in parts else -1
+        if convs_idx >= 0 and len(parts) - convs_idx >= 2:
+            bias = False
+            # The nn.Sequential index will always follow convs
+            sequential_idx = int(parts[convs_idx + 1])
+            if sequential_idx == 0:
+                if key.endswith(".weight"):
+                    new_key = "motion_encoder.conv_in.weight"
+                elif key.endswith(".bias"):
+                    new_key = "motion_encoder.conv_in.act_fn.bias"
+                    bias = True
+            elif sequential_idx == final_conv_idx:
+                if key.endswith(".weight"):
+                    new_key = "motion_encoder.conv_out.weight"
+            else:
+                # Intermediate .convs. layers, which get mapped to .res_blocks.
+                prefix = "motion_encoder.res_blocks."
+
+                layer_name = parts[convs_idx + 2]
+                if layer_name == "skip":
+                    layer_name = "conv_skip"
+
+                if key.endswith(".weight"):
+                    param_name = "weight"
+                elif key.endswith(".bias"):
+                    param_name = "act_fn.bias"
+                    bias = True
+
+                suffix_parts = [str(sequential_idx - 1), layer_name, param_name]
+                suffix = ".".join(suffix_parts)
+                new_key = prefix + suffix
+
+            param = state_dict.pop(key)
+            if bias:
+                param = param.squeeze()
+            state_dict[new_key] = param
+            return
+        return
+    return
+
+
+def convert_animate_face_adapter_weights(key: str, state_dict: Dict[str, Any]) -> None:
+    """
+    Convert face adapter weights for the Animate model.
+
+    The original model uses a fused KV projection but the diffusers models uses separate K and V projections.
+    """
+    # Skip if not a weight or bias
+    if ".weight" not in key and ".bias" not in key:
+        return
+
+    prefix = "face_adapter."
+    if ".fuser_blocks." in key:
+        parts = key.split(".")
+
+        module_list_idx = parts.index("fuser_blocks") if "fuser_blocks" in parts else -1
+        if module_list_idx >= 0 and (len(parts) - 1) - module_list_idx == 3:
+            block_idx = parts[module_list_idx + 1]
+            layer_name = parts[module_list_idx + 2]
+            param_name = parts[module_list_idx + 3]
+
+            if layer_name == "linear1_kv":
+                layer_name_k = "to_k"
+                layer_name_v = "to_v"
+
+                suffix_k = ".".join([block_idx, layer_name_k, param_name])
+                suffix_v = ".".join([block_idx, layer_name_v, param_name])
+                new_key_k = prefix + suffix_k
+                new_key_v = prefix + suffix_v
+
+                kv_proj = state_dict.pop(key)
+                k_proj, v_proj = torch.chunk(kv_proj, 2, dim=0)
+                state_dict[new_key_k] = k_proj
+                state_dict[new_key_v] = v_proj
+                return
+            else:
+                if layer_name == "q_norm":
+                    new_layer_name = "norm_q"
+                elif layer_name == "k_norm":
+                    new_layer_name = "norm_k"
+                elif layer_name == "linear1_q":
+                    new_layer_name = "to_q"
+                elif layer_name == "linear2":
+                    new_layer_name = "to_out"
+
+                suffix_parts = [block_idx, new_layer_name, param_name]
+                suffix = ".".join(suffix_parts)
+                new_key = prefix + suffix
+                state_dict[new_key] = state_dict.pop(key)
+                return
+    return
+
+
 TRANSFORMER_SPECIAL_KEYS_REMAP = {}
 VACE_TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+ANIMATE_TRANSFORMER_SPECIAL_KEYS_REMAP = {
+    "motion_encoder": convert_animate_motion_encoder_weights,
+    "face_adapter": convert_animate_face_adapter_weights,
+}
 
 
 def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
@@ -364,6 +568,37 @@ def get_transformer_config(model_type: str) -> Tuple[Dict[str, Any], ...]:
         }
         RENAME_DICT = TRANSFORMER_KEYS_RENAME_DICT
         SPECIAL_KEYS_REMAP = TRANSFORMER_SPECIAL_KEYS_REMAP
+    elif model_type == "Wan2.2-Animate-14B":
+        config = {
+            "model_id": "Wan-AI/Wan2.2-Animate-14B",
+            "diffusers_config": {
+                "image_dim": 1280,
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": (1, 2, 2),
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "rope_max_seq_len": 1024,
+                "pos_embed_seq_len": None,
+                "motion_encoder_size": 512,  # Start of Wan Animate-specific configs
+                "motion_style_dim": 512,
+                "motion_dim": 20,
+                "motion_encoder_dim": 512,
+                "face_encoder_hidden_dim": 1024,
+                "face_encoder_num_heads": 4,
+                "inject_face_latents_blocks": 5,
+            },
+        }
+        RENAME_DICT = ANIMATE_TRANSFORMER_KEYS_RENAME_DICT
+        SPECIAL_KEYS_REMAP = ANIMATE_TRANSFORMER_SPECIAL_KEYS_REMAP
     return config, RENAME_DICT, SPECIAL_KEYS_REMAP
 
 
@@ -380,10 +615,12 @@ def convert_transformer(model_type: str, stage: str = None):
     original_state_dict = load_sharded_safetensors(model_dir)
 
     with init_empty_weights():
-        if "VACE" not in model_type:
-            transformer = WanTransformer3DModel.from_config(diffusers_config)
-        else:
+        if "Animate" in model_type:
+            transformer = WanAnimateTransformer3DModel.from_config(diffusers_config)
+        elif "VACE" in model_type:
             transformer = WanVACETransformer3DModel.from_config(diffusers_config)
+        else:
+            transformer = WanTransformer3DModel.from_config(diffusers_config)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -397,7 +634,12 @@ def convert_transformer(model_type: str, stage: str = None):
                 continue
             handler_fn_inplace(key, original_state_dict)
 
+    # Load state dict into the meta model, which will materialize the tensors
     transformer.load_state_dict(original_state_dict, strict=True, assign=True)
+
+    # Move to CPU to ensure all tensors are materialized
+    transformer = transformer.to("cpu")
+
     return transformer
 
 
@@ -926,7 +1168,7 @@ DTYPE_MAPPING = {
 if __name__ == "__main__":
     args = get_args()
 
-    if "Wan2.2" in args.model_type and "TI2V" not in args.model_type:
+    if "Wan2.2" in args.model_type and "TI2V" not in args.model_type and "Animate" not in args.model_type:
         transformer = convert_transformer(args.model_type, stage="high_noise_model")
         transformer_2 = convert_transformer(args.model_type, stage="low_noise_model")
     else:
@@ -942,7 +1184,7 @@ if __name__ == "__main__":
     tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
     if "FLF2V" in args.model_type:
         flow_shift = 16.0
-    elif "TI2V" in args.model_type:
+    elif "TI2V" in args.model_type or "Animate" in args.model_type:
         flow_shift = 5.0
     else:
         flow_shift = 3.0
@@ -954,6 +1196,8 @@ if __name__ == "__main__":
     if args.dtype != "none":
         dtype = DTYPE_MAPPING[args.dtype]
         transformer.to(dtype)
+        if transformer_2 is not None:
+            transformer_2.to(dtype)
 
     if "Wan2.2" and "I2V" in args.model_type and "TI2V" not in args.model_type:
         pipe = WanImageToVideoPipeline(
@@ -1016,6 +1260,21 @@ if __name__ == "__main__":
             vae=vae,
             scheduler=scheduler,
         )
+    elif "Animate" in args.model_type:
+        image_encoder = CLIPVisionModel.from_pretrained(
+            "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", torch_dtype=torch.bfloat16
+        )
+        image_processor = CLIPImageProcessor.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+
+        pipe = WanAnimatePipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            image_processor=image_processor,
+        )
     else:
         pipe = WanPipeline(
             transformer=transformer,

src/diffusers/__init__.py

@@ -186,6 +186,7 @@ else:
             "AutoencoderKLAllegro",
             "AutoencoderKLCogVideoX",
             "AutoencoderKLCosmos",
+            "AutoencoderKLFlux2",
             "AutoencoderKLHunyuanImage",
             "AutoencoderKLHunyuanImageRefiner",
             "AutoencoderKLHunyuanVideo",
@@ -198,9 +199,11 @@ else:
             "AutoencoderOobleck",
             "AutoencoderTiny",
             "AutoModel",
+            "BriaFiboTransformer2DModel",
             "BriaTransformer2DModel",
             "CacheMixin",
             "ChromaTransformer2DModel",
+            "ChronoEditTransformer3DModel",
             "CogVideoXTransformer3DModel",
             "CogView3PlusTransformer2DModel",
             "CogView4Transformer2DModel",
@@ -213,6 +216,7 @@ else:
             "CosmosTransformer3DModel",
             "DiTTransformer2DModel",
             "EasyAnimateTransformer3DModel",
+            "Flux2Transformer2DModel",
             "FluxControlNetModel",
             "FluxMultiControlNetModel",
             "FluxTransformer2DModel",
@@ -245,6 +249,7 @@ else:
             "QwenImageTransformer2DModel",
             "SanaControlNetModel",
             "SanaTransformer2DModel",
+            "SanaVideoTransformer3DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
             "SD3Transformer2DModel",
@@ -265,8 +270,10 @@ else:
             "UNetSpatioTemporalConditionModel",
             "UVit2DModel",
             "VQModel",
+            "WanAnimateTransformer3DModel",
             "WanTransformer3DModel",
             "WanVACETransformer3DModel",
+            "ZImageTransformer2DModel",
             "attention_backend",
         ]
     )
@@ -404,6 +411,7 @@ else:
             "QwenImageModularPipeline",
             "StableDiffusionXLAutoBlocks",
             "StableDiffusionXLModularPipeline",
+            "Wan22AutoBlocks",
             "WanAutoBlocks",
             "WanModularPipeline",
         ]
@@ -430,9 +438,11 @@ else:
             "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
+            "BriaFiboPipeline",
             "BriaPipeline",
             "ChromaImg2ImgPipeline",
             "ChromaPipeline",
+            "ChronoEditPipeline",
             "CLIPImageProjection",
             "CogVideoXFunControlPipeline",
             "CogVideoXImageToVideoPipeline",
@@ -450,6 +460,7 @@ else:
             "EasyAnimateControlPipeline",
             "EasyAnimateInpaintPipeline",
             "EasyAnimatePipeline",
+            "Flux2Pipeline",
             "FluxControlImg2ImgPipeline",
             "FluxControlInpaintPipeline",
             "FluxControlNetImg2ImgPipeline",
@@ -538,10 +549,13 @@ else:
             "QwenImagePipeline",
             "ReduxImageEncoder",
             "SanaControlNetPipeline",
+            "SanaImageToVideoPipeline",
             "SanaPAGPipeline",
             "SanaPipeline",
             "SanaSprintImg2ImgPipeline",
             "SanaSprintPipeline",
+            "SanaVideoPipeline",
+            "SanaVideoPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -629,6 +643,7 @@ else:
             "VisualClozeGenerationPipeline",
             "VisualClozePipeline",
             "VQDiffusionPipeline",
+            "WanAnimatePipeline",
             "WanImageToVideoPipeline",
             "WanPipeline",
             "WanVACEPipeline",
@@ -636,6 +651,7 @@ else:
             "WuerstchenCombinedPipeline",
             "WuerstchenDecoderPipeline",
             "WuerstchenPriorPipeline",
+            "ZImagePipeline",
         ]
     )
 
@@ -889,6 +905,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
             AutoencoderKLCosmos,
+            AutoencoderKLFlux2,
             AutoencoderKLHunyuanImage,
             AutoencoderKLHunyuanImageRefiner,
             AutoencoderKLHunyuanVideo,
@@ -901,9 +918,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderOobleck,
             AutoencoderTiny,
             AutoModel,
+            BriaFiboTransformer2DModel,
             BriaTransformer2DModel,
             CacheMixin,
             ChromaTransformer2DModel,
+            ChronoEditTransformer3DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,
@@ -916,6 +935,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CosmosTransformer3DModel,
             DiTTransformer2DModel,
             EasyAnimateTransformer3DModel,
+            Flux2Transformer2DModel,
             FluxControlNetModel,
             FluxMultiControlNetModel,
             FluxTransformer2DModel,
@@ -948,6 +968,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageTransformer2DModel,
             SanaControlNetModel,
             SanaTransformer2DModel,
+            SanaVideoTransformer3DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
             SD3Transformer2DModel,
@@ -967,6 +988,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             UNetSpatioTemporalConditionModel,
             UVit2DModel,
             VQModel,
+            WanAnimateTransformer3DModel,
             WanTransformer3DModel,
             WanVACETransformer3DModel,
             attention_backend,
@@ -1081,6 +1103,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageModularPipeline,
             StableDiffusionXLAutoBlocks,
             StableDiffusionXLModularPipeline,
+            Wan22AutoBlocks,
             WanAutoBlocks,
             WanModularPipeline,
         )
@@ -1103,9 +1126,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
             AuraFlowPipeline,
+            BriaFiboPipeline,
             BriaPipeline,
             ChromaImg2ImgPipeline,
             ChromaPipeline,
+            ChronoEditPipeline,
             CLIPImageProjection,
             CogVideoXFunControlPipeline,
             CogVideoXImageToVideoPipeline,
@@ -1123,6 +1148,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             EasyAnimateControlPipeline,
             EasyAnimateInpaintPipeline,
             EasyAnimatePipeline,
+            Flux2Pipeline,
             FluxControlImg2ImgPipeline,
             FluxControlInpaintPipeline,
             FluxControlNetImg2ImgPipeline,
@@ -1211,10 +1237,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImagePipeline,
             ReduxImageEncoder,
             SanaControlNetPipeline,
+            SanaImageToVideoPipeline,
             SanaPAGPipeline,
             SanaPipeline,
             SanaSprintImg2ImgPipeline,
             SanaSprintPipeline,
+            SanaVideoPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,
@@ -1301,6 +1329,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VisualClozeGenerationPipeline,
             VisualClozePipeline,
             VQDiffusionPipeline,
+            WanAnimatePipeline,
             WanImageToVideoPipeline,
             WanPipeline,
             WanVACEPipeline,
@@ -1308,6 +1337,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
+            ZImagePipeline,
         )
 
     try:

src/diffusers/guiders/adaptive_projected_guidance.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -88,6 +88,19 @@ class AdaptiveProjectedGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        if self._step == 0:
+            if self.adaptive_projected_guidance_momentum is not None:
+                self.momentum_buffer = MomentumBuffer(self.adaptive_projected_guidance_momentum)
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/adaptive_projected_guidance_mix.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -99,6 +99,19 @@ class AdaptiveProjectedMixGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        if self._step == 0:
+            if self.adaptive_projected_guidance_momentum is not None:
+                self.momentum_buffer = MomentumBuffer(self.adaptive_projected_guidance_momentum)
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/auto_guidance.py

@@ -141,6 +141,16 @@ class AutoGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/classifier_free_guidance.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -99,6 +99,16 @@ class ClassifierFreeGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/classifier_free_zero_star_guidance.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -85,6 +85,16 @@ class ClassifierFreeZeroStarGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/frequency_decoupled_guidance.py

@@ -226,6 +226,16 @@ class FrequencyDecoupledGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/guiders/guider_utils.py

@@ -166,6 +166,11 @@ class BaseGuidance(ConfigMixin, PushToHubMixin):
     def prepare_inputs(self, data: "BlockState") -> List["BlockState"]:
         raise NotImplementedError("BaseGuidance::prepare_inputs must be implemented in subclasses.")
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        raise NotImplementedError("BaseGuidance::prepare_inputs_from_block_state must be implemented in subclasses.")
+
     def __call__(self, data: List["BlockState"]) -> Any:
         if not all(hasattr(d, "noise_pred") for d in data):
             raise ValueError("Expected all data to have `noise_pred` attribute.")
@@ -234,6 +239,51 @@ class BaseGuidance(ConfigMixin, PushToHubMixin):
         data_batch[cls._identifier_key] = identifier
         return BlockState(**data_batch)
 
+    @classmethod
+    def _prepare_batch_from_block_state(
+        cls,
+        input_fields: Dict[str, Union[str, Tuple[str, str]]],
+        data: "BlockState",
+        tuple_index: int,
+        identifier: str,
+    ) -> "BlockState":
+        """
+        Prepares a batch of data for the guidance technique. This method is used in the `prepare_inputs` method of the
+        `BaseGuidance` class. It prepares the batch based on the provided tuple index.
+
+        Args:
+            input_fields (`Dict[str, Union[str, Tuple[str, str]]]`):
+                A dictionary where the keys are the names of the fields that will be used to store the data once it is
+                prepared with `prepare_inputs`. The values can be either a string or a tuple of length 2, which is used
+                to look up the required data provided for preparation. If a string is provided, it will be used as the
+                conditional data (or unconditional if used with a guidance method that requires it). If a tuple of
+                length 2 is provided, the first element must be the conditional data identifier and the second element
+                must be the unconditional data identifier or None.
+            data (`BlockState`):
+                The input data to be prepared.
+            tuple_index (`int`):
+                The index to use when accessing input fields that are tuples.
+
+        Returns:
+            `BlockState`: The prepared batch of data.
+        """
+        from ..modular_pipelines.modular_pipeline import BlockState
+
+        data_batch = {}
+        for key, value in input_fields.items():
+            try:
+                if isinstance(value, str):
+                    data_batch[key] = getattr(data, value)
+                elif isinstance(value, tuple):
+                    data_batch[key] = getattr(data, value[tuple_index])
+                else:
+                    # We've already checked that value is a string or a tuple of strings with length 2
+                    pass
+            except AttributeError:
+                logger.debug(f"`data` does not have attribute(s) {value}, skipping.")
+        data_batch[cls._identifier_key] = identifier
+        return BlockState(**data_batch)
+
     @classmethod
     @validate_hf_hub_args
     def from_pretrained(
@@ -323,7 +373,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
     r"""
     Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
     Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
-    Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+    Flawed](https://huggingface.co/papers/2305.08891).
 
     Args:
         noise_cfg (`torch.Tensor`):

src/diffusers/guiders/perturbed_attention_guidance.py

@@ -187,6 +187,26 @@ class PerturbedAttentionGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        if self.num_conditions == 1:
+            tuple_indices = [0]
+            input_predictions = ["pred_cond"]
+        elif self.num_conditions == 2:
+            tuple_indices = [0, 1]
+            input_predictions = (
+                ["pred_cond", "pred_uncond"] if self._is_cfg_enabled() else ["pred_cond", "pred_cond_skip"]
+            )
+        else:
+            tuple_indices = [0, 1, 0]
+            input_predictions = ["pred_cond", "pred_uncond", "pred_cond_skip"]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     # Copied from diffusers.guiders.skip_layer_guidance.SkipLayerGuidance.forward
     def forward(
         self,

src/diffusers/guiders/skip_layer_guidance.py

@@ -183,6 +183,26 @@ class SkipLayerGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        if self.num_conditions == 1:
+            tuple_indices = [0]
+            input_predictions = ["pred_cond"]
+        elif self.num_conditions == 2:
+            tuple_indices = [0, 1]
+            input_predictions = (
+                ["pred_cond", "pred_uncond"] if self._is_cfg_enabled() else ["pred_cond", "pred_cond_skip"]
+            )
+        else:
+            tuple_indices = [0, 1, 0]
+            input_predictions = ["pred_cond", "pred_uncond", "pred_cond_skip"]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(
         self,
         pred_cond: torch.Tensor,

src/diffusers/guiders/smoothed_energy_guidance.py

@@ -172,6 +172,26 @@ class SmoothedEnergyGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        if self.num_conditions == 1:
+            tuple_indices = [0]
+            input_predictions = ["pred_cond"]
+        elif self.num_conditions == 2:
+            tuple_indices = [0, 1]
+            input_predictions = (
+                ["pred_cond", "pred_uncond"] if self._is_cfg_enabled() else ["pred_cond", "pred_cond_seg"]
+            )
+        else:
+            tuple_indices = [0, 1, 0]
+            input_predictions = ["pred_cond", "pred_uncond", "pred_cond_seg"]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(
         self,
         pred_cond: torch.Tensor,

src/diffusers/guiders/tangential_classifier_free_guidance.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import TYPE_CHECKING, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -74,6 +74,16 @@ class TangentialClassifierFreeGuidance(BaseGuidance):
             data_batches.append(data_batch)
         return data_batches
 
+    def prepare_inputs_from_block_state(
+        self, data: "BlockState", input_fields: Dict[str, Union[str, Tuple[str, str]]]
+    ) -> List["BlockState"]:
+        tuple_indices = [0] if self.num_conditions == 1 else [0, 1]
+        data_batches = []
+        for tuple_idx, input_prediction in zip(tuple_indices, self._input_predictions):
+            data_batch = self._prepare_batch_from_block_state(input_fields, data, tuple_idx, input_prediction)
+            data_batches.append(data_batch)
+        return data_batches
+
     def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> GuiderOutput:
         pred = None
 

src/diffusers/hooks/_helpers.py

@@ -111,6 +111,7 @@ def _register_attention_processors_metadata():
     from ..models.transformers.transformer_hunyuanimage import HunyuanImageAttnProcessor
     from ..models.transformers.transformer_qwenimage import QwenDoubleStreamAttnProcessor2_0
     from ..models.transformers.transformer_wan import WanAttnProcessor2_0
+    from ..models.transformers.transformer_z_image import ZSingleStreamAttnProcessor
 
     # AttnProcessor2_0
     AttentionProcessorRegistry.register(
@@ -158,6 +159,14 @@ def _register_attention_processors_metadata():
         ),
     )
 
+    # ZSingleStreamAttnProcessor
+    AttentionProcessorRegistry.register(
+        model_class=ZSingleStreamAttnProcessor,
+        metadata=AttentionProcessorMetadata(
+            skip_processor_output_fn=_skip_proc_output_fn_Attention_ZSingleStreamAttnProcessor,
+        ),
+    )
+
 
 def _register_transformer_blocks_metadata():
     from ..models.attention import BasicTransformerBlock
@@ -179,6 +188,7 @@ def _register_transformer_blocks_metadata():
     from ..models.transformers.transformer_mochi import MochiTransformerBlock
     from ..models.transformers.transformer_qwenimage import QwenImageTransformerBlock
     from ..models.transformers.transformer_wan import WanTransformerBlock
+    from ..models.transformers.transformer_z_image import ZImageTransformerBlock
 
     # BasicTransformerBlock
     TransformerBlockRegistry.register(
@@ -312,6 +322,15 @@ def _register_transformer_blocks_metadata():
         ),
     )
 
+    # ZImage
+    TransformerBlockRegistry.register(
+        model_class=ZImageTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
+
 
 # fmt: off
 def _skip_attention___ret___hidden_states(self, *args, **kwargs):
@@ -338,4 +357,5 @@ _skip_proc_output_fn_Attention_WanAttnProcessor2_0 = _skip_attention___ret___hid
 _skip_proc_output_fn_Attention_FluxAttnProcessor = _skip_attention___ret___hidden_states
 _skip_proc_output_fn_Attention_QwenDoubleStreamAttnProcessor2_0 = _skip_attention___ret___hidden_states
 _skip_proc_output_fn_Attention_HunyuanImageAttnProcessor = _skip_attention___ret___hidden_states
+_skip_proc_output_fn_Attention_ZSingleStreamAttnProcessor = _skip_attention___ret___hidden_states
 # fmt: on

src/diffusers/hooks/context_parallel.py

@@ -203,10 +203,12 @@ class ContextParallelSplitHook(ModelHook):
 
     def _prepare_cp_input(self, x: torch.Tensor, cp_input: ContextParallelInput) -> torch.Tensor:
         if cp_input.expected_dims is not None and x.dim() != cp_input.expected_dims:
-            raise ValueError(
-                f"Expected input tensor to have {cp_input.expected_dims} dimensions, but got {x.dim()} dimensions."
+            logger.warning_once(
+                f"Expected input tensor to have {cp_input.expected_dims} dimensions, but got {x.dim()} dimensions, split will not be applied."
             )
-        return EquipartitionSharder.shard(x, cp_input.split_dim, self.parallel_config._flattened_mesh)
+            return x
+        else:
+            return EquipartitionSharder.shard(x, cp_input.split_dim, self.parallel_config._flattened_mesh)
 
 
 class ContextParallelGatherHook(ModelHook):

src/diffusers/image_processor.py

@@ -409,7 +409,7 @@ class VaeImageProcessor(ConfigMixin):
         src_w = width if ratio < src_ratio else image.width * height // image.height
         src_h = height if ratio >= src_ratio else image.height * width // image.width
 
-        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION[self.config.resample])
         res = Image.new("RGB", (width, height))
         res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
 
@@ -460,7 +460,7 @@ class VaeImageProcessor(ConfigMixin):
         src_w = width if ratio > src_ratio else image.width * height // image.height
         src_h = height if ratio <= src_ratio else image.height * width // image.width
 
-        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION[self.config.resample])
         res = Image.new("RGB", (width, height))
         res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
         return res
@@ -1045,16 +1045,39 @@ class VaeImageProcessorLDM3D(VaeImageProcessor):
     def rgblike_to_depthmap(image: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
         r"""
         Convert an RGB-like depth image to a depth map.
-
-        Args:
-            image (`Union[np.ndarray, torch.Tensor]`):
-                The RGB-like depth image to convert.
-
-        Returns:
-            `Union[np.ndarray, torch.Tensor]`:
-                The corresponding depth map.
         """
-        return image[:, :, 1] * 2**8 + image[:, :, 2]
+        # 1. Cast the tensor to a larger integer type (e.g., int32)
+        #    to safely perform the multiplication by 256.
+        # 2. Perform the 16-bit combination: High-byte * 256 + Low-byte.
+        # 3. Cast the final result to the desired depth map type (uint16) if needed
+        #    before returning, though leaving it as int32/int64 is often safer
+        #    for return value from a library function.
+
+        if isinstance(image, torch.Tensor):
+            # Cast to a safe dtype (e.g., int32 or int64) for the calculation
+            original_dtype = image.dtype
+            image_safe = image.to(torch.int32)
+
+            # Calculate the depth map
+            depth_map = image_safe[:, :, 1] * 256 + image_safe[:, :, 2]
+
+            # You may want to cast the final result to uint16, but casting to a
+            # larger int type (like int32) is sufficient to fix the overflow.
+            # depth_map = depth_map.to(torch.uint16) # Uncomment if uint16 is strictly required
+            return depth_map.to(original_dtype)
+
+        elif isinstance(image, np.ndarray):
+            # NumPy equivalent: Cast to a safe dtype (e.g., np.int32)
+            original_dtype = image.dtype
+            image_safe = image.astype(np.int32)
+
+            # Calculate the depth map
+            depth_map = image_safe[:, :, 1] * 256 + image_safe[:, :, 2]
+
+            # depth_map = depth_map.astype(np.uint16) # Uncomment if uint16 is strictly required
+            return depth_map.astype(original_dtype)
+        else:
+            raise TypeError("Input image must be a torch.Tensor or np.ndarray")
 
     def numpy_to_depth(self, images: np.ndarray) -> List[PIL.Image.Image]:
         r"""

src/diffusers/loaders/__init__.py

@@ -81,6 +81,7 @@ if is_torch_available():
             "HiDreamImageLoraLoaderMixin",
             "SkyReelsV2LoraLoaderMixin",
             "QwenImageLoraLoaderMixin",
+            "Flux2LoraLoaderMixin",
         ]
         _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
         _import_structure["ip_adapter"] = [
@@ -113,6 +114,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 AuraFlowLoraLoaderMixin,
                 CogVideoXLoraLoaderMixin,
                 CogView4LoraLoaderMixin,
+                Flux2LoraLoaderMixin,
                 FluxLoraLoaderMixin,
                 HiDreamImageLoraLoaderMixin,
                 HunyuanVideoLoraLoaderMixin,

src/diffusers/loaders/lora_conversion_utils.py

@@ -2213,6 +2213,10 @@ def _convert_non_diffusers_qwen_lora_to_diffusers(state_dict):
 
         state_dict = {convert_key(k): v for k, v in state_dict.items()}
 
+    has_default = any("default." in k for k in state_dict)
+    if has_default:
+        state_dict = {k.replace("default.", ""): v for k, v in state_dict.items()}
+
     converted_state_dict = {}
     all_keys = list(state_dict.keys())
     down_key = ".lora_down.weight"
@@ -2261,3 +2265,89 @@ def _convert_non_diffusers_qwen_lora_to_diffusers(state_dict):
 
     converted_state_dict = {f"transformer.{k}": v for k, v in converted_state_dict.items()}
     return converted_state_dict
+
+
+def _convert_non_diffusers_flux2_lora_to_diffusers(state_dict):
+    converted_state_dict = {}
+
+    prefix = "diffusion_model."
+    original_state_dict = {k[len(prefix) :]: v for k, v in state_dict.items()}
+
+    num_double_layers = 8
+    num_single_layers = 48
+    lora_keys = ("lora_A", "lora_B")
+    attn_types = ("img_attn", "txt_attn")
+
+    for sl in range(num_single_layers):
+        single_block_prefix = f"single_blocks.{sl}"
+        attn_prefix = f"single_transformer_blocks.{sl}.attn"
+
+        for lora_key in lora_keys:
+            converted_state_dict[f"{attn_prefix}.to_qkv_mlp_proj.{lora_key}.weight"] = original_state_dict.pop(
+                f"{single_block_prefix}.linear1.{lora_key}.weight"
+            )
+
+            converted_state_dict[f"{attn_prefix}.to_out.{lora_key}.weight"] = original_state_dict.pop(
+                f"{single_block_prefix}.linear2.{lora_key}.weight"
+            )
+
+    for dl in range(num_double_layers):
+        transformer_block_prefix = f"transformer_blocks.{dl}"
+
+        for lora_key in lora_keys:
+            for attn_type in attn_types:
+                attn_prefix = f"{transformer_block_prefix}.attn"
+                qkv_key = f"double_blocks.{dl}.{attn_type}.qkv.{lora_key}.weight"
+                fused_qkv_weight = original_state_dict.pop(qkv_key)
+
+                if lora_key == "lora_A":
+                    diff_attn_proj_keys = (
+                        ["to_q", "to_k", "to_v"]
+                        if attn_type == "img_attn"
+                        else ["add_q_proj", "add_k_proj", "add_v_proj"]
+                    )
+                    for proj_key in diff_attn_proj_keys:
+                        converted_state_dict[f"{attn_prefix}.{proj_key}.{lora_key}.weight"] = torch.cat(
+                            [fused_qkv_weight]
+                        )
+                else:
+                    sample_q, sample_k, sample_v = torch.chunk(fused_qkv_weight, 3, dim=0)
+
+                    if attn_type == "img_attn":
+                        converted_state_dict[f"{attn_prefix}.to_q.{lora_key}.weight"] = torch.cat([sample_q])
+                        converted_state_dict[f"{attn_prefix}.to_k.{lora_key}.weight"] = torch.cat([sample_k])
+                        converted_state_dict[f"{attn_prefix}.to_v.{lora_key}.weight"] = torch.cat([sample_v])
+                    else:
+                        converted_state_dict[f"{attn_prefix}.add_q_proj.{lora_key}.weight"] = torch.cat([sample_q])
+                        converted_state_dict[f"{attn_prefix}.add_k_proj.{lora_key}.weight"] = torch.cat([sample_k])
+                        converted_state_dict[f"{attn_prefix}.add_v_proj.{lora_key}.weight"] = torch.cat([sample_v])
+
+        proj_mappings = [
+            ("img_attn.proj", "attn.to_out.0"),
+            ("txt_attn.proj", "attn.to_add_out"),
+        ]
+        for org_proj, diff_proj in proj_mappings:
+            for lora_key in lora_keys:
+                original_key = f"double_blocks.{dl}.{org_proj}.{lora_key}.weight"
+                diffusers_key = f"{transformer_block_prefix}.{diff_proj}.{lora_key}.weight"
+                converted_state_dict[diffusers_key] = original_state_dict.pop(original_key)
+
+        mlp_mappings = [
+            ("img_mlp.0", "ff.linear_in"),
+            ("img_mlp.2", "ff.linear_out"),
+            ("txt_mlp.0", "ff_context.linear_in"),
+            ("txt_mlp.2", "ff_context.linear_out"),
+        ]
+        for org_mlp, diff_mlp in mlp_mappings:
+            for lora_key in lora_keys:
+                original_key = f"double_blocks.{dl}.{org_mlp}.{lora_key}.weight"
+                diffusers_key = f"{transformer_block_prefix}.{diff_mlp}.{lora_key}.weight"
+                converted_state_dict[diffusers_key] = original_state_dict.pop(original_key)
+
+    if len(original_state_dict) > 0:
+        raise ValueError(f"`original_state_dict` should be empty at this point but has {original_state_dict.keys()=}.")
+
+    for key in list(converted_state_dict.keys()):
+        converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
+
+    return converted_state_dict

src/diffusers/loaders/lora_pipeline.py

@@ -45,6 +45,7 @@ from .lora_conversion_utils import (
     _convert_hunyuan_video_lora_to_diffusers,
     _convert_kohya_flux_lora_to_diffusers,
     _convert_musubi_wan_lora_to_diffusers,
+    _convert_non_diffusers_flux2_lora_to_diffusers,
     _convert_non_diffusers_hidream_lora_to_diffusers,
     _convert_non_diffusers_lora_to_diffusers,
     _convert_non_diffusers_ltxv_lora_to_diffusers,
@@ -4940,7 +4941,8 @@ class QwenImageLoraLoaderMixin(LoraBaseMixin):
         has_alphas_in_sd = any(k.endswith(".alpha") for k in state_dict)
         has_lora_unet = any(k.startswith("lora_unet_") for k in state_dict)
         has_diffusion_model = any(k.startswith("diffusion_model.") for k in state_dict)
-        if has_alphas_in_sd or has_lora_unet or has_diffusion_model:
+        has_default = any("default." in k for k in state_dict)
+        if has_alphas_in_sd or has_lora_unet or has_diffusion_model or has_default:
             state_dict = _convert_non_diffusers_qwen_lora_to_diffusers(state_dict)
 
         out = (state_dict, metadata) if return_lora_metadata else state_dict
@@ -5083,6 +5085,209 @@ class QwenImageLoraLoaderMixin(LoraBaseMixin):
         super().unfuse_lora(components=components, **kwargs)
 
 
+class Flux2LoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`Flux2Transformer2DModel`]. Specific to [`Flux2Pipeline`].
+    """
+
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+
+    @classmethod
+    @validate_hf_hub_args
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details.
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
+
+        state_dict, metadata = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+
+        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
+        if is_dora_scale_present:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+
+        is_ai_toolkit = any(k.startswith("diffusion_model.") for k in state_dict)
+        if is_ai_toolkit:
+            state_dict = _convert_non_diffusers_flux2_lora_to_diffusers(state_dict)
+
+        out = (state_dict, metadata) if return_lora_metadata else state_dict
+        return out
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        adapter_name: Optional[str] = None,
+        hotswap: bool = False,
+        **kwargs,
+    ):
+        """
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for more details.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        kwargs["return_lora_metadata"] = True
+        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->CogView4Transformer2DModel
+    def load_lora_into_transformer(
+        cls,
+        state_dict,
+        transformer,
+        adapter_name=None,
+        _pipeline=None,
+        low_cpu_mem_usage=False,
+        hotswap: bool = False,
+        metadata=None,
+    ):
+        """
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_unet`] for more details.
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        transformer_lora_adapter_metadata: Optional[dict] = None,
+    ):
+        r"""
+        See [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for more information.
+        """
+        lora_layers = {}
+        lora_metadata = {}
+
+        if transformer_lora_layers:
+            lora_layers[cls.transformer_name] = transformer_lora_layers
+            lora_metadata[cls.transformer_name] = transformer_lora_adapter_metadata
+
+        if not lora_layers:
+            raise ValueError("You must pass at least one of `transformer_lora_layers` or `text_encoder_lora_layers`.")
+
+        cls._save_lora_weights(
+            save_directory=save_directory,
+            lora_layers=lora_layers,
+            lora_metadata=lora_metadata,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        See [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] for more details.
+        """
+        super().fuse_lora(
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
+    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
+        r"""
+        See [`~loaders.StableDiffusionLoraLoaderMixin.unfuse_lora`] for more details.
+        """
+        super().unfuse_lora(components=components, **kwargs)
+
+
 class LoraLoaderMixin(StableDiffusionLoraLoaderMixin):
     def __init__(self, *args, **kwargs):
         deprecation_message = "LoraLoaderMixin is deprecated and this will be removed in a future version. Please use `StableDiffusionLoraLoaderMixin`, instead."

src/diffusers/loaders/peft.py

@@ -62,6 +62,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "WanVACETransformer3DModel": lambda model_cls, weights: weights,
     "ChromaTransformer2DModel": lambda model_cls, weights: weights,
     "QwenImageTransformer2DModel": lambda model_cls, weights: weights,
+    "Flux2Transformer2DModel": lambda model_cls, weights: weights,
 }
 
 

src/diffusers/loaders/single_file_model.py

@@ -34,6 +34,7 @@ from .single_file_utils import (
     convert_chroma_transformer_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
     convert_cosmos_transformer_checkpoint_to_diffusers,
+    convert_flux2_transformer_checkpoint_to_diffusers,
     convert_flux_transformer_checkpoint_to_diffusers,
     convert_hidream_transformer_to_diffusers,
     convert_hunyuan_video_transformer_to_diffusers,
@@ -162,6 +163,10 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": lambda x: x,
         "default_subfolder": "transformer",
     },
+    "Flux2Transformer2DModel": {
+        "checkpoint_mapping_fn": convert_flux2_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -140,6 +140,7 @@ CHECKPOINT_KEY_NAMES = {
         "net.blocks.0.self_attn.q_proj.weight",
         "net.pos_embedder.dim_spatial_range",
     ],
+    "flux2": ["model.diffusion_model.single_stream_modulation.lin.weight", "single_stream_modulation.lin.weight"],
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -189,6 +190,7 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "flux-fill": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-Fill-dev"},
     "flux-depth": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-Depth-dev"},
     "flux-schnell": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-schnell"},
+    "flux-2-dev": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.2-dev"},
     "ltx-video": {"pretrained_model_name_or_path": "diffusers/LTX-Video-0.9.0"},
     "ltx-video-0.9.1": {"pretrained_model_name_or_path": "diffusers/LTX-Video-0.9.1"},
     "ltx-video-0.9.5": {"pretrained_model_name_or_path": "Lightricks/LTX-Video-0.9.5"},
@@ -649,6 +651,9 @@ def infer_diffusers_model_type(checkpoint):
         else:
             model_type = "animatediff_v3"
 
+    elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["flux2"]):
+        model_type = "flux-2-dev"
+
     elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["flux"]):
         if any(
             g in checkpoint for g in ["guidance_in.in_layer.bias", "model.diffusion_model.guidance_in.in_layer.bias"]
@@ -3647,3 +3652,168 @@ def convert_cosmos_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
             handler_fn_inplace(key, converted_state_dict)
 
     return converted_state_dict
+
+
+def convert_flux2_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    FLUX2_TRANSFORMER_KEYS_RENAME_DICT = {
+        # Image and text input projections
+        "img_in": "x_embedder",
+        "txt_in": "context_embedder",
+        # Timestep and guidance embeddings
+        "time_in.in_layer": "time_guidance_embed.timestep_embedder.linear_1",
+        "time_in.out_layer": "time_guidance_embed.timestep_embedder.linear_2",
+        "guidance_in.in_layer": "time_guidance_embed.guidance_embedder.linear_1",
+        "guidance_in.out_layer": "time_guidance_embed.guidance_embedder.linear_2",
+        # Modulation parameters
+        "double_stream_modulation_img.lin": "double_stream_modulation_img.linear",
+        "double_stream_modulation_txt.lin": "double_stream_modulation_txt.linear",
+        "single_stream_modulation.lin": "single_stream_modulation.linear",
+        # Final output layer
+        # "final_layer.adaLN_modulation.1": "norm_out.linear",  # Handle separately since we need to swap mod params
+        "final_layer.linear": "proj_out",
+    }
+
+    FLUX2_TRANSFORMER_ADA_LAYER_NORM_KEY_MAP = {
+        "final_layer.adaLN_modulation.1": "norm_out.linear",
+    }
+
+    FLUX2_TRANSFORMER_DOUBLE_BLOCK_KEY_MAP = {
+        # Handle fused QKV projections separately as we need to break into Q, K, V projections
+        "img_attn.norm.query_norm": "attn.norm_q",
+        "img_attn.norm.key_norm": "attn.norm_k",
+        "img_attn.proj": "attn.to_out.0",
+        "img_mlp.0": "ff.linear_in",
+        "img_mlp.2": "ff.linear_out",
+        "txt_attn.norm.query_norm": "attn.norm_added_q",
+        "txt_attn.norm.key_norm": "attn.norm_added_k",
+        "txt_attn.proj": "attn.to_add_out",
+        "txt_mlp.0": "ff_context.linear_in",
+        "txt_mlp.2": "ff_context.linear_out",
+    }
+
+    FLUX2_TRANSFORMER_SINGLE_BLOCK_KEY_MAP = {
+        "linear1": "attn.to_qkv_mlp_proj",
+        "linear2": "attn.to_out",
+        "norm.query_norm": "attn.norm_q",
+        "norm.key_norm": "attn.norm_k",
+    }
+
+    def convert_flux2_single_stream_blocks(key: str, state_dict: dict[str, object]) -> None:
+        # Skip if not a weight, bias, or scale
+        if ".weight" not in key and ".bias" not in key and ".scale" not in key:
+            return
+
+        # Mapping:
+        #     - single_blocks.{N}.linear1               --> single_transformer_blocks.{N}.attn.to_qkv_mlp_proj
+        #     - single_blocks.{N}.linear2               --> single_transformer_blocks.{N}.attn.to_out
+        #     - single_blocks.{N}.norm.query_norm.scale --> single_transformer_blocks.{N}.attn.norm_q.weight
+        #     - single_blocks.{N}.norm.key_norm.scale   --> single_transformer_blocks.{N}.attn.norm_k.weight
+        new_prefix = "single_transformer_blocks"
+        if "single_blocks." in key:
+            parts = key.split(".")
+            block_idx = parts[1]
+            within_block_name = ".".join(parts[2:-1])
+            param_type = parts[-1]
+
+            if param_type == "scale":
+                param_type = "weight"
+
+            new_within_block_name = FLUX2_TRANSFORMER_SINGLE_BLOCK_KEY_MAP[within_block_name]
+            new_key = ".".join([new_prefix, block_idx, new_within_block_name, param_type])
+
+            param = state_dict.pop(key)
+            state_dict[new_key] = param
+
+        return
+
+    def convert_ada_layer_norm_weights(key: str, state_dict: dict[str, object]) -> None:
+        # Skip if not a weight
+        if ".weight" not in key:
+            return
+
+        # If adaLN_modulation is in the key, swap scale and shift parameters
+        # Original implementation is (shift, scale); diffusers implementation is (scale, shift)
+        if "adaLN_modulation" in key:
+            key_without_param_type, param_type = key.rsplit(".", maxsplit=1)
+            # Assume all such keys are in the AdaLayerNorm key map
+            new_key_without_param_type = FLUX2_TRANSFORMER_ADA_LAYER_NORM_KEY_MAP[key_without_param_type]
+            new_key = ".".join([new_key_without_param_type, param_type])
+
+            swapped_weight = swap_scale_shift(state_dict.pop(key), 0)
+            state_dict[new_key] = swapped_weight
+
+        return
+
+    def convert_flux2_double_stream_blocks(key: str, state_dict: dict[str, object]) -> None:
+        # Skip if not a weight, bias, or scale
+        if ".weight" not in key and ".bias" not in key and ".scale" not in key:
+            return
+
+        new_prefix = "transformer_blocks"
+        if "double_blocks." in key:
+            parts = key.split(".")
+            block_idx = parts[1]
+            modality_block_name = parts[2]  # img_attn, img_mlp, txt_attn, txt_mlp
+            within_block_name = ".".join(parts[2:-1])
+            param_type = parts[-1]
+
+            if param_type == "scale":
+                param_type = "weight"
+
+            if "qkv" in within_block_name:
+                fused_qkv_weight = state_dict.pop(key)
+                to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+                if "img" in modality_block_name:
+                    # double_blocks.{N}.img_attn.qkv --> transformer_blocks.{N}.attn.{to_q|to_k|to_v}
+                    to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+                    new_q_name = "attn.to_q"
+                    new_k_name = "attn.to_k"
+                    new_v_name = "attn.to_v"
+                elif "txt" in modality_block_name:
+                    # double_blocks.{N}.txt_attn.qkv --> transformer_blocks.{N}.attn.{add_q_proj|add_k_proj|add_v_proj}
+                    to_q_weight, to_k_weight, to_v_weight = torch.chunk(fused_qkv_weight, 3, dim=0)
+                    new_q_name = "attn.add_q_proj"
+                    new_k_name = "attn.add_k_proj"
+                    new_v_name = "attn.add_v_proj"
+                new_q_key = ".".join([new_prefix, block_idx, new_q_name, param_type])
+                new_k_key = ".".join([new_prefix, block_idx, new_k_name, param_type])
+                new_v_key = ".".join([new_prefix, block_idx, new_v_name, param_type])
+                state_dict[new_q_key] = to_q_weight
+                state_dict[new_k_key] = to_k_weight
+                state_dict[new_v_key] = to_v_weight
+            else:
+                new_within_block_name = FLUX2_TRANSFORMER_DOUBLE_BLOCK_KEY_MAP[within_block_name]
+                new_key = ".".join([new_prefix, block_idx, new_within_block_name, param_type])
+
+                param = state_dict.pop(key)
+                state_dict[new_key] = param
+        return
+
+    def update_state_dict(state_dict: dict[str, object], old_key: str, new_key: str) -> None:
+        state_dict[new_key] = state_dict.pop(old_key)
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP = {
+        "adaLN_modulation": convert_ada_layer_norm_weights,
+        "double_blocks": convert_flux2_double_stream_blocks,
+        "single_blocks": convert_flux2_single_stream_blocks,
+    }
+
+    converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys())}
+
+    # Handle official code --> diffusers key remapping via the remap dict
+    for key in list(converted_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in FLUX2_TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+
+        update_state_dict(converted_state_dict, key, new_key)
+
+    # Handle any special logic which can't be expressed by a simple 1:1 remapping with the handlers in
+    # special_keys_remap
+    for key in list(converted_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict

src/diffusers/models/__init__.py

@@ -35,6 +35,7 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_kl_allegro"] = ["AutoencoderKLAllegro"]
     _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
     _import_structure["autoencoders.autoencoder_kl_cosmos"] = ["AutoencoderKLCosmos"]
+    _import_structure["autoencoders.autoencoder_kl_flux2"] = ["AutoencoderKLFlux2"]
     _import_structure["autoencoders.autoencoder_kl_hunyuan_video"] = ["AutoencoderKLHunyuanVideo"]
     _import_structure["autoencoders.autoencoder_kl_hunyuanimage"] = ["AutoencoderKLHunyuanImage"]
     _import_structure["autoencoders.autoencoder_kl_hunyuanimage_refiner"] = ["AutoencoderKLHunyuanImageRefiner"]
@@ -84,12 +85,15 @@ if is_torch_available():
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
     _import_structure["transformers.transformer_allegro"] = ["AllegroTransformer3DModel"]
     _import_structure["transformers.transformer_bria"] = ["BriaTransformer2DModel"]
+    _import_structure["transformers.transformer_bria_fibo"] = ["BriaFiboTransformer2DModel"]
     _import_structure["transformers.transformer_chroma"] = ["ChromaTransformer2DModel"]
+    _import_structure["transformers.transformer_chronoedit"] = ["ChronoEditTransformer3DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_cogview4"] = ["CogView4Transformer2DModel"]
     _import_structure["transformers.transformer_cosmos"] = ["CosmosTransformer3DModel"]
     _import_structure["transformers.transformer_easyanimate"] = ["EasyAnimateTransformer3DModel"]
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
+    _import_structure["transformers.transformer_flux2"] = ["Flux2Transformer2DModel"]
     _import_structure["transformers.transformer_hidream_image"] = ["HiDreamImageTransformer2DModel"]
     _import_structure["transformers.transformer_hunyuan_video"] = ["HunyuanVideoTransformer3DModel"]
     _import_structure["transformers.transformer_hunyuan_video_framepack"] = ["HunyuanVideoFramepackTransformer3DModel"]
@@ -101,11 +105,14 @@ if is_torch_available():
     _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
     _import_structure["transformers.transformer_prx"] = ["PRXTransformer2DModel"]
     _import_structure["transformers.transformer_qwenimage"] = ["QwenImageTransformer2DModel"]
+    _import_structure["transformers.transformer_sana_video"] = ["SanaVideoTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["transformers.transformer_wan"] = ["WanTransformer3DModel"]
+    _import_structure["transformers.transformer_wan_animate"] = ["WanAnimateTransformer3DModel"]
     _import_structure["transformers.transformer_wan_vace"] = ["WanVACETransformer3DModel"]
+    _import_structure["transformers.transformer_z_image"] = ["ZImageTransformer2DModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
     _import_structure["unets.unet_2d_condition"] = ["UNet2DConditionModel"]
@@ -136,6 +143,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
             AutoencoderKLCosmos,
+            AutoencoderKLFlux2,
             AutoencoderKLHunyuanImage,
             AutoencoderKLHunyuanImageRefiner,
             AutoencoderKLHunyuanVideo,
@@ -174,8 +182,10 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .transformers import (
             AllegroTransformer3DModel,
             AuraFlowTransformer2DModel,
+            BriaFiboTransformer2DModel,
             BriaTransformer2DModel,
             ChromaTransformer2DModel,
+            ChronoEditTransformer3DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,
@@ -184,6 +194,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             DiTTransformer2DModel,
             DualTransformer2DModel,
             EasyAnimateTransformer3DModel,
+            Flux2Transformer2DModel,
             FluxTransformer2DModel,
             HiDreamImageTransformer2DModel,
             HunyuanDiT2DModel,
@@ -202,14 +213,17 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PRXTransformer2DModel,
             QwenImageTransformer2DModel,
             SanaTransformer2DModel,
+            SanaVideoTransformer3DModel,
             SD3Transformer2DModel,
             SkyReelsV2Transformer3DModel,
             StableAudioDiTModel,
             T5FilmDecoder,
             Transformer2DModel,
             TransformerTemporalModel,
+            WanAnimateTransformer3DModel,
             WanTransformer3DModel,
             WanVACETransformer3DModel,
+            ZImageTransformer2DModel,
         )
         from .unets import (
             I2VGenXLUNet,

src/diffusers/models/_modeling_parallel.py

@@ -44,11 +44,16 @@ class ContextParallelConfig:
 
     Args:
         ring_degree (`int`, *optional*, defaults to `1`):
-            Number of devices to use for ring attention within a context parallel region. Must be a divisor of the
-            total number of devices in the context parallel mesh.
+            Number of devices to use for Ring Attention. Sequence is split across devices. Each device computes
+            attention between its local Q and KV chunks passed sequentially around ring. Lower memory (only holds 1/N
+            of KV at a time), overlaps compute with communication, but requires N iterations to see all tokens. Best
+            for long sequences with limited memory/bandwidth. Number of devices to use for ring attention within a
+            context parallel region. Must be a divisor of the total number of devices in the context parallel mesh.
         ulysses_degree (`int`, *optional*, defaults to `1`):
-            Number of devices to use for ulysses attention within a context parallel region. Must be a divisor of the
-            total number of devices in the context parallel mesh.
+            Number of devices to use for Ulysses Attention. Sequence split is across devices. Each device computes
+            local QKV, then all-gathers all KV chunks to compute full attention in one pass. Higher memory (stores all
+            KV), requires high-bandwidth all-to-all communication, but lower latency. Best for moderate sequences with
+            good interconnect bandwidth.
         convert_to_fp32 (`bool`, *optional*, defaults to `True`):
             Whether to convert output and LSE to float32 for ring attention numerical stability.
         rotate_method (`str`, *optional*, defaults to `"allgather"`):
@@ -79,29 +84,46 @@ class ContextParallelConfig:
         if self.ulysses_degree is None:
             self.ulysses_degree = 1
 
+        if self.ring_degree == 1 and self.ulysses_degree == 1:
+            raise ValueError(
+                "Either ring_degree or ulysses_degree must be greater than 1 in order to use context parallel inference"
+            )
+        if self.ring_degree < 1 or self.ulysses_degree < 1:
+            raise ValueError("`ring_degree` and `ulysses_degree` must be greater than or equal to 1.")
+        if self.ring_degree > 1 and self.ulysses_degree > 1:
+            raise ValueError(
+                "Unified Ulysses-Ring attention is not yet supported. Please set either `ring_degree` or `ulysses_degree` to 1."
+            )
+        if self.rotate_method != "allgather":
+            raise NotImplementedError(
+                f"Only rotate_method='allgather' is supported for now, but got {self.rotate_method}."
+            )
+
+    @property
+    def mesh_shape(self) -> Tuple[int, int]:
+        return (self.ring_degree, self.ulysses_degree)
+
+    @property
+    def mesh_dim_names(self) -> Tuple[str, str]:
+        """Dimension names for the device mesh."""
+        return ("ring", "ulysses")
+
     def setup(self, rank: int, world_size: int, device: torch.device, mesh: torch.distributed.device_mesh.DeviceMesh):
         self._rank = rank
         self._world_size = world_size
         self._device = device
         self._mesh = mesh
-        if self.ring_degree is None:
-            self.ring_degree = 1
-        if self.ulysses_degree is None:
-            self.ulysses_degree = 1
-        if self.rotate_method != "allgather":
-            raise NotImplementedError(
-                f"Only rotate_method='allgather' is supported for now, but got {self.rotate_method}."
+
+        if self.ulysses_degree * self.ring_degree > world_size:
+            raise ValueError(
+                f"The product of `ring_degree` ({self.ring_degree}) and `ulysses_degree` ({self.ulysses_degree}) must not exceed the world size ({world_size})."
             )
-        if self._flattened_mesh is None:
-            self._flattened_mesh = self._mesh._flatten()
-        if self._ring_mesh is None:
-            self._ring_mesh = self._mesh["ring"]
-        if self._ulysses_mesh is None:
-            self._ulysses_mesh = self._mesh["ulysses"]
-        if self._ring_local_rank is None:
-            self._ring_local_rank = self._ring_mesh.get_local_rank()
-        if self._ulysses_local_rank is None:
-            self._ulysses_local_rank = self._ulysses_mesh.get_local_rank()
+
+        self._flattened_mesh = self._mesh._flatten()
+        self._ring_mesh = self._mesh["ring"]
+        self._ulysses_mesh = self._mesh["ulysses"]
+        self._ring_local_rank = self._ring_mesh.get_local_rank()
+        self._ulysses_local_rank = self._ulysses_mesh.get_local_rank()
 
 
 @dataclass
@@ -119,7 +141,7 @@ class ParallelConfig:
     _rank: int = None
     _world_size: int = None
     _device: torch.device = None
-    _cp_mesh: torch.distributed.device_mesh.DeviceMesh = None
+    _mesh: torch.distributed.device_mesh.DeviceMesh = None
 
     def setup(
         self,
@@ -127,14 +149,14 @@ class ParallelConfig:
         world_size: int,
         device: torch.device,
         *,
-        cp_mesh: Optional[torch.distributed.device_mesh.DeviceMesh] = None,
+        mesh: Optional[torch.distributed.device_mesh.DeviceMesh] = None,
     ):
         self._rank = rank
         self._world_size = world_size
         self._device = device
-        self._cp_mesh = cp_mesh
+        self._mesh = mesh
         if self.context_parallel_config is not None:
-            self.context_parallel_config.setup(rank, world_size, device, cp_mesh)
+            self.context_parallel_config.setup(rank, world_size, device, mesh)
 
 
 @dataclass(frozen=True)

src/diffusers/models/attention.py

@@ -105,7 +105,7 @@ class AttentionMixin:
                 raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
 
         for module in self.modules():
-            if isinstance(module, AttentionModuleMixin):
+            if isinstance(module, AttentionModuleMixin) and module._supports_qkv_fusion:
                 module.fuse_projections()
 
     def unfuse_qkv_projections(self):
@@ -114,13 +114,14 @@ class AttentionMixin:
         > [!WARNING] > This API is 🧪 experimental.
         """
         for module in self.modules():
-            if isinstance(module, AttentionModuleMixin):
+            if isinstance(module, AttentionModuleMixin) and module._supports_qkv_fusion:
                 module.unfuse_projections()
 
 
 class AttentionModuleMixin:
     _default_processor_cls = None
     _available_processors = []
+    _supports_qkv_fusion = True
     fused_projections = False
 
     def set_processor(self, processor: AttentionProcessor) -> None:
@@ -248,6 +249,14 @@ class AttentionModuleMixin:
         """
         Fuse the query, key, and value projections into a single projection for efficiency.
         """
+        # Skip if the AttentionModuleMixin subclass does not support fusion (for example, the QKV projections in Flux2
+        # single stream blocks are always fused)
+        if not self._supports_qkv_fusion:
+            logger.debug(
+                f"{self.__class__.__name__} does not support fusing QKV projections, so `fuse_projections` will no-op."
+            )
+            return
+
         # Skip if already fused
         if getattr(self, "fused_projections", False):
             return
@@ -307,6 +316,11 @@ class AttentionModuleMixin:
         """
         Unfuse the query, key, and value projections back to separate projections.
         """
+        # Skip if the AttentionModuleMixin subclass does not support fusion (for example, the QKV projections in Flux2
+        # single stream blocks are always fused)
+        if not self._supports_qkv_fusion:
+            return
+
         # Skip if not fused
         if not getattr(self, "fused_projections", False):
             return

src/diffusers/models/attention_dispatch.py

@@ -16,8 +16,9 @@ import contextlib
 import functools
 import inspect
 import math
+from dataclasses import dataclass
 from enum import Enum
-from typing import TYPE_CHECKING, Any, Callable, Dict, List, Literal, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 
@@ -42,7 +43,7 @@ from ..utils import (
     is_xformers_available,
     is_xformers_version,
 )
-from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS, DIFFUSERS_ENABLE_HUB_KERNELS
+from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS
 
 
 if TYPE_CHECKING:
@@ -82,24 +83,11 @@ else:
     flash_attn_3_func = None
     flash_attn_3_varlen_func = None
 
-
 if _CAN_USE_AITER_ATTN:
     from aiter import flash_attn_func as aiter_flash_attn_func
 else:
     aiter_flash_attn_func = None
 
-if DIFFUSERS_ENABLE_HUB_KERNELS:
-    if not is_kernels_available():
-        raise ImportError(
-            "To use FA3 kernel for your hardware from the Hub, the `kernels` library must be installed. Install with `pip install kernels`."
-        )
-    from ..utils.kernels_utils import _get_fa3_from_hub
-
-    flash_attn_interface_hub = _get_fa3_from_hub()
-    flash_attn_3_func_hub = flash_attn_interface_hub.flash_attn_func
-else:
-    flash_attn_3_func_hub = None
-
 if _CAN_USE_SAGE_ATTN:
     from sageattention import (
         sageattn,
@@ -172,16 +160,13 @@ logger = get_logger(__name__)  # pylint: disable=invalid-name
 # - CP with sage attention, flex, xformers, other missing backends
 # - Add support for normal and CP training with backends that don't support it yet
 
-_SAGE_ATTENTION_PV_ACCUM_DTYPE = Literal["fp32", "fp32+fp32"]
-_SAGE_ATTENTION_QK_QUANT_GRAN = Literal["per_thread", "per_warp"]
-_SAGE_ATTENTION_QUANTIZATION_BACKEND = Literal["cuda", "triton"]
-
 
 class AttentionBackendName(str, Enum):
     # EAGER = "eager"
 
     # `flash-attn`
     FLASH = "flash"
+    FLASH_HUB = "flash_hub"
     FLASH_VARLEN = "flash_varlen"
     _FLASH_3 = "_flash_3"
     _FLASH_VARLEN_3 = "_flash_varlen_3"
@@ -203,6 +188,7 @@ class AttentionBackendName(str, Enum):
 
     # `sageattention`
     SAGE = "sage"
+    SAGE_HUB = "sage_hub"
     SAGE_VARLEN = "sage_varlen"
     _SAGE_QK_INT8_PV_FP8_CUDA = "_sage_qk_int8_pv_fp8_cuda"
     _SAGE_QK_INT8_PV_FP8_CUDA_SM90 = "_sage_qk_int8_pv_fp8_cuda_sm90"
@@ -220,7 +206,7 @@ class _AttentionBackendRegistry:
     _backends = {}
     _constraints = {}
     _supported_arg_names = {}
-    _supports_context_parallel = {}
+    _supports_context_parallel = set()
     _active_backend = AttentionBackendName(DIFFUSERS_ATTN_BACKEND)
     _checks_enabled = DIFFUSERS_ATTN_CHECKS
 
@@ -237,7 +223,9 @@ class _AttentionBackendRegistry:
             cls._backends[backend] = func
             cls._constraints[backend] = constraints or []
             cls._supported_arg_names[backend] = set(inspect.signature(func).parameters.keys())
-            cls._supports_context_parallel[backend] = supports_context_parallel
+            if supports_context_parallel:
+                cls._supports_context_parallel.add(backend.value)
+
             return func
 
         return decorator
@@ -251,15 +239,37 @@ class _AttentionBackendRegistry:
         return list(cls._backends.keys())
 
     @classmethod
-    def _is_context_parallel_enabled(
-        cls, backend: AttentionBackendName, parallel_config: Optional["ParallelConfig"]
+    def _is_context_parallel_available(
+        cls,
+        backend: AttentionBackendName,
     ) -> bool:
-        supports_context_parallel = backend in cls._supports_context_parallel
-        is_degree_greater_than_1 = parallel_config is not None and (
-            parallel_config.context_parallel_config.ring_degree > 1
-            or parallel_config.context_parallel_config.ulysses_degree > 1
-        )
-        return supports_context_parallel and is_degree_greater_than_1
+        supports_context_parallel = backend.value in cls._supports_context_parallel
+        return supports_context_parallel
+
+
+@dataclass
+class _HubKernelConfig:
+    """Configuration for downloading and using a hub-based attention kernel."""
+
+    repo_id: str
+    function_attr: str
+    revision: Optional[str] = None
+    kernel_fn: Optional[Callable] = None
+
+
+# Registry for hub-based attention kernels
+_HUB_KERNELS_REGISTRY: Dict["AttentionBackendName", _HubKernelConfig] = {
+    # TODO: temporary revision for now. Remove when merged upstream into `main`.
+    AttentionBackendName._FLASH_3_HUB: _HubKernelConfig(
+        repo_id="kernels-community/flash-attn3", function_attr="flash_attn_func", revision="fake-ops-return-probs"
+    ),
+    AttentionBackendName.FLASH_HUB: _HubKernelConfig(
+        repo_id="kernels-community/flash-attn2", function_attr="flash_attn_func", revision=None
+    ),
+    AttentionBackendName.SAGE_HUB: _HubKernelConfig(
+        repo_id="kernels-community/sage_attention", function_attr="sageattn", revision=None
+    ),
+}
 
 
 @contextlib.contextmanager
@@ -306,14 +316,6 @@ def dispatch_attention_fn(
         backend_name = AttentionBackendName(backend)
         backend_fn = _AttentionBackendRegistry._backends.get(backend_name)
 
-    if parallel_config is not None and not _AttentionBackendRegistry._is_context_parallel_enabled(
-        backend_name, parallel_config
-    ):
-        raise ValueError(
-            f"Backend {backend_name} either does not support context parallelism or context parallelism "
-            f"was enabled with a world size of 1."
-        )
-
     kwargs = {
         "query": query,
         "key": key,
@@ -392,12 +394,18 @@ def _check_shape(
     attn_mask: Optional[torch.Tensor] = None,
     **kwargs,
 ) -> None:
+    # Expected shapes:
+    # query: (batch_size, seq_len_q, num_heads, head_dim)
+    # key:   (batch_size, seq_len_kv, num_heads, head_dim)
+    # value: (batch_size, seq_len_kv, num_heads, head_dim)
+    # attn_mask: (seq_len_q, seq_len_kv) or (batch_size, seq_len_q, seq_len_kv)
+    #            or (batch_size, num_heads, seq_len_q, seq_len_kv)
     if query.shape[-1] != key.shape[-1]:
-        raise ValueError("Query and key must have the same last dimension.")
-    if query.shape[-2] != value.shape[-2]:
-        raise ValueError("Query and value must have the same second to last dimension.")
-    if attn_mask is not None and attn_mask.shape[-1] != key.shape[-2]:
-        raise ValueError("Attention mask must match the key's second to last dimension.")
+        raise ValueError("Query and key must have the same head dimension.")
+    if key.shape[-3] != value.shape[-3]:
+        raise ValueError("Key and value must have the same sequence length.")
+    if attn_mask is not None and attn_mask.shape[-1] != key.shape[-3]:
+        raise ValueError("Attention mask must match the key's sequence length.")
 
 
 # ===== Helper functions =====
@@ -416,15 +424,11 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
                 f"Flash Attention 3 backend '{backend.value}' is not usable because of missing package or the version is too old. Please build FA3 beta release from source."
             )
 
-    # TODO: add support Hub variant of FA3 varlen later
-    elif backend in [AttentionBackendName._FLASH_3_HUB]:
-        if not DIFFUSERS_ENABLE_HUB_KERNELS:
-            raise RuntimeError(
-                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `DIFFUSERS_ENABLE_HUB_KERNELS` env var isn't set. Please set it like `export DIFFUSERS_ENABLE_HUB_KERNELS=yes`."
-            )
+    # TODO: add support Hub variant of varlen later
+    elif backend in [AttentionBackendName._FLASH_3_HUB, AttentionBackendName.FLASH_HUB, AttentionBackendName.SAGE_HUB]:
         if not is_kernels_available():
             raise RuntimeError(
-                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
+                f"Backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
             )
 
     elif backend == AttentionBackendName.AITER:
@@ -574,6 +578,29 @@ def _flex_attention_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
     return q_idx >= kv_idx
 
 
+# ===== Helpers for downloading kernels =====
+def _maybe_download_kernel_for_backend(backend: AttentionBackendName) -> None:
+    if backend not in _HUB_KERNELS_REGISTRY:
+        return
+    config = _HUB_KERNELS_REGISTRY[backend]
+
+    if config.kernel_fn is not None:
+        return
+
+    try:
+        from kernels import get_kernel
+
+        kernel_module = get_kernel(config.repo_id, revision=config.revision)
+        kernel_func = getattr(kernel_module, config.function_attr)
+
+        # Cache the downloaded kernel function in the config object
+        config.kernel_fn = kernel_func
+
+    except Exception as e:
+        logger.error(f"An error occurred while fetching kernel '{config.repo_id}' from the Hub: {e}")
+        raise
+
+
 # ===== torch op registrations =====
 # Registrations are required for fullgraph tracing compatibility
 # TODO: this is only required because the beta release FA3 does not have it. There is a PR adding
@@ -649,6 +676,86 @@ def _(
 # ===== Helper functions to use attention backends with templated CP autograd functions =====
 
 
+def _native_attention_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: Optional["ParallelConfig"] = None,
+):
+    # Native attention does not return_lse
+    if return_lse:
+        raise ValueError("Native attention does not support return_lse=True")
+
+    # used for backward pass
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value)
+        ctx.attn_mask = attn_mask
+        ctx.dropout_p = dropout_p
+        ctx.is_causal = is_causal
+        ctx.scale = scale
+        ctx.enable_gqa = enable_gqa
+
+    query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+    out = torch.nn.functional.scaled_dot_product_attention(
+        query=query,
+        key=key,
+        value=value,
+        attn_mask=attn_mask,
+        dropout_p=dropout_p,
+        is_causal=is_causal,
+        scale=scale,
+        enable_gqa=enable_gqa,
+    )
+    out = out.permute(0, 2, 1, 3)
+
+    return out
+
+
+def _native_attention_backward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    grad_out: torch.Tensor,
+    *args,
+    **kwargs,
+):
+    query, key, value = ctx.saved_tensors
+
+    query.requires_grad_(True)
+    key.requires_grad_(True)
+    value.requires_grad_(True)
+
+    query_t, key_t, value_t = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+    out = torch.nn.functional.scaled_dot_product_attention(
+        query=query_t,
+        key=key_t,
+        value=value_t,
+        attn_mask=ctx.attn_mask,
+        dropout_p=ctx.dropout_p,
+        is_causal=ctx.is_causal,
+        scale=ctx.scale,
+        enable_gqa=ctx.enable_gqa,
+    )
+    out = out.permute(0, 2, 1, 3)
+
+    grad_out_t = grad_out.permute(0, 2, 1, 3)
+    grad_query_t, grad_key_t, grad_value_t = torch.autograd.grad(
+        outputs=out, inputs=[query_t, key_t, value_t], grad_outputs=grad_out_t, retain_graph=False
+    )
+
+    grad_query = grad_query_t.permute(0, 2, 1, 3)
+    grad_key = grad_key_t.permute(0, 2, 1, 3)
+    grad_value = grad_value_t.permute(0, 2, 1, 3)
+
+    return grad_query, grad_key, grad_value
+
+
 # https://github.com/pytorch/pytorch/blob/8904ba638726f8c9a5aff5977c4aa76c9d2edfa6/aten/src/ATen/native/native_functions.yaml#L14958
 # forward declaration:
 #   aten::_scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, bool compute_log_sumexp, float dropout_p=0., bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
@@ -1247,6 +1354,38 @@ def _flash_attention(
     return (out, lse) if return_lse else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName.FLASH_HUB,
+    constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+    supports_context_parallel=False,
+)
+def _flash_attention_hub(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    return_lse: bool = False,
+    _parallel_config: Optional["ParallelConfig"] = None,
+) -> torch.Tensor:
+    lse = None
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        dropout_p=dropout_p,
+        softmax_scale=scale,
+        causal=is_causal,
+        return_attn_probs=return_lse,
+    )
+    if return_lse:
+        out, lse, *_ = out
+
+    return (out, lse) if return_lse else out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName.FLASH_VARLEN,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
@@ -1328,6 +1467,7 @@ def _flash_attention_3(
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_3_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+    supports_context_parallel=False,
 )
 def _flash_attention_3_hub(
     query: torch.Tensor,
@@ -1341,7 +1481,11 @@ def _flash_attention_3_hub(
     return_attn_probs: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    out = flash_attn_3_func_hub(
+    if _parallel_config:
+        raise NotImplementedError(f"{AttentionBackendName._FLASH_3_HUB.value} is not implemented for parallelism yet.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
+    out = func(
         q=query,
         k=key,
         v=value,
@@ -1523,6 +1667,7 @@ def _native_flex_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.NATIVE,
     constraints=[_check_device, _check_shape],
+    supports_context_parallel=True,
 )
 def _native_attention(
     query: torch.Tensor,
@@ -1538,18 +1683,35 @@ def _native_attention(
 ) -> torch.Tensor:
     if return_lse:
         raise ValueError("Native attention backend does not support setting `return_lse=True`.")
-    query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
-    out = torch.nn.functional.scaled_dot_product_attention(
-        query=query,
-        key=key,
-        value=value,
-        attn_mask=attn_mask,
-        dropout_p=dropout_p,
-        is_causal=is_causal,
-        scale=scale,
-        enable_gqa=enable_gqa,
-    )
-    out = out.permute(0, 2, 1, 3)
+    if _parallel_config is None:
+        query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+        out = torch.nn.functional.scaled_dot_product_attention(
+            query=query,
+            key=key,
+            value=value,
+            attn_mask=attn_mask,
+            dropout_p=dropout_p,
+            is_causal=is_causal,
+            scale=scale,
+            enable_gqa=enable_gqa,
+        )
+        out = out.permute(0, 2, 1, 3)
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            attn_mask,
+            dropout_p,
+            is_causal,
+            scale,
+            enable_gqa,
+            return_lse,
+            forward_op=_native_attention_forward_op,
+            backward_op=_native_attention_backward_op,
+            _parallel_config=_parallel_config,
+        )
+
     return out
 
 
@@ -1816,6 +1978,38 @@ def _sage_attention(
     return (out, lse) if return_lse else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName.SAGE_HUB,
+    constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+    supports_context_parallel=False,
+)
+def _sage_attention_hub(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    return_lse: bool = False,
+    _parallel_config: Optional["ParallelConfig"] = None,
+) -> torch.Tensor:
+    lse = None
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB].kernel_fn
+    if _parallel_config is None:
+        out = func(
+            q=query,
+            k=key,
+            v=value,
+            tensor_layout="NHD",
+            is_causal=is_causal,
+            sm_scale=scale,
+            return_lse=return_lse,
+        )
+        if return_lse:
+            out, lse, *_ = out
+
+    return (out, lse) if return_lse else out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName.SAGE_VARLEN,
     constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],

src/diffusers/models/auto_model.py

@@ -147,14 +147,13 @@ class AutoModel(ConfigMixin):
             "force_download",
             "local_files_only",
             "proxies",
-            "resume_download",
             "revision",
             "token",
         ]
         hub_kwargs = {name: kwargs.pop(name, None) for name in hub_kwargs_names}
 
         # load_config_kwargs uses the same hub kwargs minus subfolder and resume_download
-        load_config_kwargs = {k: v for k, v in hub_kwargs.items() if k not in ["subfolder", "resume_download"]}
+        load_config_kwargs = {k: v for k, v in hub_kwargs.items() if k not in ["subfolder"]}
 
         library = None
         orig_class_name = None
@@ -205,7 +204,6 @@ class AutoModel(ConfigMixin):
                 module_file=module_file,
                 class_name=class_name,
                 **hub_kwargs,
-                **kwargs,
             )
         else:
             from ..pipelines.pipeline_loading_utils import ALL_IMPORTABLE_CLASSES, get_class_obj_and_candidates

src/diffusers/models/autoencoders/__init__.py

@@ -4,6 +4,7 @@ from .autoencoder_kl import AutoencoderKL
 from .autoencoder_kl_allegro import AutoencoderKLAllegro
 from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
 from .autoencoder_kl_cosmos import AutoencoderKLCosmos
+from .autoencoder_kl_flux2 import AutoencoderKLFlux2
 from .autoencoder_kl_hunyuan_video import AutoencoderKLHunyuanVideo
 from .autoencoder_kl_hunyuanimage import AutoencoderKLHunyuanImage
 from .autoencoder_kl_hunyuanimage_refiner import AutoencoderKLHunyuanImageRefiner

src/diffusers/models/autoencoders/autoencoder_dc.py

@@ -102,7 +102,7 @@ def get_block(
     attention_head_dim: int,
     norm_type: str,
     act_fn: str,
-    qkv_mutliscales: Tuple[int] = (),
+    qkv_mutliscales: Tuple[int, ...] = (),
 ):
     if block_type == "ResBlock":
         block = ResBlock(in_channels, out_channels, norm_type, act_fn)
@@ -206,8 +206,8 @@ class Encoder(nn.Module):
         latent_channels: int,
         attention_head_dim: int = 32,
         block_type: Union[str, Tuple[str]] = "ResBlock",
-        block_out_channels: Tuple[int] = (128, 256, 512, 512, 1024, 1024),
-        layers_per_block: Tuple[int] = (2, 2, 2, 2, 2, 2),
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: Tuple[int, ...] = (2, 2, 2, 2, 2, 2),
         qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (), (5,), (5,), (5,)),
         downsample_block_type: str = "pixel_unshuffle",
         out_shortcut: bool = True,
@@ -292,8 +292,8 @@ class Decoder(nn.Module):
         latent_channels: int,
         attention_head_dim: int = 32,
         block_type: Union[str, Tuple[str]] = "ResBlock",
-        block_out_channels: Tuple[int] = (128, 256, 512, 512, 1024, 1024),
-        layers_per_block: Tuple[int] = (2, 2, 2, 2, 2, 2),
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: Tuple[int, ...] = (2, 2, 2, 2, 2, 2),
         qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (), (5,), (5,), (5,)),
         norm_type: Union[str, Tuple[str]] = "rms_norm",
         act_fn: Union[str, Tuple[str]] = "silu",
@@ -440,8 +440,8 @@ class AutoencoderDC(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModel
         decoder_block_types: Union[str, Tuple[str]] = "ResBlock",
         encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512, 1024, 1024),
         decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512, 1024, 1024),
-        encoder_layers_per_block: Tuple[int] = (2, 2, 2, 3, 3, 3),
-        decoder_layers_per_block: Tuple[int] = (3, 3, 3, 3, 3, 3),
+        encoder_layers_per_block: Tuple[int, ...] = (2, 2, 2, 3, 3, 3),
+        decoder_layers_per_block: Tuple[int, ...] = (3, 3, 3, 3, 3, 3),
         encoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (), (5,), (5,), (5,)),
         decoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (), (5,), (5,), (5,)),
         upsample_block_type: str = "pixel_shuffle",

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -78,9 +78,9 @@ class AutoencoderKL(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModel
         self,
         in_channels: int = 3,
         out_channels: int = 3,
-        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
-        up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
-        block_out_channels: Tuple[int] = (64,),
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
         layers_per_block: int = 1,
         act_fn: str = "silu",
         latent_channels: int = 4,

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -995,19 +995,19 @@ class AutoencoderKLCogVideoX(ModelMixin, AutoencoderMixin, ConfigMixin, FromOrig
         self,
         in_channels: int = 3,
         out_channels: int = 3,
-        down_block_types: Tuple[str] = (
+        down_block_types: Tuple[str, ...] = (
             "CogVideoXDownBlock3D",
             "CogVideoXDownBlock3D",
             "CogVideoXDownBlock3D",
             "CogVideoXDownBlock3D",
         ),
-        up_block_types: Tuple[str] = (
+        up_block_types: Tuple[str, ...] = (
             "CogVideoXUpBlock3D",
             "CogVideoXUpBlock3D",
             "CogVideoXUpBlock3D",
             "CogVideoXUpBlock3D",
         ),
-        block_out_channels: Tuple[int] = (128, 256, 256, 512),
+        block_out_channels: Tuple[int, ...] = (128, 256, 256, 512),
         latent_channels: int = 16,
         layers_per_block: int = 3,
         act_fn: str = "silu",

src/diffusers/models/autoencoders/autoencoder_kl_flux2.py

@@ -0,0 +1,546 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ...loaders.single_file_model import FromOriginalModelMixin
+from ...utils import deprecate
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+    FusedAttnProcessor2_0,
+)
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import AutoencoderMixin, Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class AutoencoderKLFlux2(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without losing too much precision in which case `force_upcast`
+            can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+        mid_block_add_attention (`bool`, *optional*, default to `True`):
+            If enabled, the mid_block of the Encoder and Decoder will have attention blocks. If set to false, the
+            mid_block will only have resnet blocks
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = (
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+        ),
+        up_block_types: Tuple[str, ...] = (
+            "UpDecoderBlock2D",
+            "UpDecoderBlock2D",
+            "UpDecoderBlock2D",
+            "UpDecoderBlock2D",
+        ),
+        block_out_channels: Tuple[int, ...] = (
+            128,
+            256,
+            512,
+            512,
+        ),
+        layers_per_block: int = 2,
+        act_fn: str = "silu",
+        latent_channels: int = 32,
+        norm_num_groups: int = 32,
+        sample_size: int = 1024,  # YiYi notes: not sure
+        force_upcast: bool = True,
+        use_quant_conv: bool = True,
+        use_post_quant_conv: bool = True,
+        mid_block_add_attention: bool = True,
+        batch_norm_eps: float = 1e-4,
+        batch_norm_momentum: float = 0.1,
+        patch_size: Tuple[int, int] = (2, 2),
+    ):
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+        # pass init params to Decoder
+        self.decoder = Decoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            norm_num_groups=norm_num_groups,
+            act_fn=act_fn,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) if use_quant_conv else None
+        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) if use_post_quant_conv else None
+
+        self.bn = nn.BatchNorm2d(
+            math.prod(patch_size) * latent_channels,
+            eps=batch_norm_eps,
+            momentum=batch_norm_momentum,
+            affine=False,
+            track_running_stats=True,
+        )
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # only relevant if vae tiling is enabled
+        self.tile_sample_min_size = self.config.sample_size
+        sample_size = (
+            self.config.sample_size[0]
+            if isinstance(self.config.sample_size, (list, tuple))
+            else self.config.sample_size
+        )
+        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor = 0.25
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def _encode(self, x: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, height, width = x.shape
+
+        if self.use_tiling and (width > self.tile_sample_min_size or height > self.tile_sample_min_size):
+            return self._tiled_encode(x)
+
+        enc = self.encoder(x)
+        if self.quant_conv is not None:
+            enc = self.quant_conv(enc)
+
+        return enc
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self._encode(x)
+
+        posterior = DiagonalGaussianDistribution(h)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        if self.post_quant_conv is not None:
+            z = self.post_quant_conv(z)
+
+        dec = self.decoder(z)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(
+        self, z: torch.FloatTensor, return_dict: bool = True, generator=None
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+        return b
+
+    def _tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+
+        Returns:
+            `torch.Tensor`:
+                The latent representation of the encoded videos.
+        """
+
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split the image into 512x512 tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[2], overlap_size):
+            row = []
+            for j in range(0, x.shape[3], overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                if self.config.use_quant_conv:
+                    tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        enc = torch.cat(result_rows, dim=2)
+        return enc
+
+    def tiled_encode(self, x: torch.Tensor, return_dict: bool = True) -> AutoencoderKLOutput:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
+                `tuple` is returned.
+        """
+        deprecation_message = (
+            "The tiled_encode implementation supporting the `return_dict` parameter is deprecated. In the future, the "
+            "implementation of this method will be replaced with that of `_tiled_encode` and you will no longer be able "
+            "to pass `return_dict`. You will also have to create a `DiagonalGaussianDistribution()` from the returned value."
+        )
+        deprecate("tiled_encode", "1.0.0", deprecation_message, standard_warn=False)
+
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split the image into 512x512 tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[2], overlap_size):
+            row = []
+            for j in range(0, x.shape[3], overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                if self.config.use_quant_conv:
+                    tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        moments = torch.cat(result_rows, dim=2)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_sample_min_size - blend_extent
+
+        # Split z into overlapping 64x64 tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, z.shape[2], overlap_size):
+            row = []
+            for j in range(0, z.shape[3], overlap_size):
+                tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
+                if self.config.use_post_quant_conv:
+                    tile = self.post_quant_conv(tile)
+                decoded = self.decoder(tile)
+                row.append(decoded)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        dec = torch.cat(result_rows, dim=2)
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Args:
+            sample (`torch.Tensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        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.
+
+        > [!WARNING] > This API is 🧪 experimental.
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        > [!WARNING] > This API is 🧪 experimental.
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)

src/diffusers/models/autoencoders/autoencoder_kl_hunyuan_video.py

@@ -653,7 +653,7 @@ class AutoencoderKLHunyuanVideo(ModelMixin, AutoencoderMixin, ConfigMixin):
             "HunyuanVideoUpBlock3D",
             "HunyuanVideoUpBlock3D",
         ),
-        block_out_channels: Tuple[int] = (128, 256, 512, 512),
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
         layers_per_block: int = 2,
         act_fn: str = "silu",
         norm_num_groups: int = 32,

src/diffusers/models/autoencoders/autoencoder_kl_hunyuanimage_refiner.py

@@ -601,7 +601,7 @@ class AutoencoderKLHunyuanImageRefiner(ModelMixin, ConfigMixin):
         in_channels: int = 3,
         out_channels: int = 3,
         latent_channels: int = 32,
-        block_out_channels: Tuple[int] = (128, 256, 512, 1024, 1024),
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024, 1024),
         layers_per_block: int = 2,
         spatial_compression_ratio: int = 16,
         temporal_compression_ratio: int = 4,

src/diffusers/models/autoencoders/autoencoder_kl_mochi.py

@@ -688,8 +688,8 @@ class AutoencoderKLMochi(ModelMixin, AutoencoderMixin, ConfigMixin):
         self,
         in_channels: int = 15,
         out_channels: int = 3,
-        encoder_block_out_channels: Tuple[int] = (64, 128, 256, 384),
-        decoder_block_out_channels: Tuple[int] = (128, 256, 512, 768),
+        encoder_block_out_channels: Tuple[int, ...] = (64, 128, 256, 384),
+        decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 768),
         latent_channels: int = 12,
         layers_per_block: Tuple[int, ...] = (3, 3, 4, 6, 3),
         act_fn: str = "silu",

src/diffusers/models/autoencoders/autoencoder_kl_qwenimage.py

@@ -16,7 +16,7 @@
 # QwenImageVAE is further fine-tuned from the Wan Video VAE to achieve improved performance.
 # For more information about the Wan VAE, please refer to:
 # - GitHub: https://github.com/Wan-Video/Wan2.1
-# - arXiv: https://arxiv.org/abs/2503.20314
+# - Paper: https://huggingface.co/papers/2503.20314
 
 from typing import List, Optional, Tuple, Union
 
@@ -679,7 +679,7 @@ class AutoencoderKLQwenImage(ModelMixin, AutoencoderMixin, ConfigMixin, FromOrig
         self,
         base_dim: int = 96,
         z_dim: int = 16,
-        dim_mult: Tuple[int] = [1, 2, 4, 4],
+        dim_mult: Tuple[int, ...] = (1, 2, 4, 4),
         num_res_blocks: int = 2,
         attn_scales: List[float] = [],
         temperal_downsample: List[bool] = [False, True, True],

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -31,7 +31,7 @@ class TemporalDecoder(nn.Module):
         self,
         in_channels: int = 4,
         out_channels: int = 3,
-        block_out_channels: Tuple[int] = (128, 256, 512, 512),
+        block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
         layers_per_block: int = 2,
     ):
         super().__init__()
@@ -172,8 +172,8 @@ class AutoencoderKLTemporalDecoder(ModelMixin, AutoencoderMixin, ConfigMixin):
         self,
         in_channels: int = 3,
         out_channels: int = 3,
-        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
-        block_out_channels: Tuple[int] = (64,),
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+        block_out_channels: Tuple[int, ...] = (64,),
         layers_per_block: int = 1,
         latent_channels: int = 4,
         sample_size: int = 32,

src/diffusers/models/autoencoders/autoencoder_kl_wan.py

@@ -971,7 +971,7 @@ class AutoencoderKLWan(ModelMixin, AutoencoderMixin, ConfigMixin, FromOriginalMo
         base_dim: int = 96,
         decoder_base_dim: Optional[int] = None,
         z_dim: int = 16,
-        dim_mult: Tuple[int] = [1, 2, 4, 4],
+        dim_mult: List[int] = [1, 2, 4, 4],
         num_res_blocks: int = 2,
         attn_scales: List[float] = [],
         temperal_downsample: List[bool] = [False, True, True],

src/diffusers/models/autoencoders/vae.py

@@ -286,11 +286,9 @@ class Decoder(nn.Module):
 
         sample = self.conv_in(sample)
 
-        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
         if torch.is_grad_enabled() and self.gradient_checkpointing:
             # middle
             sample = self._gradient_checkpointing_func(self.mid_block, sample, latent_embeds)
-            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks:
@@ -298,7 +296,6 @@ class Decoder(nn.Module):
         else:
             # middle
             sample = self.mid_block(sample, latent_embeds)
-            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks:

src/diffusers/models/controlnets/controlnet_xs.py

@@ -293,14 +293,14 @@ class ControlNetXSAdapter(ModelMixin, ConfigMixin):
         self,
         conditioning_channels: int = 3,
         conditioning_channel_order: str = "rgb",
-        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
+        conditioning_embedding_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
         time_embedding_mix: float = 1.0,
         learn_time_embedding: bool = False,
         num_attention_heads: Union[int, Tuple[int]] = 4,
-        block_out_channels: Tuple[int] = (4, 8, 16, 16),
-        base_block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        block_out_channels: Tuple[int, ...] = (4, 8, 16, 16),
+        base_block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
         cross_attention_dim: int = 1024,
-        down_block_types: Tuple[str] = (
+        down_block_types: Tuple[str, ...] = (
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
@@ -436,7 +436,7 @@ class ControlNetXSAdapter(ModelMixin, ConfigMixin):
         time_embedding_mix: int = 1.0,
         conditioning_channels: int = 3,
         conditioning_channel_order: str = "rgb",
-        conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
+        conditioning_embedding_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
     ):
         r"""
         Instantiate a [`ControlNetXSAdapter`] from a [`UNet2DConditionModel`].
@@ -529,14 +529,19 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
         self,
         # unet configs
         sample_size: Optional[int] = 96,
-        down_block_types: Tuple[str] = (
+        down_block_types: Tuple[str, ...] = (
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "CrossAttnDownBlock2D",
             "DownBlock2D",
         ),
-        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
-        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        up_block_types: Tuple[str, ...] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
         norm_num_groups: Optional[int] = 32,
         cross_attention_dim: Union[int, Tuple[int]] = 1024,
         transformer_layers_per_block: Union[int, Tuple[int]] = 1,
@@ -550,10 +555,10 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
         # additional controlnet configs
         time_embedding_mix: float = 1.0,
         ctrl_conditioning_channels: int = 3,
-        ctrl_conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
+        ctrl_conditioning_embedding_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
         ctrl_conditioning_channel_order: str = "rgb",
         ctrl_learn_time_embedding: bool = False,
-        ctrl_block_out_channels: Tuple[int] = (4, 8, 16, 16),
+        ctrl_block_out_channels: Tuple[int, ...] = (4, 8, 16, 16),
         ctrl_num_attention_heads: Union[int, Tuple[int]] = 4,
         ctrl_max_norm_num_groups: int = 32,
     ):

src/diffusers/models/modeling_utils.py

@@ -595,7 +595,11 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 attention as backend.
         """
         from .attention import AttentionModuleMixin
-        from .attention_dispatch import AttentionBackendName, _check_attention_backend_requirements
+        from .attention_dispatch import (
+            AttentionBackendName,
+            _check_attention_backend_requirements,
+            _maybe_download_kernel_for_backend,
+        )
 
         # TODO: the following will not be required when everything is refactored to AttentionModuleMixin
         from .attention_processor import Attention, MochiAttention
@@ -606,8 +610,10 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         available_backends = {x.value for x in AttentionBackendName.__members__.values()}
         if backend not in available_backends:
             raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
+
         backend = AttentionBackendName(backend)
         _check_attention_backend_requirements(backend)
+        _maybe_download_kernel_for_backend(backend)
 
         attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
         for module in self.modules():
@@ -1484,59 +1490,71 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         config: Union[ParallelConfig, ContextParallelConfig],
         cp_plan: Optional[Dict[str, ContextParallelModelPlan]] = None,
     ):
-        from ..hooks.context_parallel import apply_context_parallel
-        from .attention import AttentionModuleMixin
-        from .attention_processor import Attention, MochiAttention
-
         logger.warning(
             "`enable_parallelism` is an experimental feature. The API may change in the future and breaking changes may be introduced at any time without warning."
         )
 
+        if not torch.distributed.is_available() and not torch.distributed.is_initialized():
+            raise RuntimeError(
+                "torch.distributed must be available and initialized before calling `enable_parallelism`."
+            )
+
+        from ..hooks.context_parallel import apply_context_parallel
+        from .attention import AttentionModuleMixin
+        from .attention_dispatch import AttentionBackendName, _AttentionBackendRegistry
+        from .attention_processor import Attention, MochiAttention
+
         if isinstance(config, ContextParallelConfig):
             config = ParallelConfig(context_parallel_config=config)
 
-        if not torch.distributed.is_initialized():
-            raise RuntimeError("torch.distributed must be initialized before calling `enable_parallelism`.")
-
         rank = torch.distributed.get_rank()
         world_size = torch.distributed.get_world_size()
         device_type = torch._C._get_accelerator().type
         device_module = torch.get_device_module(device_type)
         device = torch.device(device_type, rank % device_module.device_count())
 
-        cp_mesh = None
+        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
+
+        if config.context_parallel_config is not None:
+            for module in self.modules():
+                if not isinstance(module, attention_classes):
+                    continue
+
+                processor = module.processor
+                if processor is None or not hasattr(processor, "_attention_backend"):
+                    continue
+
+                attention_backend = processor._attention_backend
+                if attention_backend is None:
+                    attention_backend, _ = _AttentionBackendRegistry.get_active_backend()
+                else:
+                    attention_backend = AttentionBackendName(attention_backend)
+
+                if not _AttentionBackendRegistry._is_context_parallel_available(attention_backend):
+                    compatible_backends = sorted(_AttentionBackendRegistry._supports_context_parallel)
+                    raise ValueError(
+                        f"Context parallelism is enabled but the attention processor '{processor.__class__.__name__}' "
+                        f"is using backend '{attention_backend.value}' which does not support context parallelism. "
+                        f"Please set a compatible attention backend: {compatible_backends} using `model.set_attention_backend()` before "
+                        f"calling `enable_parallelism()`."
+                    )
+
+                # All modules use the same attention processor and backend. We don't need to
+                # iterate over all modules after checking the first processor
+                break
+
+        mesh = None
         if config.context_parallel_config is not None:
             cp_config = config.context_parallel_config
-            if cp_config.ring_degree < 1 or cp_config.ulysses_degree < 1:
-                raise ValueError("`ring_degree` and `ulysses_degree` must be greater than or equal to 1.")
-            if cp_config.ring_degree > 1 and cp_config.ulysses_degree > 1:
-                raise ValueError(
-                    "Unified Ulysses-Ring attention is not yet supported. Please set either `ring_degree` or `ulysses_degree` to 1."
-                )
-            if cp_config.ring_degree * cp_config.ulysses_degree > world_size:
-                raise ValueError(
-                    f"The product of `ring_degree` ({cp_config.ring_degree}) and `ulysses_degree` ({cp_config.ulysses_degree}) must not exceed the world size ({world_size})."
-                )
-            cp_mesh = torch.distributed.device_mesh.init_device_mesh(
+            mesh = torch.distributed.device_mesh.init_device_mesh(
                 device_type=device_type,
-                mesh_shape=(cp_config.ring_degree, cp_config.ulysses_degree),
-                mesh_dim_names=("ring", "ulysses"),
+                mesh_shape=cp_config.mesh_shape,
+                mesh_dim_names=cp_config.mesh_dim_names,
             )
 
-        config.setup(rank, world_size, device, cp_mesh=cp_mesh)
-
-        if cp_plan is None and self._cp_plan is None:
-            raise ValueError(
-                "`cp_plan` must be provided either as an argument or set in the model's `_cp_plan` attribute."
-            )
-        cp_plan = cp_plan if cp_plan is not None else self._cp_plan
-
-        if config.context_parallel_config is not None:
-            apply_context_parallel(self, config.context_parallel_config, cp_plan)
-
+        config.setup(rank, world_size, device, mesh=mesh)
         self._parallel_config = config
 
-        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
         for module in self.modules():
             if not isinstance(module, attention_classes):
                 continue
@@ -1545,6 +1563,14 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 continue
             processor._parallel_config = config
 
+        if config.context_parallel_config is not None:
+            if cp_plan is None and self._cp_plan is None:
+                raise ValueError(
+                    "`cp_plan` must be provided either as an argument or set in the model's `_cp_plan` attribute."
+                )
+            cp_plan = cp_plan if cp_plan is not None else self._cp_plan
+            apply_context_parallel(self, config.context_parallel_config, cp_plan)
+
     @classmethod
     def _load_pretrained_model(
         cls,

src/diffusers/models/transformers/__init__.py

@@ -18,12 +18,15 @@ if is_torch_available():
     from .transformer_2d import Transformer2DModel
     from .transformer_allegro import AllegroTransformer3DModel
     from .transformer_bria import BriaTransformer2DModel
+    from .transformer_bria_fibo import BriaFiboTransformer2DModel
     from .transformer_chroma import ChromaTransformer2DModel
+    from .transformer_chronoedit import ChronoEditTransformer3DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_cogview4 import CogView4Transformer2DModel
     from .transformer_cosmos import CosmosTransformer3DModel
     from .transformer_easyanimate import EasyAnimateTransformer3DModel
     from .transformer_flux import FluxTransformer2DModel
+    from .transformer_flux2 import Flux2Transformer2DModel
     from .transformer_hidream_image import HiDreamImageTransformer2DModel
     from .transformer_hunyuan_video import HunyuanVideoTransformer3DModel
     from .transformer_hunyuan_video_framepack import HunyuanVideoFramepackTransformer3DModel
@@ -35,8 +38,11 @@ if is_torch_available():
     from .transformer_omnigen import OmniGenTransformer2DModel
     from .transformer_prx import PRXTransformer2DModel
     from .transformer_qwenimage import QwenImageTransformer2DModel
+    from .transformer_sana_video import SanaVideoTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel
     from .transformer_wan import WanTransformer3DModel
+    from .transformer_wan_animate import WanAnimateTransformer3DModel
     from .transformer_wan_vace import WanVACETransformer3DModel
+    from .transformer_z_image import ZImageTransformer2DModel

src/diffusers/models/transformers/transformer_bria_fibo.py

@@ -0,0 +1,655 @@
+# Copyright (c) Bria.ai. All rights reserved.
+#
+# This file is licensed under the Creative Commons Attribution-NonCommercial 4.0 International Public License (CC-BY-NC-4.0).
+# You may obtain a copy of the license at https://creativecommons.org/licenses/by-nc/4.0/
+#
+# You are free to share and adapt this material for non-commercial purposes provided you give appropriate credit,
+# indicate if changes were made, and do not use the material for commercial purposes.
+#
+# See the license for further details.
+import inspect
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...models.attention_processor import Attention
+from ...models.embeddings import TimestepEmbedding, apply_rotary_emb, get_1d_rotary_pos_embed, get_timestep_embedding
+from ...models.modeling_outputs import Transformer2DModelOutput
+from ...models.modeling_utils import ModelMixin
+from ...models.transformers.transformer_bria import BriaAttnProcessor
+from ...utils import (
+    USE_PEFT_BACKEND,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import AttentionModuleMixin, FeedForward
+from ..attention_dispatch import dispatch_attention_fn
+from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def _get_projections(attn: "BriaFiboAttention", hidden_states, encoder_hidden_states=None):
+    query = attn.to_q(hidden_states)
+    key = attn.to_k(hidden_states)
+    value = attn.to_v(hidden_states)
+
+    encoder_query = encoder_key = encoder_value = None
+    if encoder_hidden_states is not None and attn.added_kv_proj_dim is not None:
+        encoder_query = attn.add_q_proj(encoder_hidden_states)
+        encoder_key = attn.add_k_proj(encoder_hidden_states)
+        encoder_value = attn.add_v_proj(encoder_hidden_states)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_fused_projections(attn: "BriaFiboAttention", hidden_states, encoder_hidden_states=None):
+    query, key, value = attn.to_qkv(hidden_states).chunk(3, dim=-1)
+
+    encoder_query = encoder_key = encoder_value = (None,)
+    if encoder_hidden_states is not None and hasattr(attn, "to_added_qkv"):
+        encoder_query, encoder_key, encoder_value = attn.to_added_qkv(encoder_hidden_states).chunk(3, dim=-1)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_qkv_projections(attn: "BriaFiboAttention", hidden_states, encoder_hidden_states=None):
+    if attn.fused_projections:
+        return _get_fused_projections(attn, hidden_states, encoder_hidden_states)
+    return _get_projections(attn, hidden_states, encoder_hidden_states)
+
+
+# Copied from diffusers.models.transformers.transformer_flux.FluxAttnProcessor with FluxAttnProcessor->BriaFiboAttnProcessor, FluxAttention->BriaFiboAttention
+class BriaFiboAttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
+
+    def __call__(
+        self,
+        attn: "BriaFiboAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
+            attn, hidden_states, encoder_hidden_states
+        )
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if attn.added_kv_proj_dim is not None:
+            encoder_query = encoder_query.unflatten(-1, (attn.heads, -1))
+            encoder_key = encoder_key.unflatten(-1, (attn.heads, -1))
+            encoder_value = encoder_value.unflatten(-1, (attn.heads, -1))
+
+            encoder_query = attn.norm_added_q(encoder_query)
+            encoder_key = attn.norm_added_k(encoder_key)
+
+            query = torch.cat([encoder_query, query], dim=1)
+            key = torch.cat([encoder_key, key], dim=1)
+            value = torch.cat([encoder_value, value], dim=1)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+            hidden_states = attn.to_out[0](hidden_states)
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
+# Based on https://github.com/huggingface/diffusers/blob/55d49d4379007740af20629bb61aba9546c6b053/src/diffusers/models/transformers/transformer_flux.py
+class BriaFiboAttention(torch.nn.Module, AttentionModuleMixin):
+    _default_processor_cls = BriaFiboAttnProcessor
+    _available_processors = [BriaFiboAttnProcessor]
+
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        out_bias: bool = True,
+        eps: float = 1e-5,
+        out_dim: int = None,
+        context_pre_only: Optional[bool] = None,
+        pre_only: bool = False,
+        elementwise_affine: bool = True,
+        processor=None,
+    ):
+        super().__init__()
+
+        self.head_dim = dim_head
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.dropout = dropout
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.context_pre_only = context_pre_only
+        self.pre_only = pre_only
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.added_proj_bias = added_proj_bias
+
+        self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.to_q = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_k = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_v = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+
+        if not self.pre_only:
+            self.to_out = torch.nn.ModuleList([])
+            self.to_out.append(torch.nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
+            self.to_out.append(torch.nn.Dropout(dropout))
+
+        if added_kv_proj_dim is not None:
+            self.norm_added_q = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.norm_added_k = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.add_q_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.to_add_out = torch.nn.Linear(self.inner_dim, query_dim, bias=out_bias)
+
+        if processor is None:
+            processor = self._default_processor_cls()
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"}
+        unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
+        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs)
+
+
+class BriaFiboEmbedND(torch.nn.Module):
+    # modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        cos_out = []
+        sin_out = []
+        pos = ids.float()
+        is_mps = ids.device.type == "mps"
+        freqs_dtype = torch.float32 if is_mps else torch.float64
+        for i in range(n_axes):
+            cos, sin = get_1d_rotary_pos_embed(
+                self.axes_dim[i],
+                pos[:, i],
+                theta=self.theta,
+                repeat_interleave_real=True,
+                use_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            cos_out.append(cos)
+            sin_out.append(sin)
+        freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
+        freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
+        return freqs_cos, freqs_sin
+
+
+@maybe_allow_in_graph
+class BriaFiboSingleTransformerBlock(nn.Module):
+    def __init__(self, dim: int, num_attention_heads: int, attention_head_dim: int, mlp_ratio: float = 4.0):
+        super().__init__()
+        self.mlp_hidden_dim = int(dim * mlp_ratio)
+
+        self.norm = AdaLayerNormZeroSingle(dim)
+        self.proj_mlp = nn.Linear(dim, self.mlp_hidden_dim)
+        self.act_mlp = nn.GELU(approximate="tanh")
+        self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
+
+        processor = BriaAttnProcessor()
+
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            processor=processor,
+            qk_norm="rms_norm",
+            eps=1e-6,
+            pre_only=True,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+        gate = gate.unsqueeze(1)
+        hidden_states = gate * self.proj_out(hidden_states)
+        hidden_states = residual + hidden_states
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return hidden_states
+
+
+class BriaFiboTextProjection(nn.Module):
+    def __init__(self, in_features, hidden_size):
+        super().__init__()
+        self.linear = nn.Linear(in_features=in_features, out_features=hidden_size, bias=False)
+
+    def forward(self, caption):
+        hidden_states = self.linear(caption)
+        return hidden_states
+
+
+@maybe_allow_in_graph
+# Based on from diffusers.models.transformers.transformer_flux.FluxTransformerBlock
+class BriaFiboTransformerBlock(nn.Module):
+    def __init__(
+        self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str = "rms_norm", eps: float = 1e-6
+    ):
+        super().__init__()
+
+        self.norm1 = AdaLayerNormZero(dim)
+        self.norm1_context = AdaLayerNormZero(dim)
+
+        self.attn = BriaFiboAttention(
+            query_dim=dim,
+            added_kv_proj_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            context_pre_only=False,
+            bias=True,
+            processor=BriaFiboAttnProcessor(),
+            eps=eps,
+        )
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+
+        norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+            encoder_hidden_states, emb=temb
+        )
+        joint_attention_kwargs = joint_attention_kwargs or {}
+
+        # Attention.
+        attention_outputs = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        if len(attention_outputs) == 2:
+            attn_output, context_attn_output = attention_outputs
+        elif len(attention_outputs) == 3:
+            attn_output, context_attn_output, ip_attn_output = attention_outputs
+
+        # Process attention outputs for the `hidden_states`.
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = hidden_states + ff_output
+        if len(attention_outputs) == 3:
+            hidden_states = hidden_states + ip_attn_output
+
+        # Process attention outputs for the `encoder_hidden_states`.
+        context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+        encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+        return encoder_hidden_states, hidden_states
+
+
+class BriaFiboTimesteps(nn.Module):
+    def __init__(
+        self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float, scale: int = 1, time_theta=10000
+    ):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+        self.downscale_freq_shift = downscale_freq_shift
+        self.scale = scale
+        self.time_theta = time_theta
+
+    def forward(self, timesteps):
+        t_emb = get_timestep_embedding(
+            timesteps,
+            self.num_channels,
+            flip_sin_to_cos=self.flip_sin_to_cos,
+            downscale_freq_shift=self.downscale_freq_shift,
+            scale=self.scale,
+            max_period=self.time_theta,
+        )
+        return t_emb
+
+
+class BriaFiboTimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim, time_theta):
+        super().__init__()
+
+        self.time_proj = BriaFiboTimesteps(
+            num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, time_theta=time_theta
+        )
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(self, timestep, dtype):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=dtype))  # (N, D)
+        return timesteps_emb
+
+
+class BriaFiboTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+    """
+    Parameters:
+        patch_size (`int`): Patch size to turn the input data into small patches.
+        in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
+        num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
+        num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
+        attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
+        num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
+        joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
+        guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
+        ...
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 1,
+        in_channels: int = 64,
+        num_layers: int = 19,
+        num_single_layers: int = 38,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 4096,
+        pooled_projection_dim: int = None,
+        guidance_embeds: bool = False,
+        axes_dims_rope: List[int] = [16, 56, 56],
+        rope_theta=10000,
+        time_theta=10000,
+        text_encoder_dim: int = 2048,
+    ):
+        super().__init__()
+        self.out_channels = in_channels
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+
+        self.pos_embed = BriaFiboEmbedND(theta=rope_theta, axes_dim=axes_dims_rope)
+
+        self.time_embed = BriaFiboTimestepProjEmbeddings(embedding_dim=self.inner_dim, time_theta=time_theta)
+
+        if guidance_embeds:
+            self.guidance_embed = BriaFiboTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+
+        self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
+        self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BriaFiboTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.config.num_attention_heads,
+                    attention_head_dim=self.config.attention_head_dim,
+                )
+                for i in range(self.config.num_layers)
+            ]
+        )
+
+        self.single_transformer_blocks = nn.ModuleList(
+            [
+                BriaFiboSingleTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.config.num_attention_heads,
+                    attention_head_dim=self.config.attention_head_dim,
+                )
+                for i in range(self.config.num_single_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+
+        self.gradient_checkpointing = False
+
+        caption_projection = [
+            BriaFiboTextProjection(in_features=text_encoder_dim, hidden_size=self.inner_dim // 2)
+            for i in range(self.config.num_layers + self.config.num_single_layers)
+        ]
+        self.caption_projection = nn.ModuleList(caption_projection)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        text_encoder_layers: list = None,
+        pooled_projections: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_ids: torch.Tensor = None,
+        txt_ids: torch.Tensor = None,
+        guidance: torch.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
+        """
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
+                from the embeddings of input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+        hidden_states = self.x_embedder(hidden_states)
+
+        timestep = timestep.to(hidden_states.dtype)
+        if guidance is not None:
+            guidance = guidance.to(hidden_states.dtype)
+        else:
+            guidance = None
+
+        temb = self.time_embed(timestep, dtype=hidden_states.dtype)
+
+        if guidance:
+            temb += self.guidance_embed(guidance, dtype=hidden_states.dtype)
+
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        if len(txt_ids.shape) == 3:
+            txt_ids = txt_ids[0]
+
+        if len(img_ids.shape) == 3:
+            img_ids = img_ids[0]
+
+        ids = torch.cat((txt_ids, img_ids), dim=0)
+        image_rotary_emb = self.pos_embed(ids)
+
+        new_text_encoder_layers = []
+        for i, text_encoder_layer in enumerate(text_encoder_layers):
+            text_encoder_layer = self.caption_projection[i](text_encoder_layer)
+            new_text_encoder_layers.append(text_encoder_layer)
+        text_encoder_layers = new_text_encoder_layers
+
+        block_id = 0
+        for index_block, block in enumerate(self.transformer_blocks):
+            current_text_encoder_layer = text_encoder_layers[block_id]
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states[:, :, : self.inner_dim // 2], current_text_encoder_layer], dim=-1
+            )
+            block_id += 1
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    image_rotary_emb,
+                    joint_attention_kwargs,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+        for index_block, block in enumerate(self.single_transformer_blocks):
+            current_text_encoder_layer = text_encoder_layers[block_id]
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states[:, :, : self.inner_dim // 2], current_text_encoder_layer], dim=-1
+            )
+            block_id += 1
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    temb,
+                    image_rotary_emb,
+                    joint_attention_kwargs,
+                )
+
+            else:
+                hidden_states = block(
+                    hidden_states=hidden_states,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+            encoder_hidden_states = hidden_states[:, : encoder_hidden_states.shape[1], ...]
+            hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+
+        hidden_states = self.norm_out(hidden_states, temb)
+        output = self.proj_out(hidden_states)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_chronoedit.py

@@ -0,0 +1,739 @@
+# Copyright 2025 The ChronoEdit Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import maybe_allow_in_graph
+from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
+from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
+from ..attention_dispatch import dispatch_attention_fn
+from ..cache_utils import CacheMixin
+from ..embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import FP32LayerNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.models.transformers.transformer_wan._get_qkv_projections
+def _get_qkv_projections(attn: "WanAttention", hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor):
+    # encoder_hidden_states is only passed for cross-attention
+    if encoder_hidden_states is None:
+        encoder_hidden_states = hidden_states
+
+    if attn.fused_projections:
+        if attn.cross_attention_dim_head is None:
+            # In self-attention layers, we can fuse the entire QKV projection into a single linear
+            query, key, value = attn.to_qkv(hidden_states).chunk(3, dim=-1)
+        else:
+            # In cross-attention layers, we can only fuse the KV projections into a single linear
+            query = attn.to_q(hidden_states)
+            key, value = attn.to_kv(encoder_hidden_states).chunk(2, dim=-1)
+    else:
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+    return query, key, value
+
+
+# Copied from diffusers.models.transformers.transformer_wan._get_added_kv_projections
+def _get_added_kv_projections(attn: "WanAttention", encoder_hidden_states_img: torch.Tensor):
+    if attn.fused_projections:
+        key_img, value_img = attn.to_added_kv(encoder_hidden_states_img).chunk(2, dim=-1)
+    else:
+        key_img = attn.add_k_proj(encoder_hidden_states_img)
+        value_img = attn.add_v_proj(encoder_hidden_states_img)
+    return key_img, value_img
+
+
+# modified from diffusers.models.transformers.transformer_wan.WanAttnProcessor
+class WanAttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "WanAttnProcessor requires PyTorch 2.0. To use it, please upgrade PyTorch to version 2.0 or higher."
+            )
+
+    def __call__(
+        self,
+        attn: "WanAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            # 512 is the context length of the text encoder, hardcoded for now
+            image_context_length = encoder_hidden_states.shape[1] - 512
+            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
+            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]
+
+        query, key, value = _get_qkv_projections(attn, hidden_states, encoder_hidden_states)
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(
+                hidden_states: torch.Tensor,
+                freqs_cos: torch.Tensor,
+                freqs_sin: torch.Tensor,
+            ):
+                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                cos = freqs_cos[..., 0::2]
+                sin = freqs_sin[..., 1::2]
+                out = torch.empty_like(hidden_states)
+                out[..., 0::2] = x1 * cos - x2 * sin
+                out[..., 1::2] = x1 * sin + x2 * cos
+                return out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, *rotary_emb)
+            key = apply_rotary_emb(key, *rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img, value_img = _get_added_kv_projections(attn, encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1))
+            value_img = value_img.unflatten(2, (attn.heads, -1))
+
+            hidden_states_img = dispatch_attention_fn(
+                query,
+                key_img,
+                value_img,
+                attn_mask=None,
+                dropout_p=0.0,
+                is_causal=False,
+                backend=self._attention_backend,
+                # Reference: https://github.com/huggingface/diffusers/pull/12660
+                parallel_config=None,
+            )
+            hidden_states_img = hidden_states_img.flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            # Reference: https://github.com/huggingface/diffusers/pull/12660
+            parallel_config=(self._parallel_config if encoder_hidden_states is None else None),
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanAttnProcessor2_0
+class WanAttnProcessor2_0:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = (
+            "The WanAttnProcessor2_0 class is deprecated and will be removed in a future version. "
+            "Please use WanAttnProcessor instead. "
+        )
+        deprecate("WanAttnProcessor2_0", "1.0.0", deprecation_message, standard_warn=False)
+        return WanAttnProcessor(*args, **kwargs)
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanAttention
+class WanAttention(torch.nn.Module, AttentionModuleMixin):
+    _default_processor_cls = WanAttnProcessor
+    _available_processors = [WanAttnProcessor]
+
+    def __init__(
+        self,
+        dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        eps: float = 1e-5,
+        dropout: float = 0.0,
+        added_kv_proj_dim: Optional[int] = None,
+        cross_attention_dim_head: Optional[int] = None,
+        processor=None,
+        is_cross_attention=None,
+    ):
+        super().__init__()
+
+        self.inner_dim = dim_head * heads
+        self.heads = heads
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.cross_attention_dim_head = cross_attention_dim_head
+        self.kv_inner_dim = self.inner_dim if cross_attention_dim_head is None else cross_attention_dim_head * heads
+
+        self.to_q = torch.nn.Linear(dim, self.inner_dim, bias=True)
+        self.to_k = torch.nn.Linear(dim, self.kv_inner_dim, bias=True)
+        self.to_v = torch.nn.Linear(dim, self.kv_inner_dim, bias=True)
+        self.to_out = torch.nn.ModuleList(
+            [
+                torch.nn.Linear(self.inner_dim, dim, bias=True),
+                torch.nn.Dropout(dropout),
+            ]
+        )
+        self.norm_q = torch.nn.RMSNorm(dim_head * heads, eps=eps, elementwise_affine=True)
+        self.norm_k = torch.nn.RMSNorm(dim_head * heads, eps=eps, elementwise_affine=True)
+
+        self.add_k_proj = self.add_v_proj = None
+        if added_kv_proj_dim is not None:
+            self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=True)
+            self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=True)
+            self.norm_added_k = torch.nn.RMSNorm(dim_head * heads, eps=eps)
+
+        self.is_cross_attention = cross_attention_dim_head is not None
+
+        self.set_processor(processor)
+
+    def fuse_projections(self):
+        if getattr(self, "fused_projections", False):
+            return
+
+        if self.cross_attention_dim_head is None:
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_qkv = nn.Linear(in_features, out_features, bias=True)
+            self.to_qkv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+        else:
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_kv = nn.Linear(in_features, out_features, bias=True)
+            self.to_kv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+
+        if self.added_kv_proj_dim is not None:
+            concatenated_weights = torch.cat([self.add_k_proj.weight.data, self.add_v_proj.weight.data])
+            concatenated_bias = torch.cat([self.add_k_proj.bias.data, self.add_v_proj.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_added_kv = nn.Linear(in_features, out_features, bias=True)
+            self.to_added_kv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+
+        self.fused_projections = True
+
+    @torch.no_grad()
+    def unfuse_projections(self):
+        if not getattr(self, "fused_projections", False):
+            return
+
+        if hasattr(self, "to_qkv"):
+            delattr(self, "to_qkv")
+        if hasattr(self, "to_kv"):
+            delattr(self, "to_kv")
+        if hasattr(self, "to_added_kv"):
+            delattr(self, "to_added_kv")
+
+        self.fused_projections = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, rotary_emb, **kwargs)
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanImageEmbedding
+class WanImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int, pos_embed_seq_len=None):
+        super().__init__()
+
+        self.norm1 = FP32LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = FP32LayerNorm(out_features)
+        if pos_embed_seq_len is not None:
+            self.pos_embed = nn.Parameter(torch.zeros(1, pos_embed_seq_len, in_features))
+        else:
+            self.pos_embed = None
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        if self.pos_embed is not None:
+            batch_size, seq_len, embed_dim = encoder_hidden_states_image.shape
+            encoder_hidden_states_image = encoder_hidden_states_image.view(-1, 2 * seq_len, embed_dim)
+            encoder_hidden_states_image = encoder_hidden_states_image + self.pos_embed
+
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanTimeTextImageEmbedding
+class WanTimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+        pos_embed_seq_len: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = WanImageEmbedding(image_embed_dim, dim, pos_embed_seq_len=pos_embed_seq_len)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        timestep_seq_len: Optional[int] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+        if timestep_seq_len is not None:
+            timestep = timestep.unflatten(0, (-1, timestep_seq_len))
+
+        time_embedder_dtype = next(iter(self.time_embedder.parameters())).dtype
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class ChronoEditRotaryPosEmbed(nn.Module):
+    def __init__(
+        self,
+        attention_head_dim: int,
+        patch_size: Tuple[int, int, int],
+        max_seq_len: int,
+        theta: float = 10000.0,
+        temporal_skip_len: int = 8,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+        self.temporal_skip_len = temporal_skip_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+        freqs_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+
+        freqs_cos = []
+        freqs_sin = []
+
+        for dim in [t_dim, h_dim, w_dim]:
+            freq_cos, freq_sin = get_1d_rotary_pos_embed(
+                dim,
+                max_seq_len,
+                theta,
+                use_real=True,
+                repeat_interleave_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            freqs_cos.append(freq_cos)
+            freqs_sin.append(freq_sin)
+
+        self.register_buffer("freqs_cos", torch.cat(freqs_cos, dim=1), persistent=False)
+        self.register_buffer("freqs_sin", torch.cat(freqs_sin, dim=1), persistent=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        split_sizes = [
+            self.attention_head_dim - 2 * (self.attention_head_dim // 3),
+            self.attention_head_dim // 3,
+            self.attention_head_dim // 3,
+        ]
+
+        freqs_cos = self.freqs_cos.split(split_sizes, dim=1)
+        freqs_sin = self.freqs_sin.split(split_sizes, dim=1)
+
+        if num_frames == 2:
+            freqs_cos_f = freqs_cos[0][: self.temporal_skip_len][[0, -1]].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        else:
+            freqs_cos_f = freqs_cos[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_h = freqs_cos[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_w = freqs_cos[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        if num_frames == 2:
+            freqs_sin_f = freqs_sin[0][: self.temporal_skip_len][[0, -1]].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        else:
+            freqs_sin_f = freqs_sin[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_h = freqs_sin[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_w = freqs_sin[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_cos = torch.cat([freqs_cos_f, freqs_cos_h, freqs_cos_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+        freqs_sin = torch.cat([freqs_sin_f, freqs_sin_h, freqs_sin_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+
+        return freqs_cos, freqs_sin
+
+
+@maybe_allow_in_graph
+# Copied from diffusers.models.transformers.transformer_wan.WanTransformerBlock
+class WanTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = WanAttention(
+            dim=dim,
+            heads=num_heads,
+            dim_head=dim // num_heads,
+            eps=eps,
+            cross_attention_dim_head=None,
+            processor=WanAttnProcessor(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = WanAttention(
+            dim=dim,
+            heads=num_heads,
+            dim_head=dim // num_heads,
+            eps=eps,
+            added_kv_proj_dim=added_kv_proj_dim,
+            cross_attention_dim_head=dim // num_heads,
+            processor=WanAttnProcessor(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        if temb.ndim == 4:
+            # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table.unsqueeze(0) + temb.float()
+            ).chunk(6, dim=2)
+            # batch_size, seq_len, 1, inner_dim
+            shift_msa = shift_msa.squeeze(2)
+            scale_msa = scale_msa.squeeze(2)
+            gate_msa = gate_msa.squeeze(2)
+            c_shift_msa = c_shift_msa.squeeze(2)
+            c_scale_msa = c_scale_msa.squeeze(2)
+            c_gate_msa = c_gate_msa.squeeze(2)
+        else:
+            # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table + temb.float()
+            ).chunk(6, dim=1)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(norm_hidden_states, encoder_hidden_states, None, None)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+        return hidden_states
+
+
+# modified from diffusers.models.transformers.transformer_wan.WanTransformer3DModel
+class ChronoEditTransformer3DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin, AttentionMixin
+):
+    r"""
+    A Transformer model for video-like data used in the ChronoEdit model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["WanTransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+    _repeated_blocks = ["WanTransformerBlock"]
+    _cp_plan = {
+        "rope": {
+            0: ContextParallelInput(split_dim=1, expected_dims=4, split_output=True),
+            1: ContextParallelInput(split_dim=1, expected_dims=4, split_output=True),
+        },
+        "blocks.0": {
+            "hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
+        },
+        # Reference: https://github.com/huggingface/diffusers/pull/12660
+        # We need to disable the splitting of encoder_hidden_states because
+        # the image_encoder consistently generates 257 tokens for image_embed. This causes
+        # the shape of encoder_hidden_states—whose token count is always 769 (512 + 257)
+        # after concatenation—to be indivisible by the number of devices in the CP.
+        "proj_out": ContextParallelOutput(gather_dim=1, expected_dims=3),
+    }
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        pos_embed_seq_len: Optional[int] = None,
+        rope_temporal_skip_len: int = 8,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = ChronoEditRotaryPosEmbed(
+            attention_head_dim, patch_size, rope_max_seq_len, temporal_skip_len=rope_temporal_skip_len
+        )
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+            pos_embed_seq_len=pos_embed_seq_len,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                WanTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        # timestep shape: batch_size, or batch_size, seq_len (wan 2.2 ti2v)
+        if timestep.ndim == 2:
+            ts_seq_len = timestep.shape[1]
+            timestep = timestep.flatten()  # batch_size * seq_len
+        else:
+            ts_seq_len = None
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image, timestep_seq_len=ts_seq_len
+        )
+        if ts_seq_len is not None:
+            # batch_size, seq_len, 6, inner_dim
+            timestep_proj = timestep_proj.unflatten(2, (6, -1))
+        else:
+            # batch_size, 6, inner_dim
+            timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        if temb.ndim == 3:
+            # batch_size, seq_len, inner_dim (wan 2.2 ti2v)
+            shift, scale = (self.scale_shift_table.unsqueeze(0).to(temb.device) + temb.unsqueeze(2)).chunk(2, dim=2)
+            shift = shift.squeeze(2)
+            scale = scale.squeeze(2)
+        else:
+            # batch_size, inner_dim
+            shift, scale = (self.scale_shift_table.to(temb.device) + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_flux2.py

@@ -0,0 +1,908 @@
+# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, is_torch_npu_available, logging, scale_lora_layers, unscale_lora_layers
+from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
+from ..attention import AttentionMixin, AttentionModuleMixin
+from ..attention_dispatch import dispatch_attention_fn
+from ..cache_utils import CacheMixin
+from ..embeddings import (
+    TimestepEmbedding,
+    Timesteps,
+    apply_rotary_emb,
+    get_1d_rotary_pos_embed,
+)
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormContinuous
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def _get_projections(attn: "Flux2Attention", hidden_states, encoder_hidden_states=None):
+    query = attn.to_q(hidden_states)
+    key = attn.to_k(hidden_states)
+    value = attn.to_v(hidden_states)
+
+    encoder_query = encoder_key = encoder_value = None
+    if encoder_hidden_states is not None and attn.added_kv_proj_dim is not None:
+        encoder_query = attn.add_q_proj(encoder_hidden_states)
+        encoder_key = attn.add_k_proj(encoder_hidden_states)
+        encoder_value = attn.add_v_proj(encoder_hidden_states)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_fused_projections(attn: "Flux2Attention", hidden_states, encoder_hidden_states=None):
+    query, key, value = attn.to_qkv(hidden_states).chunk(3, dim=-1)
+
+    encoder_query = encoder_key = encoder_value = (None,)
+    if encoder_hidden_states is not None and hasattr(attn, "to_added_qkv"):
+        encoder_query, encoder_key, encoder_value = attn.to_added_qkv(encoder_hidden_states).chunk(3, dim=-1)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_qkv_projections(attn: "Flux2Attention", hidden_states, encoder_hidden_states=None):
+    if attn.fused_projections:
+        return _get_fused_projections(attn, hidden_states, encoder_hidden_states)
+    return _get_projections(attn, hidden_states, encoder_hidden_states)
+
+
+class Flux2SwiGLU(nn.Module):
+    """
+    Flux 2 uses a SwiGLU-style activation in the transformer feedforward sub-blocks, but with the linear projection
+    layer fused into the first linear layer of the FF sub-block. Thus, this module has no trainable parameters.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.gate_fn = nn.SiLU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        x = self.gate_fn(x1) * x2
+        return x
+
+
+class Flux2FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: float = 3.0,
+        inner_dim: Optional[int] = None,
+        bias: bool = False,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out or dim
+
+        # Flux2SwiGLU will reduce the dimension by half
+        self.linear_in = nn.Linear(dim, inner_dim * 2, bias=bias)
+        self.act_fn = Flux2SwiGLU()
+        self.linear_out = nn.Linear(inner_dim, dim_out, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.linear_in(x)
+        x = self.act_fn(x)
+        x = self.linear_out(x)
+        return x
+
+
+class Flux2AttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
+
+    def __call__(
+        self,
+        attn: "Flux2Attention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
+            attn, hidden_states, encoder_hidden_states
+        )
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if attn.added_kv_proj_dim is not None:
+            encoder_query = encoder_query.unflatten(-1, (attn.heads, -1))
+            encoder_key = encoder_key.unflatten(-1, (attn.heads, -1))
+            encoder_value = encoder_value.unflatten(-1, (attn.heads, -1))
+
+            encoder_query = attn.norm_added_q(encoder_query)
+            encoder_key = attn.norm_added_k(encoder_key)
+
+            query = torch.cat([encoder_query, query], dim=1)
+            key = torch.cat([encoder_key, key], dim=1)
+            value = torch.cat([encoder_value, value], dim=1)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
+class Flux2Attention(torch.nn.Module, AttentionModuleMixin):
+    _default_processor_cls = Flux2AttnProcessor
+    _available_processors = [Flux2AttnProcessor]
+
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        out_bias: bool = True,
+        eps: float = 1e-5,
+        out_dim: int = None,
+        elementwise_affine: bool = True,
+        processor=None,
+    ):
+        super().__init__()
+
+        self.head_dim = dim_head
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+
+        self.use_bias = bias
+        self.dropout = dropout
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.added_proj_bias = added_proj_bias
+
+        self.to_q = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_k = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_v = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+
+        # QK Norm
+        self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+
+        self.to_out = torch.nn.ModuleList([])
+        self.to_out.append(torch.nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
+        self.to_out.append(torch.nn.Dropout(dropout))
+
+        if added_kv_proj_dim is not None:
+            self.norm_added_q = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.norm_added_k = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.add_q_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.to_add_out = torch.nn.Linear(self.inner_dim, query_dim, bias=out_bias)
+
+        if processor is None:
+            processor = self._default_processor_cls()
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
+        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs)
+
+
+class Flux2ParallelSelfAttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
+
+    def __call__(
+        self,
+        attn: "Flux2ParallelSelfAttention",
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Parallel in (QKV + MLP in) projection
+        hidden_states = attn.to_qkv_mlp_proj(hidden_states)
+        qkv, mlp_hidden_states = torch.split(
+            hidden_states, [3 * attn.inner_dim, attn.mlp_hidden_dim * attn.mlp_mult_factor], dim=-1
+        )
+
+        # Handle the attention logic
+        query, key, value = qkv.chunk(3, dim=-1)
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Handle the feedforward (FF) logic
+        mlp_hidden_states = attn.mlp_act_fn(mlp_hidden_states)
+
+        # Concatenate and parallel output projection
+        hidden_states = torch.cat([hidden_states, mlp_hidden_states], dim=-1)
+        hidden_states = attn.to_out(hidden_states)
+
+        return hidden_states
+
+
+class Flux2ParallelSelfAttention(torch.nn.Module, AttentionModuleMixin):
+    """
+    Flux 2 parallel self-attention for the Flux 2 single-stream transformer blocks.
+
+    This implements a parallel transformer block, where the attention QKV projections are fused to the feedforward (FF)
+    input projections, and the attention output projections are fused to the FF output projections. See the [ViT-22B
+    paper](https://arxiv.org/abs/2302.05442) for a visual depiction of this type of transformer block.
+    """
+
+    _default_processor_cls = Flux2ParallelSelfAttnProcessor
+    _available_processors = [Flux2ParallelSelfAttnProcessor]
+    # Does not support QKV fusion as the QKV projections are always fused
+    _supports_qkv_fusion = False
+
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        out_bias: bool = True,
+        eps: float = 1e-5,
+        out_dim: int = None,
+        elementwise_affine: bool = True,
+        mlp_ratio: float = 4.0,
+        mlp_mult_factor: int = 2,
+        processor=None,
+    ):
+        super().__init__()
+
+        self.head_dim = dim_head
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+
+        self.use_bias = bias
+        self.dropout = dropout
+
+        self.mlp_ratio = mlp_ratio
+        self.mlp_hidden_dim = int(query_dim * self.mlp_ratio)
+        self.mlp_mult_factor = mlp_mult_factor
+
+        # Fused QKV projections + MLP input projection
+        self.to_qkv_mlp_proj = torch.nn.Linear(
+            self.query_dim, self.inner_dim * 3 + self.mlp_hidden_dim * self.mlp_mult_factor, bias=bias
+        )
+        self.mlp_act_fn = Flux2SwiGLU()
+
+        # QK Norm
+        self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+
+        # Fused attention output projection + MLP output projection
+        self.to_out = torch.nn.Linear(self.inner_dim + self.mlp_hidden_dim, self.out_dim, bias=out_bias)
+
+        if processor is None:
+            processor = self._default_processor_cls()
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
+        return self.processor(self, hidden_states, attention_mask, image_rotary_emb, **kwargs)
+
+
+class Flux2SingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float = 3.0,
+        eps: float = 1e-6,
+        bias: bool = False,
+    ):
+        super().__init__()
+
+        self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+
+        # Note that the MLP in/out linear layers are fused with the attention QKV/out projections, respectively; this
+        # is often called a "parallel" transformer block. See the [ViT-22B paper](https://arxiv.org/abs/2302.05442)
+        # for a visual depiction of this type of transformer block.
+        self.attn = Flux2ParallelSelfAttention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=bias,
+            out_bias=bias,
+            eps=eps,
+            mlp_ratio=mlp_ratio,
+            mlp_mult_factor=2,
+            processor=Flux2ParallelSelfAttnProcessor(),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor],
+        temb_mod_params: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        split_hidden_states: bool = False,
+        text_seq_len: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # If encoder_hidden_states is None, hidden_states is assumed to have encoder_hidden_states already
+        # concatenated
+        if encoder_hidden_states is not None:
+            text_seq_len = encoder_hidden_states.shape[1]
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        mod_shift, mod_scale, mod_gate = temb_mod_params
+
+        norm_hidden_states = self.norm(hidden_states)
+        norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift
+
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        hidden_states = hidden_states + mod_gate * attn_output
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        if split_hidden_states:
+            encoder_hidden_states, hidden_states = hidden_states[:, :text_seq_len], hidden_states[:, text_seq_len:]
+            return encoder_hidden_states, hidden_states
+        else:
+            return hidden_states
+
+
+class Flux2TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float = 3.0,
+        eps: float = 1e-6,
+        bias: bool = False,
+    ):
+        super().__init__()
+        self.mlp_hidden_dim = int(dim * mlp_ratio)
+
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.norm1_context = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+
+        self.attn = Flux2Attention(
+            query_dim=dim,
+            added_kv_proj_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=bias,
+            added_proj_bias=bias,
+            out_bias=bias,
+            eps=eps,
+            processor=Flux2AttnProcessor(),
+        )
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.ff = Flux2FeedForward(dim=dim, dim_out=dim, mult=mlp_ratio, bias=bias)
+
+        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.ff_context = Flux2FeedForward(dim=dim, dim_out=dim, mult=mlp_ratio, bias=bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb_mod_params_img: Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
+        temb_mod_params_txt: Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        joint_attention_kwargs = joint_attention_kwargs or {}
+
+        # Modulation parameters shape: [1, 1, self.dim]
+        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = temb_mod_params_img
+        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = temb_mod_params_txt
+
+        # Img stream
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = (1 + scale_msa) * norm_hidden_states + shift_msa
+
+        # Conditioning txt stream
+        norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states)
+        norm_encoder_hidden_states = (1 + c_scale_msa) * norm_encoder_hidden_states + c_shift_msa
+
+        # Attention on concatenated img + txt stream
+        attention_outputs = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        attn_output, context_attn_output = attention_outputs
+
+        # Process attention outputs for the image stream (`hidden_states`).
+        attn_output = gate_msa * attn_output
+        hidden_states = hidden_states + attn_output
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states)
+        hidden_states = hidden_states + gate_mlp * ff_output
+
+        # Process attention outputs for the text stream (`encoder_hidden_states`).
+        context_attn_output = c_gate_msa * context_attn_output
+        encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp) + c_shift_mlp
+
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp * context_ff_output
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+        return encoder_hidden_states, hidden_states
+
+
+class Flux2PosEmbed(nn.Module):
+    # modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        # Expected ids shape: [S, len(self.axes_dim)]
+        cos_out = []
+        sin_out = []
+        pos = ids.float()
+        is_mps = ids.device.type == "mps"
+        is_npu = ids.device.type == "npu"
+        freqs_dtype = torch.float32 if (is_mps or is_npu) else torch.float64
+        # Unlike Flux 1, loop over len(self.axes_dim) rather than ids.shape[-1]
+        for i in range(len(self.axes_dim)):
+            cos, sin = get_1d_rotary_pos_embed(
+                self.axes_dim[i],
+                pos[..., i],
+                theta=self.theta,
+                repeat_interleave_real=True,
+                use_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            cos_out.append(cos)
+            sin_out.append(sin)
+        freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
+        freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
+        return freqs_cos, freqs_sin
+
+
+class Flux2TimestepGuidanceEmbeddings(nn.Module):
+    def __init__(self, in_channels: int = 256, embedding_dim: int = 6144, bias: bool = False):
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=in_channels, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias
+        )
+
+        self.guidance_embedder = TimestepEmbedding(
+            in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias
+        )
+
+    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor) -> torch.Tensor:
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(timestep.dtype))  # (N, D)
+
+        guidance_proj = self.time_proj(guidance)
+        guidance_emb = self.guidance_embedder(guidance_proj.to(guidance.dtype))  # (N, D)
+
+        time_guidance_emb = timesteps_emb + guidance_emb
+
+        return time_guidance_emb
+
+
+class Flux2Modulation(nn.Module):
+    def __init__(self, dim: int, mod_param_sets: int = 2, bias: bool = False):
+        super().__init__()
+        self.mod_param_sets = mod_param_sets
+
+        self.linear = nn.Linear(dim, dim * 3 * self.mod_param_sets, bias=bias)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, temb: torch.Tensor) -> Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
+        mod = self.act_fn(temb)
+        mod = self.linear(mod)
+
+        if mod.ndim == 2:
+            mod = mod.unsqueeze(1)
+        mod_params = torch.chunk(mod, 3 * self.mod_param_sets, dim=-1)
+        # Return tuple of 3-tuples of modulation params shift/scale/gate
+        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(self.mod_param_sets))
+
+
+class Flux2Transformer2DModel(
+    ModelMixin,
+    ConfigMixin,
+    PeftAdapterMixin,
+    FromOriginalModelMixin,
+    FluxTransformer2DLoadersMixin,
+    CacheMixin,
+    AttentionMixin,
+):
+    """
+    The Transformer model introduced in Flux 2.
+
+    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
+
+    Args:
+        patch_size (`int`, defaults to `1`):
+            Patch size to turn the input data into small patches.
+        in_channels (`int`, defaults to `128`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `None`):
+            The number of channels in the output. If not specified, it defaults to `in_channels`.
+        num_layers (`int`, defaults to `8`):
+            The number of layers of dual stream DiT blocks to use.
+        num_single_layers (`int`, defaults to `48`):
+            The number of layers of single stream DiT blocks to use.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of dimensions to use for each attention head.
+        num_attention_heads (`int`, defaults to `48`):
+            The number of attention heads to use.
+        joint_attention_dim (`int`, defaults to `15360`):
+            The number of dimensions to use for the joint attention (embedding/channel dimension of
+            `encoder_hidden_states`).
+        pooled_projection_dim (`int`, defaults to `768`):
+            The number of dimensions to use for the pooled projection.
+        guidance_embeds (`bool`, defaults to `True`):
+            Whether to use guidance embeddings for guidance-distilled variant of the model.
+        axes_dims_rope (`Tuple[int]`, defaults to `(32, 32, 32, 32)`):
+            The dimensions to use for the rotary positional embeddings.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["Flux2TransformerBlock", "Flux2SingleTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+    _repeated_blocks = ["Flux2TransformerBlock", "Flux2SingleTransformerBlock"]
+    _cp_plan = {
+        "": {
+            "hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
+            "encoder_hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
+            "img_ids": ContextParallelInput(split_dim=0, expected_dims=2, split_output=False),
+            "txt_ids": ContextParallelInput(split_dim=0, expected_dims=2, split_output=False),
+        },
+        "proj_out": ContextParallelOutput(gather_dim=1, expected_dims=3),
+    }
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 1,
+        in_channels: int = 128,
+        out_channels: Optional[int] = None,
+        num_layers: int = 8,
+        num_single_layers: int = 48,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 48,
+        joint_attention_dim: int = 15360,
+        timestep_guidance_channels: int = 256,
+        mlp_ratio: float = 3.0,
+        axes_dims_rope: Tuple[int, ...] = (32, 32, 32, 32),
+        rope_theta: int = 2000,
+        eps: float = 1e-6,
+    ):
+        super().__init__()
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Sinusoidal positional embedding for RoPE on image and text tokens
+        self.pos_embed = Flux2PosEmbed(theta=rope_theta, axes_dim=axes_dims_rope)
+
+        # 2. Combined timestep + guidance embedding
+        self.time_guidance_embed = Flux2TimestepGuidanceEmbeddings(
+            in_channels=timestep_guidance_channels, embedding_dim=self.inner_dim, bias=False
+        )
+
+        # 3. Modulation (double stream and single stream blocks share modulation parameters, resp.)
+        # Two sets of shift/scale/gate modulation parameters for the double stream attn and FF sub-blocks
+        self.double_stream_modulation_img = Flux2Modulation(self.inner_dim, mod_param_sets=2, bias=False)
+        self.double_stream_modulation_txt = Flux2Modulation(self.inner_dim, mod_param_sets=2, bias=False)
+        # Only one set of modulation parameters as the attn and FF sub-blocks are run in parallel for single stream
+        self.single_stream_modulation = Flux2Modulation(self.inner_dim, mod_param_sets=1, bias=False)
+
+        # 4. Input projections
+        self.x_embedder = nn.Linear(in_channels, self.inner_dim, bias=False)
+        self.context_embedder = nn.Linear(joint_attention_dim, self.inner_dim, bias=False)
+
+        # 5. Double Stream Transformer Blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                Flux2TransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    mlp_ratio=mlp_ratio,
+                    eps=eps,
+                    bias=False,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 6. Single Stream Transformer Blocks
+        self.single_transformer_blocks = nn.ModuleList(
+            [
+                Flux2SingleTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    mlp_ratio=mlp_ratio,
+                    eps=eps,
+                    bias=False,
+                )
+                for _ in range(num_single_layers)
+            ]
+        )
+
+        # 7. Output layers
+        self.norm_out = AdaLayerNormContinuous(
+            self.inner_dim, self.inner_dim, elementwise_affine=False, eps=eps, bias=False
+        )
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=False)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_ids: torch.Tensor = None,
+        txt_ids: torch.Tensor = None,
+        guidance: torch.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        """
+        The [`FluxTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
+                A list of tensors that if specified are added to the residuals of transformer blocks.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # 0. Handle input arguments
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        num_txt_tokens = encoder_hidden_states.shape[1]
+
+        # 1. Calculate timestep embedding and modulation parameters
+        timestep = timestep.to(hidden_states.dtype) * 1000
+        guidance = guidance.to(hidden_states.dtype) * 1000
+
+        temb = self.time_guidance_embed(timestep, guidance)
+
+        double_stream_mod_img = self.double_stream_modulation_img(temb)
+        double_stream_mod_txt = self.double_stream_modulation_txt(temb)
+        single_stream_mod = self.single_stream_modulation(temb)[0]
+
+        # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states)
+        hidden_states = self.x_embedder(hidden_states)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        # 3. Calculate RoPE embeddings from image and text tokens
+        # NOTE: the below logic means that we can't support batched inference with images of different resolutions or
+        # text prompts of differents lengths. Is this a use case we want to support?
+        if img_ids.ndim == 3:
+            img_ids = img_ids[0]
+        if txt_ids.ndim == 3:
+            txt_ids = txt_ids[0]
+
+        if is_torch_npu_available():
+            freqs_cos_image, freqs_sin_image = self.pos_embed(img_ids.cpu())
+            image_rotary_emb = (freqs_cos_image.npu(), freqs_sin_image.npu())
+            freqs_cos_text, freqs_sin_text = self.pos_embed(txt_ids.cpu())
+            text_rotary_emb = (freqs_cos_text.npu(), freqs_sin_text.npu())
+        else:
+            image_rotary_emb = self.pos_embed(img_ids)
+            text_rotary_emb = self.pos_embed(txt_ids)
+        concat_rotary_emb = (
+            torch.cat([text_rotary_emb[0], image_rotary_emb[0]], dim=0),
+            torch.cat([text_rotary_emb[1], image_rotary_emb[1]], dim=0),
+        )
+
+        # 4. Double Stream Transformer Blocks
+        for index_block, block in enumerate(self.transformer_blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    double_stream_mod_img,
+                    double_stream_mod_txt,
+                    concat_rotary_emb,
+                    joint_attention_kwargs,
+                )
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb_mod_params_img=double_stream_mod_img,
+                    temb_mod_params_txt=double_stream_mod_txt,
+                    image_rotary_emb=concat_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+        # Concatenate text and image streams for single-block inference
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        # 5. Single Stream Transformer Blocks
+        for index_block, block in enumerate(self.single_transformer_blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    None,
+                    single_stream_mod,
+                    concat_rotary_emb,
+                    joint_attention_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=None,
+                    temb_mod_params=single_stream_mod,
+                    image_rotary_emb=concat_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+        # Remove text tokens from concatenated stream
+        hidden_states = hidden_states[:, num_txt_tokens:, ...]
+
+        # 6. Output layers
+        hidden_states = self.norm_out(hidden_states, temb)
+        output = self.proj_out(hidden_states)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)