Merge branch 'main' into push-test-fixes

update to a place holder

.github/workflows/nightly_tests.yml

@@ -79,7 +79,7 @@ jobs:
           python utils/print_env.py
       - name: Pipeline CUDA Test
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
           CUBLAS_WORKSPACE_CONFIG: :16:8
         run: |
@@ -139,7 +139,7 @@ jobs:
     - name: Run nightly PyTorch CUDA tests for non-pipeline modules
       if: ${{ matrix.module != 'examples'}}
       env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
@@ -152,7 +152,7 @@ jobs:
     - name: Run nightly example tests with Torch
       if: ${{ matrix.module == 'examples' }}
       env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
         CUBLAS_WORKSPACE_CONFIG: :16:8
       run: |
@@ -209,7 +209,7 @@ jobs:
 
     - name: Run nightly Flax TPU tests
       env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
         python -m pytest -n 0 \
           -s -v -k "Flax" \
@@ -264,7 +264,7 @@ jobs:
 
     - name: Run Nightly ONNXRuntime CUDA tests
       env:
-        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
       run: |
         python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
           -s -v -k "Onnx" \

.github/workflows/push_tests.yml

@@ -1,4 +1,4 @@
-name: Slow Tests on main
+name: Fast GPU Tests on main
 
 on:
   push:

.github/workflows/release_tests_fast.yml

@@ -0,0 +1,389 @@
+# Duplicate workflow to push_tests.yml that is meant to run on release/patch branches as a final check
+# Creating a duplicate workflow here is simpler than adding complex path/branch parsing logic to push_tests.yml
+# Needs to be updated if push_tests.yml updated
+name: (Release) Fast GPU Tests on main
+
+on:
+  push:
+    branches:
+      - "v*.*.*-release"
+      - "v*.*.*-patch"
+
+env:
+  DIFFUSERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  PYTEST_TIMEOUT: 600
+  PIPELINE_USAGE_CUTOFF: 50000
+
+jobs:
+  setup_torch_cuda_pipeline_matrix:
+    name: Setup Torch Pipelines CUDA Slow Tests Matrix
+    runs-on:
+      group: aws-general-8-plus
+    container:
+      image: diffusers/diffusers-pytorch-cpu
+    outputs:
+      pipeline_test_matrix: ${{ steps.fetch_pipeline_matrix.outputs.pipeline_test_matrix }}
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: Install dependencies
+        run: |
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Fetch Pipeline Matrix
+        id: fetch_pipeline_matrix
+        run: |
+          matrix=$(python utils/fetch_torch_cuda_pipeline_test_matrix.py)
+          echo $matrix
+          echo "pipeline_test_matrix=$matrix" >> $GITHUB_OUTPUT
+      - name: Pipeline Tests Artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: test-pipelines.json
+          path: reports
+
+  torch_pipelines_cuda_tests:
+    name: Torch Pipelines CUDA Tests
+    needs: setup_torch_cuda_pipeline_matrix
+    strategy:
+      fail-fast: false
+      max-parallel: 8
+      matrix:
+        module: ${{ fromJson(needs.setup_torch_cuda_pipeline_matrix.outputs.pipeline_test_matrix) }}
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi
+      - name: Install dependencies
+        run: |
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+          python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Slow PyTorch CUDA checkpoint tests on Ubuntu
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+          CUBLAS_WORKSPACE_CONFIG: :16:8
+        run: |
+          python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+            -s -v -k "not Flax and not Onnx" \
+            --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+            tests/pipelines/${{ matrix.module }}
+      - name: Failure short reports
+        if: ${{ failure() }}
+        run: |
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_stats.txt
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_failures_short.txt
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: pipeline_${{ matrix.module }}_test_reports
+          path: reports
+
+  torch_cuda_tests:
+    name: Torch CUDA Tests
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
+    defaults:
+      run:
+        shell: bash
+    strategy:
+      fail-fast: false
+      max-parallel: 2
+      matrix:
+        module: [models, schedulers, lora, others, single_file]
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
+
+    - name: Environment
+      run: |
+        python utils/print_env.py
+
+    - name: Run PyTorch CUDA tests
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+        CUBLAS_WORKSPACE_CONFIG: :16:8
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "not Flax and not Onnx" \
+          --make-reports=tests_torch_cuda \
+          tests/${{ matrix.module }}
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_torch_cuda_stats.txt
+        cat reports/tests_torch_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: torch_cuda_test_reports
+        path: reports
+
+  flax_tpu_tests:
+    name: Flax TPU Tests
+    runs-on: docker-tpu
+    container:
+      image: diffusers/diffusers-flax-tpu
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/ --privileged
+    defaults:
+      run:
+        shell: bash
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+
+    - name: Environment
+      run: |
+        python utils/print_env.py
+
+    - name: Run slow Flax TPU tests
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        python -m pytest -n 0 \
+          -s -v -k "Flax" \
+          --make-reports=tests_flax_tpu \
+          tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_flax_tpu_stats.txt
+        cat reports/tests_flax_tpu_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: flax_tpu_test_reports
+        path: reports
+
+  onnx_cuda_tests:
+    name: ONNX CUDA Tests
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: diffusers/diffusers-onnxruntime-cuda
+      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/ --gpus 0
+    defaults:
+      run:
+        shell: bash
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
+
+    - name: Environment
+      run: |
+        python utils/print_env.py
+
+    - name: Run slow ONNXRuntime CUDA tests
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+          -s -v -k "Onnx" \
+          --make-reports=tests_onnx_cuda \
+          tests/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_onnx_cuda_stats.txt
+        cat reports/tests_onnx_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: onnx_cuda_test_reports
+        path: reports
+
+  run_torch_compile_tests:
+    name: PyTorch Compile CUDA tests
+
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-compile-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+    - name: Environment
+      run: |
+        python utils/print_env.py
+    - name: Run example tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        RUN_COMPILE: yes
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: cat reports/tests_torch_compile_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: torch_compile_test_reports
+        path: reports
+
+  run_xformers_tests:
+    name: PyTorch xformers CUDA tests
+
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-xformers-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+    - name: Environment
+      run: |
+        python utils/print_env.py
+    - name: Run example tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -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
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: torch_xformers_test_reports
+        path: reports
+
+  run_examples_tests:
+    name: Examples PyTorch CUDA tests on Ubuntu
+
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+
+    - name: Environment
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python utils/print_env.py
+
+    - name: Run example tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install timm
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/examples_torch_cuda_stats.txt
+        cat reports/examples_torch_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v2
+      with:
+        name: examples_test_reports
+        path: reports

CONTRIBUTING.md

@@ -57,7 +57,7 @@ Any question or comment related to the Diffusers library can be asked on the [di
 - ...
 
 Every question that is asked on the forum or on Discord actively encourages the community to publicly
-share knowledge and might very well help a beginner in the future that has the same question you're
+share knowledge and might very well help a beginner in the future who has the same question you're
 having. Please do pose any questions you might have.
 In the same spirit, you are of immense help to the community by answering such questions because this way you are publicly documenting knowledge for everybody to learn from.
 
@@ -503,4 +503,4 @@ $ git push --set-upstream origin your-branch-for-syncing
 
 ### Style guide
 
-For documentation strings, 🧨 Diffusers follows the [Google style](https://google.github.io/styleguide/pyguide.html).
+For documentation strings, 🧨 Diffusers follows the [Google style](https://google.github.io/styleguide/pyguide.html).

PHILOSOPHY.md

@@ -15,7 +15,7 @@ specific language governing permissions and limitations under the License.
 🧨 Diffusers provides **state-of-the-art** pretrained diffusion models across multiple modalities.
 Its purpose is to serve as a **modular toolbox** for both inference and training.
 
-We aim at building a library that stands the test of time and therefore take API design very seriously.
+We aim to build a library that stands the test of time and therefore take API design very seriously.
 
 In a nutshell, Diffusers is built to be a natural extension of PyTorch. Therefore, most of our design choices are based on [PyTorch's Design Principles](https://pytorch.org/docs/stable/community/design.html#pytorch-design-philosophy). Let's go over the most important ones:
 
@@ -107,4 +107,4 @@ The following design principles are followed:
 - Every scheduler exposes the timesteps to be "looped over" via a `timesteps` attribute, which is an array of timesteps the model will be called upon.
 - The `step(...)` function takes a predicted model output and the "current" sample (x_t) and returns the "previous", slightly more denoised sample (x_t-1).
 - Given the complexity of diffusion schedulers, the `step` function does not expose all the complexity and can be a bit of a "black box".
-- In almost all cases, novel schedulers shall be implemented in a new scheduling file.
+- In almost all cases, novel schedulers shall be implemented in a new scheduling file.

docs/source/en/_toctree.yml

@@ -226,6 +226,8 @@
     - sections:
       - local: api/models/controlnet
         title: ControlNetModel
+      - local: api/models/controlnet_flux
+        title: FluxControlNetModel
       - local: api/models/controlnet_hunyuandit
         title: HunyuanDiT2DControlNetModel
       - local: api/models/controlnet_sd3
@@ -320,6 +322,8 @@
       title: Consistency Models
     - local: api/pipelines/controlnet
       title: ControlNet
+    - local: api/pipelines/controlnet_flux
+      title: ControlNet with Flux.1
     - local: api/pipelines/controlnet_hunyuandit
       title: ControlNet with Hunyuan-DiT
     - local: api/pipelines/controlnet_sd3

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

@@ -0,0 +1,45 @@
+<!--Copyright 2024 The HuggingFace Team and The InstantX Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# FluxControlNetModel
+
+FluxControlNetModel is an implementation of ControlNet for Flux.1.
+
+The ControlNet model was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, Maneesh Agrawala. It provides a greater degree of control over text-to-image generation by conditioning the model on additional inputs such as edge maps, depth maps, segmentation maps, and keypoints for pose detection.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+## Loading from the original format
+
+By default the [`FluxControlNetModel`] should be loaded with [`~ModelMixin.from_pretrained`].
+
+```py
+from diffusers import FluxControlNetPipeline
+from diffusers.models import FluxControlNetModel, FluxMultiControlNetModel
+
+controlnet = FluxControlNetModel.from_pretrained("InstantX/FLUX.1-dev-Controlnet-Canny")
+pipe = FluxControlNetPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", controlnet=controlnet)
+
+controlnet = FluxControlNetModel.from_pretrained("InstantX/FLUX.1-dev-Controlnet-Canny")
+controlnet = FluxMultiControlNetModel([controlnet])
+pipe = FluxControlNetPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", controlnet=controlnet)
+```
+
+## FluxControlNetModel
+
+[[autodoc]] FluxControlNetModel
+
+## FluxControlNetOutput
+
+[[autodoc]] models.controlnet_flux.FluxControlNetOutput

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

@@ -29,6 +29,10 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
 
+There are two models available that can be used with the CogVideoX pipeline:
+- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b)
+- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b)
+
 ## Inference
 
 Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
@@ -68,15 +72,26 @@ With torch.compile(): Average inference time: 76.27 seconds.
 
 ### Memory optimization
 
-CogVideoX requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
+CogVideoX-2b requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
 
 - `pipe.enable_model_cpu_offload()`:
   - Without enabling cpu offloading, memory usage is `33 GB`
   - With enabling cpu offloading, memory usage is `19 GB`
+- `pipe.enable_sequential_cpu_offload()`:
+  - Similar to `enable_model_cpu_offload` but can significantly reduce memory usage at the cost of slow inference
+  - When enabled, memory usage is under `4 GB`
 - `pipe.vae.enable_tiling()`:
   - With enabling cpu offloading and tiling, memory usage is `11 GB`
 - `pipe.vae.enable_slicing()`
 
+### Quantized inference
+
+[torchao](https://github.com/pytorch/ao) and [optimum-quanto](https://github.com/huggingface/optimum-quanto/) can be used to quantize the text encoder, transformer and VAE modules to lower the memory requirements. This makes it possible to run the model on a free-tier T4 Colab or lower VRAM GPUs!
+
+It is also worth noting that torchao quantization is fully compatible with [torch.compile](/optimization/torch2.0#torchcompile), which allows for much faster inference speed. Additionally, models can be serialized and stored in a quantized datatype to save disk space with torchao. Find examples and benchmarks in the gists below.
+- [torchao](https://gist.github.com/a-r-r-o-w/4d9732d17412888c885480c6521a9897)
+- [quanto](https://gist.github.com/a-r-r-o-w/31be62828b00a9292821b85c1017effa)
+
 ## CogVideoXPipeline
 
 [[autodoc]] CogVideoXPipeline

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

@@ -0,0 +1,48 @@
+<!--Copyright 2024 The HuggingFace Team and The InstantX Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ControlNet with Flux.1
+
+FluxControlNetPipeline is an implementation of ControlNet for Flux.1.
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This controlnet code is implemented by [The InstantX Team](https://huggingface.co/InstantX). You can find pre-trained checkpoints for Flux-ControlNet in the table below:
+
+
+| ControlNet type | Developer | Link |
+| -------- | ---------- | ---- |
+| Canny | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny) |
+| Depth | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Depth) |
+| Union | [The InstantX Team](https://huggingface.co/InstantX) | [Link](https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union) |
+
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+## FluxControlNetPipeline
+[[autodoc]] FluxControlNetPipeline
+	- all
+	- __call__
+
+
+## FluxPipelineOutput
+[[autodoc]] pipelines.flux.pipeline_output.FluxPipelineOutput

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

@@ -30,63 +30,64 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 
 | Pipeline | Tasks |
 |---|---|
-| [AltDiffusion](alt_diffusion) | image2image |
+| [aMUSEd](amused) | text2image |
 | [AnimateDiff](animatediff) | text2video |
 | [Attend-and-Excite](attend_and_excite) | text2image |
-| [Audio Diffusion](audio_diffusion) | image2audio |
 | [AudioLDM](audioldm) | text2audio |
 | [AudioLDM2](audioldm2) | text2audio |
+| [AuraFlow](auraflow) | text2image |
 | [BLIP Diffusion](blip_diffusion) | text2image |
+| [CogVideoX](cogvideox) | text2video |
 | [Consistency Models](consistency_models) | unconditional image generation |
 | [ControlNet](controlnet) | text2image, image2image, inpainting |
+| [ControlNet with Flux.1](controlnet_flux) | text2image |
+| [ControlNet with Hunyuan-DiT](controlnet_hunyuandit) | text2image |
+| [ControlNet with Stable Diffusion 3](controlnet_sd3) | text2image |
 | [ControlNet with Stable Diffusion XL](controlnet_sdxl) | text2image |
 | [ControlNet-XS](controlnetxs) | text2image |
 | [ControlNet-XS with Stable Diffusion XL](controlnetxs_sdxl) | text2image |
-| [Cycle Diffusion](cycle_diffusion) | image2image |
 | [Dance Diffusion](dance_diffusion) | unconditional audio generation |
 | [DDIM](ddim) | unconditional image generation |
 | [DDPM](ddpm) | unconditional image generation |
 | [DeepFloyd IF](deepfloyd_if) | text2image, image2image, inpainting, super-resolution |
 | [DiffEdit](diffedit) | inpainting |
 | [DiT](dit) | text2image |
-| [GLIGEN](stable_diffusion/gligen) | text2image |
+| [Flux](flux) | text2image |
+| [Hunyuan-DiT](hunyuandit) | text2image |
+| [I2VGen-XL](i2vgenxl) | text2video |
 | [InstructPix2Pix](pix2pix) | image editing |
 | [Kandinsky 2.1](kandinsky) | text2image, image2image, inpainting, interpolation |
 | [Kandinsky 2.2](kandinsky_v22) | text2image, image2image, inpainting |
 | [Kandinsky 3](kandinsky3) | text2image, image2image |
+| [Kolors](kolors) | text2image |
 | [Latent Consistency Models](latent_consistency_models) | text2image |
 | [Latent Diffusion](latent_diffusion) | text2image, super-resolution |
-| [LDM3D](stable_diffusion/ldm3d_diffusion) | text2image, text-to-3D, text-to-pano, upscaling |
+| [Latte](latte) | text2image |
 | [LEDITS++](ledits_pp) | image editing |
+| [Lumina-T2X](lumina) | text2image |
+| [Marigold](marigold) | depth |
 | [MultiDiffusion](panorama) | text2image |
 | [MusicLDM](musicldm) | text2audio |
+| [PAG](pag) | text2image |
 | [Paint by Example](paint_by_example) | inpainting |
-| [ParaDiGMS](paradigms) | text2image |
-| [Pix2Pix Zero](pix2pix_zero) | image editing |
+| [PIA](pia) | image2video |
 | [PixArt-α](pixart) | text2image |
-| [PNDM](pndm) | unconditional image generation |
-| [RePaint](repaint) | inpainting |
-| [Score SDE VE](score_sde_ve) | unconditional image generation |
+| [PixArt-Σ](pixart_sigma) | text2image |
 | [Self-Attention Guidance](self_attention_guidance) | text2image |
 | [Semantic Guidance](semantic_stable_diffusion) | text2image |
 | [Shap-E](shap_e) | text-to-3D, image-to-3D |
-| [Spectrogram Diffusion](spectrogram_diffusion) |  |
 | [Stable Audio](stable_audio) | text2audio |
+| [Stable Cascade](stable_cascade) | text2image |
 | [Stable Diffusion](stable_diffusion/overview) | text2image, image2image, depth2image, inpainting, image variation, latent upscaler, super-resolution |
-| [Stable Diffusion Model Editing](model_editing) | model editing |
 | [Stable Diffusion XL](stable_diffusion/stable_diffusion_xl) | text2image, image2image, inpainting |
 | [Stable Diffusion XL Turbo](stable_diffusion/sdxl_turbo) | text2image, image2image, inpainting |
 | [Stable unCLIP](stable_unclip) | text2image, image variation |
-| [Stochastic Karras VE](stochastic_karras_ve) | unconditional image generation |
 | [T2I-Adapter](stable_diffusion/adapter) | text2image |
 | [Text2Video](text_to_video) | text2video, video2video |
 | [Text2Video-Zero](text_to_video_zero) | text2video |
 | [unCLIP](unclip) | text2image, image variation |
-| [Unconditional Latent Diffusion](latent_diffusion_uncond) | unconditional image generation |
 | [UniDiffuser](unidiffuser) | text2image, image2text, image variation, text variation, unconditional image generation, unconditional audio generation |
 | [Value-guided planning](value_guided_sampling) | value guided sampling |
-| [Versatile Diffusion](versatile_diffusion) | text2image, image variation |
-| [VQ Diffusion](vq_diffusion) | text2image |
 | [Wuerstchen](wuerstchen) | text2image |
 
 ## DiffusionPipeline

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

@@ -78,6 +78,10 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
 	- all
 	- __call__
 
+## StableDiffusionXLControlNetPAGImg2ImgPipeline
+[[autodoc]] StableDiffusionXLControlNetPAGImg2ImgPipeline
+	- all
+	- __call__
 
 ## StableDiffusion3PAGPipeline
 [[autodoc]] StableDiffusion3PAGPipeline

docs/source/en/stable_diffusion.md

@@ -238,7 +238,7 @@ Pretty impressive! Let's tweak the second image - corresponding to the `Generato
 ```python
 prompts = [
     "portrait photo of the oldest warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",
-    "portrait photo of a old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",
+    "portrait photo of an old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",
     "portrait photo of a warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",
     "portrait photo of a young warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",
 ]

examples/custom_diffusion/train_custom_diffusion.py

@@ -314,11 +314,12 @@ def save_new_embed(text_encoder, modifier_token_id, accelerator, args, output_di
     for x, y in zip(modifier_token_id, args.modifier_token):
         learned_embeds_dict = {}
         learned_embeds_dict[y] = learned_embeds[x]
-        filename = f"{output_dir}/{y}.bin"
 
         if safe_serialization:
+            filename = f"{output_dir}/{y}.safetensors"
             safetensors.torch.save_file(learned_embeds_dict, filename, metadata={"format": "pt"})
         else:
+            filename = f"{output_dir}/{y}.bin"
             torch.save(learned_embeds_dict, filename)
 
 
@@ -1040,17 +1041,22 @@ def main(args):
     )
 
     # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
     )
 
     # Prepare everything with our `accelerator`.
@@ -1065,8 +1071,14 @@ def main(args):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 

examples/dreambooth/train_dreambooth_flux.py

@@ -842,7 +842,7 @@ class PromptDataset(Dataset):
         return example
 
 
-def tokenize_prompt(tokenizer, prompt, max_sequence_length=512):
+def tokenize_prompt(tokenizer, prompt, max_sequence_length):
     text_inputs = tokenizer(
         prompt,
         padding="max_length",
@@ -863,20 +863,26 @@ def _encode_prompt_with_t5(
     prompt=None,
     num_images_per_prompt=1,
     device=None,
+    text_input_ids=None,
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
     batch_size = len(prompt)
 
-    text_inputs = tokenizer(
-        prompt,
-        padding="max_length",
-        max_length=max_sequence_length,
-        truncation=True,
-        return_length=False,
-        return_overflowing_tokens=False,
-        return_tensors="pt",
-    )
-    text_input_ids = text_inputs.input_ids
+    if tokenizer is not None:
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+    else:
+        if text_input_ids is None:
+            raise ValueError("text_input_ids must be provided when the tokenizer is not specified")
+
     prompt_embeds = text_encoder(text_input_ids.to(device))[0]
 
     dtype = text_encoder.dtype
@@ -896,22 +902,28 @@ def _encode_prompt_with_clip(
     tokenizer,
     prompt: str,
     device=None,
+    text_input_ids=None,
     num_images_per_prompt: int = 1,
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
     batch_size = len(prompt)
 
-    text_inputs = tokenizer(
-        prompt,
-        padding="max_length",
-        max_length=77,
-        truncation=True,
-        return_overflowing_tokens=False,
-        return_length=False,
-        return_tensors="pt",
-    )
+    if tokenizer is not None:
+        text_inputs = tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=77,
+            truncation=True,
+            return_overflowing_tokens=False,
+            return_length=False,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+    else:
+        if text_input_ids is None:
+            raise ValueError("text_input_ids must be provided when the tokenizer is not specified")
 
-    text_input_ids = text_inputs.input_ids
     prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=False)
 
     # Use pooled output of CLIPTextModel
@@ -932,17 +944,19 @@ def encode_prompt(
     max_sequence_length,
     device=None,
     num_images_per_prompt: int = 1,
+    text_input_ids_list=None,
 ):
     prompt = [prompt] if isinstance(prompt, str) else prompt
     batch_size = len(prompt)
     dtype = text_encoders[0].dtype
-
+    device = device if device is not None else text_encoders[1].device
     pooled_prompt_embeds = _encode_prompt_with_clip(
         text_encoder=text_encoders[0],
         tokenizer=tokenizers[0],
         prompt=prompt,
-        device=device if device is not None else text_encoders[0].device,
+        device=device,
         num_images_per_prompt=num_images_per_prompt,
+        text_input_ids=text_input_ids_list[0] if text_input_ids_list else None,
     )
 
     prompt_embeds = _encode_prompt_with_t5(
@@ -951,7 +965,8 @@ def encode_prompt(
         max_sequence_length=max_sequence_length,
         prompt=prompt,
         num_images_per_prompt=num_images_per_prompt,
-        device=device if device is not None else text_encoders[1].device,
+        device=device,
+        text_input_ids=text_input_ids_list[1] if text_input_ids_list else None,
     )
 
     text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
@@ -1499,7 +1514,25 @@ def main(args):
                         )
                     else:
                         tokens_one = tokenize_prompt(tokenizer_one, prompts, max_sequence_length=77)
-                        tokens_two = tokenize_prompt(tokenizer_two, prompts, max_sequence_length=512)
+                        tokens_two = tokenize_prompt(
+                            tokenizer_two, prompts, max_sequence_length=args.max_sequence_length
+                        )
+                        prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
+                            text_encoders=[text_encoder_one, text_encoder_two],
+                            tokenizers=[None, None],
+                            text_input_ids_list=[tokens_one, tokens_two],
+                            max_sequence_length=args.max_sequence_length,
+                            prompt=prompts,
+                        )
+                else:
+                    if args.train_text_encoder:
+                        prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
+                            text_encoders=[text_encoder_one, text_encoder_two],
+                            tokenizers=[None, None],
+                            text_input_ids_list=[tokens_one, tokens_two],
+                            max_sequence_length=args.max_sequence_length,
+                            prompt=args.instance_prompt,
+                        )
 
                 # Convert images to latent space
                 model_input = vae.encode(pixel_values).latent_dist.sample()
@@ -1553,41 +1586,22 @@ def main(args):
                     guidance = None
 
                 # Predict the noise residual
-                if not args.train_text_encoder:
-                    model_pred = transformer(
-                        hidden_states=packed_noisy_model_input,
-                        # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
-                        timestep=timesteps / 1000,
-                        guidance=guidance,
-                        pooled_projections=pooled_prompt_embeds,
-                        encoder_hidden_states=prompt_embeds,
-                        txt_ids=text_ids,
-                        img_ids=latent_image_ids,
-                        return_dict=False,
-                    )[0]
-                else:
-                    prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
-                        text_encoders=[text_encoder_one, text_encoder_two],
-                        tokenizers=None,
-                        prompt=None,
-                        text_input_ids_list=[tokens_one, tokens_two],
-                    )
-                    model_pred = transformer(
-                        hidden_states=packed_noisy_model_input,
-                        # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
-                        timestep=timesteps / 1000,
-                        guidance=guidance,
-                        pooled_projections=pooled_prompt_embeds,
-                        encoder_hidden_states=prompt_embeds,
-                        txt_ids=text_ids,
-                        img_ids=latent_image_ids,
-                        return_dict=False,
-                    )[0]
-
+                model_pred = transformer(
+                    hidden_states=packed_noisy_model_input,
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    timestep=timesteps / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    return_dict=False,
+                )[0]
+                # upscaling height & width as discussed in https://github.com/huggingface/diffusers/pull/9257#discussion_r1731108042
                 model_pred = FluxPipeline._unpack_latents(
                     model_pred,
-                    height=int(model_input.shape[2]),
-                    width=int(model_input.shape[3]),
+                    height=int(model_input.shape[2] * vae_scale_factor / 2),
+                    width=int(model_input.shape[3] * vae_scale_factor / 2),
                     vae_scale_factor=vae_scale_factor,
                 )
 

scripts/convert_cogvideox_to_diffusers.py

@@ -86,6 +86,9 @@ TRANSFORMER_SPECIAL_KEYS_REMAP = {
     "key_layernorm_list": reassign_query_key_layernorm_inplace,
     "adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
     "embed_tokens": remove_keys_inplace,
+    "freqs_sin": remove_keys_inplace,
+    "freqs_cos": remove_keys_inplace,
+    "position_embedding": remove_keys_inplace,
 }
 
 VAE_KEYS_RENAME_DICT = {
@@ -123,11 +126,21 @@ def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key:
     state_dict[new_key] = state_dict.pop(old_key)
 
 
-def convert_transformer(ckpt_path: str):
+def convert_transformer(
+    ckpt_path: str,
+    num_layers: int,
+    num_attention_heads: int,
+    use_rotary_positional_embeddings: bool,
+    dtype: torch.dtype,
+):
     PREFIX_KEY = "model.diffusion_model."
 
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    transformer = CogVideoXTransformer3DModel()
+    transformer = CogVideoXTransformer3DModel(
+        num_layers=num_layers,
+        num_attention_heads=num_attention_heads,
+        use_rotary_positional_embeddings=use_rotary_positional_embeddings,
+    ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[len(PREFIX_KEY) :]
@@ -145,9 +158,9 @@ def convert_transformer(ckpt_path: str):
     return transformer
 
 
-def convert_vae(ckpt_path: str):
+def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    vae = AutoencoderKLCogVideoX()
+    vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -172,13 +185,26 @@ def get_args():
     )
     parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
     parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
-    parser.add_argument("--fp16", action="store_true", default=True, help="Whether to save the model weights in fp16")
+    parser.add_argument("--fp16", action="store_true", default=False, help="Whether to save the model weights in fp16")
+    parser.add_argument("--bf16", action="store_true", default=False, help="Whether to save the model weights in bf16")
     parser.add_argument(
         "--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving"
     )
     parser.add_argument(
         "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
     )
+    # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
+    parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
+    # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
+    parser.add_argument("--num_attention_heads", type=int, default=30, help="Number of attention heads")
+    # For CogVideoX-2B, use_rotary_positional_embeddings is False. For 5B, it is True
+    parser.add_argument(
+        "--use_rotary_positional_embeddings", action="store_true", default=False, help="Whether to use RoPE or not"
+    )
+    # For CogVideoX-2B, scaling_factor is 1.15258426. For 5B, it is 0.7
+    parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
+    # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
+    parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
     return parser.parse_args()
 
 
@@ -188,18 +214,33 @@ if __name__ == "__main__":
     transformer = None
     vae = None
 
+    if args.fp16 and args.bf16:
+        raise ValueError("You cannot pass both --fp16 and --bf16 at the same time.")
+
+    dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
+
     if args.transformer_ckpt_path is not None:
-        transformer = convert_transformer(args.transformer_ckpt_path)
+        transformer = convert_transformer(
+            args.transformer_ckpt_path,
+            args.num_layers,
+            args.num_attention_heads,
+            args.use_rotary_positional_embeddings,
+            dtype,
+        )
     if args.vae_ckpt_path is not None:
-        vae = convert_vae(args.vae_ckpt_path)
+        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype)
 
     text_encoder_id = "google/t5-v1_1-xxl"
     tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
     text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
 
+    # Apparently, the conversion does not work any more without this :shrug:
+    for param in text_encoder.parameters():
+        param.data = param.data.contiguous()
+
     scheduler = CogVideoXDDIMScheduler.from_config(
         {
-            "snr_shift_scale": 3.0,
+            "snr_shift_scale": args.snr_shift_scale,
             "beta_end": 0.012,
             "beta_schedule": "scaled_linear",
             "beta_start": 0.00085,
@@ -208,7 +249,7 @@ if __name__ == "__main__":
             "prediction_type": "v_prediction",
             "rescale_betas_zero_snr": True,
             "set_alpha_to_one": True,
-            "timestep_spacing": "linspace",
+            "timestep_spacing": "trailing",
         }
     )
 
@@ -218,5 +259,10 @@ if __name__ == "__main__":
 
     if args.fp16:
         pipe = pipe.to(dtype=torch.float16)
+    if args.bf16:
+        pipe = pipe.to(dtype=torch.bfloat16)
 
+    # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
+    # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
+    # is either fp16/bf16 here).
     pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)

src/diffusers/__init__.py

@@ -89,6 +89,7 @@ else:
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
             "FluxControlNetModel",
+            "FluxMultiControlNetModel",
             "FluxTransformer2DModel",
             "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
@@ -346,6 +347,7 @@ else:
             "StableDiffusionXLAdapterPipeline",
             "StableDiffusionXLControlNetImg2ImgPipeline",
             "StableDiffusionXLControlNetInpaintPipeline",
+            "StableDiffusionXLControlNetPAGImg2ImgPipeline",
             "StableDiffusionXLControlNetPAGPipeline",
             "StableDiffusionXLControlNetPipeline",
             "StableDiffusionXLControlNetXSPipeline",
@@ -553,6 +555,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ControlNetXSAdapter,
             DiTTransformer2DModel,
             FluxControlNetModel,
+            FluxMultiControlNetModel,
             FluxTransformer2DModel,
             HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
@@ -787,6 +790,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLAdapterPipeline,
             StableDiffusionXLControlNetImg2ImgPipeline,
             StableDiffusionXLControlNetInpaintPipeline,
+            StableDiffusionXLControlNetPAGImg2ImgPipeline,
             StableDiffusionXLControlNetPAGPipeline,
             StableDiffusionXLControlNetPipeline,
             StableDiffusionXLControlNetXSPipeline,

src/diffusers/loaders/ip_adapter.py

@@ -208,6 +208,8 @@ class IPAdapterMixin:
                             pretrained_model_name_or_path_or_dict,
                             subfolder=image_encoder_subfolder,
                             low_cpu_mem_usage=low_cpu_mem_usage,
+                            cache_dir=cache_dir,
+                            local_files_only=local_files_only,
                         ).to(self.device, dtype=self.dtype)
                         self.register_modules(image_encoder=image_encoder)
                     else:

src/diffusers/loaders/lora_conversion_utils.py

@@ -14,6 +14,8 @@
 
 import re
 
+import torch
+
 from ..utils import is_peft_version, logging
 
 
@@ -326,3 +328,294 @@ def _get_alpha_name(lora_name_alpha, diffusers_name, alpha):
         prefix = "text_encoder_2."
     new_name = prefix + diffusers_name.split(".lora.")[0] + ".alpha"
     return {new_name: alpha}
+
+
+# The utilities under `_convert_kohya_flux_lora_to_diffusers()`
+# are taken from https://github.com/kohya-ss/sd-scripts/blob/a61cf73a5cb5209c3f4d1a3688dd276a4dfd1ecb/networks/convert_flux_lora.py
+# All credits go to `kohya-ss`.
+def _convert_kohya_flux_lora_to_diffusers(state_dict):
+    def _convert_to_ai_toolkit(sds_sd, ait_sd, sds_key, ait_key):
+        if sds_key + ".lora_down.weight" not in sds_sd:
+            return
+        down_weight = sds_sd.pop(sds_key + ".lora_down.weight")
+
+        # scale weight by alpha and dim
+        rank = down_weight.shape[0]
+        alpha = sds_sd.pop(sds_key + ".alpha").item()  # alpha is scalar
+        scale = alpha / rank  # LoRA is scaled by 'alpha / rank' in forward pass, so we need to scale it back here
+
+        # calculate scale_down and scale_up to keep the same value. if scale is 4, scale_down is 2 and scale_up is 2
+        scale_down = scale
+        scale_up = 1.0
+        while scale_down * 2 < scale_up:
+            scale_down *= 2
+            scale_up /= 2
+
+        ait_sd[ait_key + ".lora_A.weight"] = down_weight * scale_down
+        ait_sd[ait_key + ".lora_B.weight"] = sds_sd.pop(sds_key + ".lora_up.weight") * scale_up
+
+    def _convert_to_ai_toolkit_cat(sds_sd, ait_sd, sds_key, ait_keys, dims=None):
+        if sds_key + ".lora_down.weight" not in sds_sd:
+            return
+        down_weight = sds_sd.pop(sds_key + ".lora_down.weight")
+        up_weight = sds_sd.pop(sds_key + ".lora_up.weight")
+        sd_lora_rank = down_weight.shape[0]
+
+        # scale weight by alpha and dim
+        alpha = sds_sd.pop(sds_key + ".alpha")
+        scale = alpha / sd_lora_rank
+
+        # calculate scale_down and scale_up
+        scale_down = scale
+        scale_up = 1.0
+        while scale_down * 2 < scale_up:
+            scale_down *= 2
+            scale_up /= 2
+
+        down_weight = down_weight * scale_down
+        up_weight = up_weight * scale_up
+
+        # calculate dims if not provided
+        num_splits = len(ait_keys)
+        if dims is None:
+            dims = [up_weight.shape[0] // num_splits] * num_splits
+        else:
+            assert sum(dims) == up_weight.shape[0]
+
+        # check upweight is sparse or not
+        is_sparse = False
+        if sd_lora_rank % num_splits == 0:
+            ait_rank = sd_lora_rank // num_splits
+            is_sparse = True
+            i = 0
+            for j in range(len(dims)):
+                for k in range(len(dims)):
+                    if j == k:
+                        continue
+                    is_sparse = is_sparse and torch.all(
+                        up_weight[i : i + dims[j], k * ait_rank : (k + 1) * ait_rank] == 0
+                    )
+                i += dims[j]
+            if is_sparse:
+                logger.info(f"weight is sparse: {sds_key}")
+
+        # make ai-toolkit weight
+        ait_down_keys = [k + ".lora_A.weight" for k in ait_keys]
+        ait_up_keys = [k + ".lora_B.weight" for k in ait_keys]
+        if not is_sparse:
+            # down_weight is copied to each split
+            ait_sd.update({k: down_weight for k in ait_down_keys})
+
+            # up_weight is split to each split
+            ait_sd.update({k: v for k, v in zip(ait_up_keys, torch.split(up_weight, dims, dim=0))})  # noqa: C416
+        else:
+            # down_weight is chunked to each split
+            ait_sd.update({k: v for k, v in zip(ait_down_keys, torch.chunk(down_weight, num_splits, dim=0))})  # noqa: C416
+
+            # up_weight is sparse: only non-zero values are copied to each split
+            i = 0
+            for j in range(len(dims)):
+                ait_sd[ait_up_keys[j]] = up_weight[i : i + dims[j], j * ait_rank : (j + 1) * ait_rank].contiguous()
+                i += dims[j]
+
+    def _convert_sd_scripts_to_ai_toolkit(sds_sd):
+        ait_sd = {}
+        for i in range(19):
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_img_attn_proj",
+                f"transformer.transformer_blocks.{i}.attn.to_out.0",
+            )
+            _convert_to_ai_toolkit_cat(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_img_attn_qkv",
+                [
+                    f"transformer.transformer_blocks.{i}.attn.to_q",
+                    f"transformer.transformer_blocks.{i}.attn.to_k",
+                    f"transformer.transformer_blocks.{i}.attn.to_v",
+                ],
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_img_mlp_0",
+                f"transformer.transformer_blocks.{i}.ff.net.0.proj",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_img_mlp_2",
+                f"transformer.transformer_blocks.{i}.ff.net.2",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_img_mod_lin",
+                f"transformer.transformer_blocks.{i}.norm1.linear",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_txt_attn_proj",
+                f"transformer.transformer_blocks.{i}.attn.to_add_out",
+            )
+            _convert_to_ai_toolkit_cat(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_txt_attn_qkv",
+                [
+                    f"transformer.transformer_blocks.{i}.attn.add_q_proj",
+                    f"transformer.transformer_blocks.{i}.attn.add_k_proj",
+                    f"transformer.transformer_blocks.{i}.attn.add_v_proj",
+                ],
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_txt_mlp_0",
+                f"transformer.transformer_blocks.{i}.ff_context.net.0.proj",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_txt_mlp_2",
+                f"transformer.transformer_blocks.{i}.ff_context.net.2",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_double_blocks_{i}_txt_mod_lin",
+                f"transformer.transformer_blocks.{i}.norm1_context.linear",
+            )
+
+        for i in range(38):
+            _convert_to_ai_toolkit_cat(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_single_blocks_{i}_linear1",
+                [
+                    f"transformer.single_transformer_blocks.{i}.attn.to_q",
+                    f"transformer.single_transformer_blocks.{i}.attn.to_k",
+                    f"transformer.single_transformer_blocks.{i}.attn.to_v",
+                    f"transformer.single_transformer_blocks.{i}.proj_mlp",
+                ],
+                dims=[3072, 3072, 3072, 12288],
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_single_blocks_{i}_linear2",
+                f"transformer.single_transformer_blocks.{i}.proj_out",
+            )
+            _convert_to_ai_toolkit(
+                sds_sd,
+                ait_sd,
+                f"lora_unet_single_blocks_{i}_modulation_lin",
+                f"transformer.single_transformer_blocks.{i}.norm.linear",
+            )
+
+        if len(sds_sd) > 0:
+            logger.warning(f"Unsuppored keys for ai-toolkit: {sds_sd.keys()}")
+
+        return ait_sd
+
+    return _convert_sd_scripts_to_ai_toolkit(state_dict)
+
+
+# Adapted from https://gist.github.com/Leommm-byte/6b331a1e9bd53271210b26543a7065d6
+# Some utilities were reused from
+# https://github.com/kohya-ss/sd-scripts/blob/a61cf73a5cb5209c3f4d1a3688dd276a4dfd1ecb/networks/convert_flux_lora.py
+def _convert_xlabs_flux_lora_to_diffusers(old_state_dict):
+    new_state_dict = {}
+    orig_keys = list(old_state_dict.keys())
+
+    def handle_qkv(sds_sd, ait_sd, sds_key, ait_keys, dims=None):
+        down_weight = sds_sd.pop(sds_key)
+        up_weight = sds_sd.pop(sds_key.replace(".down.weight", ".up.weight"))
+
+        # calculate dims if not provided
+        num_splits = len(ait_keys)
+        if dims is None:
+            dims = [up_weight.shape[0] // num_splits] * num_splits
+        else:
+            assert sum(dims) == up_weight.shape[0]
+
+        # make ai-toolkit weight
+        ait_down_keys = [k + ".lora_A.weight" for k in ait_keys]
+        ait_up_keys = [k + ".lora_B.weight" for k in ait_keys]
+
+        # down_weight is copied to each split
+        ait_sd.update({k: down_weight for k in ait_down_keys})
+
+        # up_weight is split to each split
+        ait_sd.update({k: v for k, v in zip(ait_up_keys, torch.split(up_weight, dims, dim=0))})  # noqa: C416
+
+    for old_key in orig_keys:
+        # Handle double_blocks
+        if old_key.startswith(("diffusion_model.double_blocks", "double_blocks")):
+            block_num = re.search(r"double_blocks\.(\d+)", old_key).group(1)
+            new_key = f"transformer.transformer_blocks.{block_num}"
+
+            if "processor.proj_lora1" in old_key:
+                new_key += ".attn.to_out.0"
+            elif "processor.proj_lora2" in old_key:
+                new_key += ".attn.to_add_out"
+            elif "processor.qkv_lora1" in old_key and "up" not in old_key:
+                handle_qkv(
+                    old_state_dict,
+                    new_state_dict,
+                    old_key,
+                    [
+                        f"transformer.transformer_blocks.{block_num}.attn.add_q_proj",
+                        f"transformer.transformer_blocks.{block_num}.attn.add_k_proj",
+                        f"transformer.transformer_blocks.{block_num}.attn.add_v_proj",
+                    ],
+                )
+                # continue
+            elif "processor.qkv_lora2" in old_key and "up" not in old_key:
+                handle_qkv(
+                    old_state_dict,
+                    new_state_dict,
+                    old_key,
+                    [
+                        f"transformer.transformer_blocks.{block_num}.attn.to_q",
+                        f"transformer.transformer_blocks.{block_num}.attn.to_k",
+                        f"transformer.transformer_blocks.{block_num}.attn.to_v",
+                    ],
+                )
+                # continue
+
+            if "down" in old_key:
+                new_key += ".lora_A.weight"
+            elif "up" in old_key:
+                new_key += ".lora_B.weight"
+
+        # Handle single_blocks
+        elif old_key.startswith("diffusion_model.single_blocks", "single_blocks"):
+            block_num = re.search(r"single_blocks\.(\d+)", old_key).group(1)
+            new_key = f"transformer.single_transformer_blocks.{block_num}"
+
+            if "proj_lora1" in old_key or "proj_lora2" in old_key:
+                new_key += ".proj_out"
+            elif "qkv_lora1" in old_key or "qkv_lora2" in old_key:
+                new_key += ".norm.linear"
+
+            if "down" in old_key:
+                new_key += ".lora_A.weight"
+            elif "up" in old_key:
+                new_key += ".lora_B.weight"
+
+        else:
+            # Handle other potential key patterns here
+            new_key = old_key
+
+        # Since we already handle qkv above.
+        if "qkv" not in old_key:
+            new_state_dict[new_key] = old_state_dict.pop(old_key)
+
+    if len(old_state_dict) > 0:
+        raise ValueError(f"`old_state_dict` should be at this point but has: {list(old_state_dict.keys())}.")
+
+    return new_state_dict

src/diffusers/loaders/lora_pipeline.py

@@ -31,7 +31,12 @@ from ..utils import (
     scale_lora_layers,
 )
 from .lora_base import LoraBaseMixin
-from .lora_conversion_utils import _convert_non_diffusers_lora_to_diffusers, _maybe_map_sgm_blocks_to_diffusers
+from .lora_conversion_utils import (
+    _convert_kohya_flux_lora_to_diffusers,
+    _convert_non_diffusers_lora_to_diffusers,
+    _convert_xlabs_flux_lora_to_diffusers,
+    _maybe_map_sgm_blocks_to_diffusers,
+)
 
 
 if is_transformers_available():
@@ -280,7 +285,9 @@ class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
                 A standard state dict containing the lora layer parameters. The key should be prefixed with an
                 additional `text_encoder` to distinguish between unet lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             text_encoder (`CLIPTextModel`):
                 The text encoder model to load the LoRA layers into.
             prefix (`str`):
@@ -753,7 +760,9 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
                 A standard state dict containing the lora layer parameters. The key should be prefixed with an
                 additional `text_encoder` to distinguish between unet lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             text_encoder (`CLIPTextModel`):
                 The text encoder model to load the LoRA layers into.
             prefix (`str`):
@@ -1249,7 +1258,9 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
                 A standard state dict containing the lora layer parameters. The key should be prefixed with an
                 additional `text_encoder` to distinguish between unet lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             text_encoder (`CLIPTextModel`):
                 The text encoder model to load the LoRA layers into.
             prefix (`str`):
@@ -1489,10 +1500,10 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
 
     @classmethod
     @validate_hf_hub_args
-    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
     def lora_state_dict(
         cls,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        return_alphas: bool = False,
         **kwargs,
     ):
         r"""
@@ -1577,7 +1588,40 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
             allow_pickle=allow_pickle,
         )
 
-        return state_dict
+        # TODO (sayakpaul): to a follow-up to clean and try to unify the conditions.
+
+        is_kohya = any(".lora_down.weight" in k for k in state_dict)
+        if is_kohya:
+            state_dict = _convert_kohya_flux_lora_to_diffusers(state_dict)
+            # Kohya already takes care of scaling the LoRA parameters with alpha.
+            return (state_dict, None) if return_alphas else state_dict
+
+        is_xlabs = any("processor" in k for k in state_dict)
+        if is_xlabs:
+            state_dict = _convert_xlabs_flux_lora_to_diffusers(state_dict)
+            # xlabs doesn't use `alpha`.
+            return (state_dict, None) if return_alphas else state_dict
+
+        # For state dicts like
+        # https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA
+        keys = list(state_dict.keys())
+        network_alphas = {}
+        for k in keys:
+            if "alpha" in k:
+                alpha_value = state_dict.get(k)
+                if (torch.is_tensor(alpha_value) and torch.is_floating_point(alpha_value)) or isinstance(
+                    alpha_value, float
+                ):
+                    network_alphas[k] = state_dict.pop(k)
+                else:
+                    raise ValueError(
+                        f"The alpha key ({k}) seems to be incorrect. If you think this error is unexpected, please open as issue."
+                    )
+
+        if return_alphas:
+            return state_dict, network_alphas
+        else:
+            return state_dict
 
     def load_lora_weights(
         self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
@@ -1611,7 +1655,9 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
             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.
-        state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+        state_dict, network_alphas = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict, return_alphas=True, **kwargs
+        )
 
         is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
         if not is_correct_format:
@@ -1619,6 +1665,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
 
         self.load_lora_into_transformer(
             state_dict,
+            network_alphas=network_alphas,
             transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
             adapter_name=adapter_name,
             _pipeline=self,
@@ -1628,7 +1675,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
         if len(text_encoder_state_dict) > 0:
             self.load_lora_into_text_encoder(
                 text_encoder_state_dict,
-                network_alphas=None,
+                network_alphas=network_alphas,
                 text_encoder=self.text_encoder,
                 prefix="text_encoder",
                 lora_scale=self.lora_scale,
@@ -1637,8 +1684,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
             )
 
     @classmethod
-    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer
-    def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=None):
+    def load_lora_into_transformer(cls, state_dict, network_alphas, transformer, adapter_name=None, _pipeline=None):
         """
         This will load the LoRA layers specified in `state_dict` into `transformer`.
 
@@ -1647,6 +1693,10 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
                 A standard state dict containing the lora layer parameters. The keys can either be indexed directly
                 into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                 encoder lora layers.
+            network_alphas (`Dict[str, float]`):
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             transformer (`SD3Transformer2DModel`):
                 The Transformer model to load the LoRA layers into.
             adapter_name (`str`, *optional*):
@@ -1678,7 +1728,12 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
                 if "lora_B" in key:
                     rank[key] = val.shape[1]
 
-            lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
+            if network_alphas is not None and len(network_alphas) >= 1:
+                prefix = cls.transformer_name
+                alpha_keys = [k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix]
+                network_alphas = {k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys}
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=network_alphas, peft_state_dict=state_dict)
             if "use_dora" in lora_config_kwargs:
                 if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
                     raise ValueError(
@@ -1735,7 +1790,9 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
                 A standard state dict containing the lora layer parameters. The key should be prefixed with an
                 additional `text_encoder` to distinguish between unet lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             text_encoder (`CLIPTextModel`):
                 The text encoder model to load the LoRA layers into.
             prefix (`str`):
@@ -1968,7 +2025,9 @@ class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
                 into the unet or prefixed with an additional `unet` which can be used to distinguish between text
                 encoder lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             unet (`UNet2DConditionModel`):
                 The UNet model to load the LoRA layers into.
             adapter_name (`str`, *optional*):
@@ -2061,7 +2120,9 @@ class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
                 A standard state dict containing the lora layer parameters. The key should be prefixed with an
                 additional `text_encoder` to distinguish between unet lora layers.
             network_alphas (`Dict[str, float]`):
-                See `LoRALinearLayer` for more details.
+                The value of the network alpha used for stable learning and preventing underflow. This value has the
+                same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
+                link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
             text_encoder (`CLIPTextModel`):
                 The text encoder model to load the LoRA layers into.
             prefix (`str`):

src/diffusers/loaders/single_file.py

@@ -23,6 +23,7 @@ from packaging import version
 from ..utils import deprecate, is_transformers_available, logging
 from .single_file_utils import (
     SingleFileComponentError,
+    _is_legacy_scheduler_kwargs,
     _is_model_weights_in_cached_folder,
     _legacy_load_clip_tokenizer,
     _legacy_load_safety_checker,
@@ -42,7 +43,6 @@ logger = logging.get_logger(__name__)
 # Legacy behaviour. `from_single_file` does not load the safety checker unless explicitly provided
 SINGLE_FILE_OPTIONAL_COMPONENTS = ["safety_checker"]
 
-
 if is_transformers_available():
     import transformers
     from transformers import PreTrainedModel, PreTrainedTokenizer
@@ -135,7 +135,7 @@ def load_single_file_sub_model(
             class_obj, checkpoint=checkpoint, config=cached_model_config_path, local_files_only=local_files_only
         )
 
-    elif is_diffusers_scheduler and is_legacy_loading:
+    elif is_diffusers_scheduler and (is_legacy_loading or _is_legacy_scheduler_kwargs(kwargs)):
         loaded_sub_model = _legacy_load_scheduler(
             class_obj, checkpoint=checkpoint, component_name=name, original_config=original_config, **kwargs
         )

src/diffusers/loaders/single_file_utils.py

@@ -91,11 +91,11 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "xl_inpaint": {"pretrained_model_name_or_path": "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"},
     "playground-v2-5": {"pretrained_model_name_or_path": "playgroundai/playground-v2.5-1024px-aesthetic"},
     "upscale": {"pretrained_model_name_or_path": "stabilityai/stable-diffusion-x4-upscaler"},
-    "inpainting": {"pretrained_model_name_or_path": "runwayml/stable-diffusion-inpainting"},
+    "inpainting": {"pretrained_model_name_or_path": "Lykon/dreamshaper-8-inpainting"},
     "inpainting_v2": {"pretrained_model_name_or_path": "stabilityai/stable-diffusion-2-inpainting"},
     "controlnet": {"pretrained_model_name_or_path": "lllyasviel/control_v11p_sd15_canny"},
     "v2": {"pretrained_model_name_or_path": "stabilityai/stable-diffusion-2-1"},
-    "v1": {"pretrained_model_name_or_path": "runwayml/stable-diffusion-v1-5"},
+    "v1": {"pretrained_model_name_or_path": "Lykon/dreamshaper-8"},
     "stable_cascade_stage_b": {"pretrained_model_name_or_path": "stabilityai/stable-cascade", "subfolder": "decoder"},
     "stable_cascade_stage_b_lite": {
         "pretrained_model_name_or_path": "stabilityai/stable-cascade",
@@ -271,6 +271,7 @@ LDM_CLIP_PREFIX_TO_REMOVE = [
     "conditioner.embedders.0.transformer.",
 ]
 LDM_OPEN_CLIP_TEXT_PROJECTION_DIM = 1024
+SCHEDULER_LEGACY_KWARGS = ["prediction_type", "scheduler_type"]
 
 VALID_URL_PREFIXES = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
 
@@ -320,6 +321,10 @@ def _is_model_weights_in_cached_folder(cached_folder, name):
     return weights_exist
 
 
+def _is_legacy_scheduler_kwargs(kwargs):
+    return any(k in SCHEDULER_LEGACY_KWARGS for k in kwargs.keys())
+
+
 def load_single_file_checkpoint(
     pretrained_model_link_or_path,
     force_download=False,
@@ -1487,14 +1492,22 @@ def _legacy_load_scheduler(
 
     if scheduler_type is not None:
         deprecation_message = (
-            "Please pass an instance of a Scheduler object directly to the `scheduler` argument in `from_single_file`."
+            "Please pass an instance of a Scheduler object directly to the `scheduler` argument in `from_single_file`\n\n"
+            "Example:\n\n"
+            "from diffusers import StableDiffusionPipeline, DDIMScheduler\n\n"
+            "scheduler = DDIMScheduler()\n"
+            "pipe = StableDiffusionPipeline.from_single_file(<checkpoint path>, scheduler=scheduler)\n"
         )
         deprecate("scheduler_type", "1.0.0", deprecation_message)
 
     if prediction_type is not None:
         deprecation_message = (
-            "Please configure an instance of a Scheduler with the appropriate `prediction_type` "
-            "and pass the object directly to the `scheduler` argument in `from_single_file`."
+            "Please configure an instance of a Scheduler with the appropriate `prediction_type` and "
+            "pass the object directly to the `scheduler` argument in `from_single_file`.\n\n"
+            "Example:\n\n"
+            "from diffusers import StableDiffusionPipeline, DDIMScheduler\n\n"
+            'scheduler = DDIMScheduler(prediction_type="v_prediction")\n'
+            "pipe = StableDiffusionPipeline.from_single_file(<checkpoint path>, scheduler=scheduler)\n"
         )
         deprecate("prediction_type", "1.0.0", deprecation_message)
 

src/diffusers/models/__init__.py

@@ -35,7 +35,7 @@ if is_torch_available():
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["autoencoders.vq_model"] = ["VQModel"]
     _import_structure["controlnet"] = ["ControlNetModel"]
-    _import_structure["controlnet_flux"] = ["FluxControlNetModel"]
+    _import_structure["controlnet_flux"] = ["FluxControlNetModel", "FluxMultiControlNetModel"]
     _import_structure["controlnet_hunyuan"] = ["HunyuanDiT2DControlNetModel", "HunyuanDiT2DMultiControlNetModel"]
     _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
@@ -88,7 +88,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VQModel,
         )
         from .controlnet import ControlNetModel
-        from .controlnet_flux import FluxControlNetModel
+        from .controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel
         from .controlnet_hunyuan import HunyuanDiT2DControlNetModel, HunyuanDiT2DMultiControlNetModel
         from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
         from .controlnet_sparsectrl import SparseControlNetModel

src/diffusers/models/attention.py

@@ -972,15 +972,32 @@ class FreeNoiseTransformerBlock(nn.Module):
         return frame_indices
 
     def _get_frame_weights(self, num_frames: int, weighting_scheme: str = "pyramid") -> List[float]:
-        if weighting_scheme == "pyramid":
+        if weighting_scheme == "flat":
+            weights = [1.0] * num_frames
+
+        elif weighting_scheme == "pyramid":
             if num_frames % 2 == 0:
                 # num_frames = 4 => [1, 2, 2, 1]
-                weights = list(range(1, num_frames // 2 + 1))
+                mid = num_frames // 2
+                weights = list(range(1, mid + 1))
                 weights = weights + weights[::-1]
             else:
                 # num_frames = 5 => [1, 2, 3, 2, 1]
-                weights = list(range(1, num_frames // 2 + 1))
-                weights = weights + [num_frames // 2 + 1] + weights[::-1]
+                mid = (num_frames + 1) // 2
+                weights = list(range(1, mid))
+                weights = weights + [mid] + weights[::-1]
+
+        elif weighting_scheme == "delayed_reverse_sawtooth":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [0.01, 2, 2, 1]
+                mid = num_frames // 2
+                weights = [0.01] * (mid - 1) + [mid]
+                weights = weights + list(range(mid, 0, -1))
+            else:
+                # num_frames = 5 => [0.01, 0.01, 3, 2, 1]
+                mid = (num_frames + 1) // 2
+                weights = [0.01] * mid
+                weights = weights + list(range(mid, 0, -1))
         else:
             raise ValueError(f"Unsupported value for weighting_scheme={weighting_scheme}")
 

src/diffusers/models/attention_processor.py

@@ -1695,81 +1695,6 @@ class FusedAuraFlowAttnProcessor2_0:
             return hidden_states
 
 
-# YiYi to-do: refactor rope related functions/classes
-def apply_rope(xq, xk, freqs_cis):
-    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
-    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
-    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
-    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
-    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
-
-
-class FluxSingleAttnProcessor2_0:
-    r"""
-    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
-    """
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        input_ndim = hidden_states.ndim
-
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        query = attn.to_q(hidden_states)
-        if encoder_hidden_states is None:
-            encoder_hidden_states = hidden_states
-
-        key = attn.to_k(encoder_hidden_states)
-        value = attn.to_v(encoder_hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # Apply RoPE if needed
-        if image_rotary_emb is not None:
-            # YiYi to-do: update uising apply_rotary_emb
-            # from ..embeddings import apply_rotary_emb
-            # query = apply_rotary_emb(query, image_rotary_emb)
-            # key = apply_rotary_emb(key, image_rotary_emb)
-            query, key = apply_rope(query, key, image_rotary_emb)
-
-        # the output of sdp = (batch, num_heads, seq_len, head_dim)
-        # TODO: add support for attn.scale when we move to Torch 2.1
-        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
-
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
-
-        return hidden_states
-
-
 class FluxAttnProcessor2_0:
     """Attention processor used typically in processing the SD3-like self-attention projections."""
 
@@ -1785,16 +1710,7 @@ class FluxAttnProcessor2_0:
         attention_mask: Optional[torch.FloatTensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
-        input_ndim = hidden_states.ndim
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-        context_input_ndim = encoder_hidden_states.ndim
-        if context_input_ndim == 4:
-            batch_size, channel, height, width = encoder_hidden_states.shape
-            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-
-        batch_size = encoder_hidden_states.shape[0]
+        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
 
         # `sample` projections.
         query = attn.to_q(hidden_states)
@@ -1813,58 +1729,293 @@ class FluxAttnProcessor2_0:
         if attn.norm_k is not None:
             key = attn.norm_k(key)
 
-        # `context` projections.
-        encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
-        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
-        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+        if encoder_hidden_states is not None:
+            # `context` projections.
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
 
-        encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-            batch_size, -1, attn.heads, head_dim
-        ).transpose(1, 2)
-        encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-            batch_size, -1, attn.heads, head_dim
-        ).transpose(1, 2)
-        encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-            batch_size, -1, attn.heads, head_dim
-        ).transpose(1, 2)
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
 
-        if attn.norm_added_q is not None:
-            encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-        if attn.norm_added_k is not None:
-            encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
 
-        # attention
-        query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-        key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-        value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+            # attention
+            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
 
         if image_rotary_emb is not None:
-            # YiYi to-do: update uising apply_rotary_emb
-            # from ..embeddings import apply_rotary_emb
-            # query = apply_rotary_emb(query, image_rotary_emb)
-            # key = apply_rotary_emb(key, image_rotary_emb)
-            query, key = apply_rope(query, key, image_rotary_emb)
+            from .embeddings import apply_rotary_emb
+
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
 
         hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
         hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
         hidden_states = hidden_states.to(query.dtype)
 
-        encoder_hidden_states, hidden_states = (
-            hidden_states[:, : encoder_hidden_states.shape[1]],
-            hidden_states[:, encoder_hidden_states.shape[1] :],
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            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
+
+
+class FusedFluxAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "FusedFluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+
+        # `sample` projections.
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+        # `context` projections.
+        if encoder_hidden_states is not None:
+            encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
+            split_size = encoder_qkv.shape[-1] // 3
+            (
+                encoder_hidden_states_query_proj,
+                encoder_hidden_states_key_proj,
+                encoder_hidden_states_value_proj,
+            ) = torch.split(encoder_qkv, split_size, dim=-1)
+
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+
+            # attention
+            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            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
+
+
+class CogVideoXAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
+    query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
         )
 
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
         # linear proj
         hidden_states = attn.to_out[0](hidden_states)
         # dropout
         hidden_states = attn.to_out[1](hidden_states)
-        encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
 
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
-        if context_input_ndim == 4:
-            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
 
+
+class FusedCogVideoXAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
+    query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
         return hidden_states, encoder_hidden_states
 
 
@@ -4105,6 +4256,17 @@ class LoRAAttnAddedKVProcessor:
         pass
 
 
+class FluxSingleAttnProcessor2_0(FluxAttnProcessor2_0):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self):
+        deprecation_message = "`FluxSingleAttnProcessor2_0` is deprecated and will be removed in a future version. Please use `FluxAttnProcessor2_0` instead."
+        deprecate("FluxSingleAttnProcessor2_0", "0.32.0", deprecation_message)
+        super().__init__()
+
+
 ADDED_KV_ATTENTION_PROCESSORS = (
     AttnAddedKVProcessor,
     SlicedAttnAddedKVProcessor,

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -902,7 +902,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             Tuple of block output channels.
         act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
         sample_size (`int`, *optional*, defaults to `32`): Sample input size.
-        scaling_factor (`float`, *optional*, defaults to 0.18215):
+        scaling_factor (`float`, *optional*, defaults to `1.15258426`):
             The component-wise standard deviation of the trained latent space computed using the first batch of the
             training set. This is used to scale the latent space to have unit variance when training the diffusion
             model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the

src/diffusers/models/controlnet_flux.py

@@ -24,9 +24,9 @@ from ..models.attention_processor import AttentionProcessor
 from ..models.modeling_utils import ModelMixin
 from ..utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from .controlnet import BaseOutput, zero_module
-from .embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
+from .embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from .modeling_outputs import Transformer2DModelOutput
-from .transformers.transformer_flux import EmbedND, FluxSingleTransformerBlock, FluxTransformerBlock
+from .transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -54,12 +54,13 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
         axes_dims_rope: List[int] = [16, 56, 56],
+        num_mode: int = None,
     ):
         super().__init__()
         self.out_channels = in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
 
-        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
+        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
         text_time_guidance_cls = (
             CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
         )
@@ -101,6 +102,10 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         for _ in range(len(self.single_transformer_blocks)):
             self.controlnet_single_blocks.append(zero_module(nn.Linear(self.inner_dim, self.inner_dim)))
 
+        self.union = num_mode is not None
+        if self.union:
+            self.controlnet_mode_embedder = nn.Embedding(num_mode, self.inner_dim)
+
         self.controlnet_x_embedder = zero_module(torch.nn.Linear(in_channels, self.inner_dim))
 
         self.gradient_checkpointing = False
@@ -173,8 +178,8 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
     def from_transformer(
         cls,
         transformer,
-        num_layers=4,
-        num_single_layers=10,
+        num_layers: int = 4,
+        num_single_layers: int = 10,
         attention_head_dim: int = 128,
         num_attention_heads: int = 24,
         load_weights_from_transformer=True,
@@ -205,6 +210,7 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         self,
         hidden_states: torch.Tensor,
         controlnet_cond: torch.Tensor,
+        controlnet_mode: torch.Tensor = None,
         conditioning_scale: float = 1.0,
         encoder_hidden_states: torch.Tensor = None,
         pooled_projections: torch.Tensor = None,
@@ -221,6 +227,12 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         Args:
             hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
                 Input `hidden_states`.
+            controlnet_cond (`torch.Tensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            controlnet_mode (`torch.Tensor`):
+                The mode tensor of shape `(batch_size, 1)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
             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
@@ -272,8 +284,29 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         )
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
-        txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
-        ids = torch.cat((txt_ids, img_ids), dim=1)
+        if self.union:
+            # union mode
+            if controlnet_mode is None:
+                raise ValueError("`controlnet_mode` cannot be `None` when applying ControlNet-Union")
+            # union mode emb
+            controlnet_mode_emb = self.controlnet_mode_embedder(controlnet_mode)
+            encoder_hidden_states = torch.cat([controlnet_mode_emb, encoder_hidden_states], dim=1)
+            txt_ids = torch.cat([txt_ids[:1], txt_ids], dim=0)
+
+        if txt_ids.ndim == 3:
+            logger.warning(
+                "Passing `txt_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            txt_ids = txt_ids[0]
+        if img_ids.ndim == 3:
+            logger.warning(
+                "Passing `img_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            img_ids = img_ids[0]
+
+        ids = torch.cat((txt_ids, img_ids), dim=0)
         image_rotary_emb = self.pos_embed(ids)
 
         block_samples = ()
@@ -355,7 +388,6 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         controlnet_block_samples = [sample * conditioning_scale for sample in controlnet_block_samples]
         controlnet_single_block_samples = [sample * conditioning_scale for sample in controlnet_single_block_samples]
 
-        #
         controlnet_block_samples = None if len(controlnet_block_samples) == 0 else controlnet_block_samples
         controlnet_single_block_samples = (
             None if len(controlnet_single_block_samples) == 0 else controlnet_single_block_samples
@@ -372,3 +404,114 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             controlnet_block_samples=controlnet_block_samples,
             controlnet_single_block_samples=controlnet_single_block_samples,
         )
+
+
+class FluxMultiControlNetModel(ModelMixin):
+    r"""
+    `FluxMultiControlNetModel` wrapper class for Multi-FluxControlNetModel
+
+    This module is a wrapper for multiple instances of the `FluxControlNetModel`. The `forward()` API is designed to be
+    compatible with `FluxControlNetModel`.
+
+    Args:
+        controlnets (`List[FluxControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `FluxControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets):
+        super().__init__()
+        self.nets = nn.ModuleList(controlnets)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        controlnet_cond: List[torch.tensor],
+        controlnet_mode: List[torch.tensor],
+        conditioning_scale: List[float],
+        encoder_hidden_states: torch.Tensor = 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[FluxControlNetOutput, Tuple]:
+        # ControlNet-Union with multiple conditions
+        # only load one ControlNet for saving memories
+        if len(self.nets) == 1 and self.nets[0].union:
+            controlnet = self.nets[0]
+
+            for i, (image, mode, scale) in enumerate(zip(controlnet_cond, controlnet_mode, conditioning_scale)):
+                block_samples, single_block_samples = controlnet(
+                    hidden_states=hidden_states,
+                    controlnet_cond=image,
+                    controlnet_mode=mode[:, None],
+                    conditioning_scale=scale,
+                    timestep=timestep,
+                    guidance=guidance,
+                    pooled_projections=pooled_projections,
+                    encoder_hidden_states=encoder_hidden_states,
+                    txt_ids=txt_ids,
+                    img_ids=img_ids,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                    return_dict=return_dict,
+                )
+
+                # merge samples
+                if i == 0:
+                    control_block_samples = block_samples
+                    control_single_block_samples = single_block_samples
+                else:
+                    control_block_samples = [
+                        control_block_sample + block_sample
+                        for control_block_sample, block_sample in zip(control_block_samples, block_samples)
+                    ]
+
+                    control_single_block_samples = [
+                        control_single_block_sample + block_sample
+                        for control_single_block_sample, block_sample in zip(
+                            control_single_block_samples, single_block_samples
+                        )
+                    ]
+
+        # Regular Multi-ControlNets
+        # load all ControlNets into memories
+        else:
+            for i, (image, mode, scale, controlnet) in enumerate(
+                zip(controlnet_cond, controlnet_mode, conditioning_scale, self.nets)
+            ):
+                block_samples, single_block_samples = controlnet(
+                    hidden_states=hidden_states,
+                    controlnet_cond=image,
+                    controlnet_mode=mode[:, None],
+                    conditioning_scale=scale,
+                    timestep=timestep,
+                    guidance=guidance,
+                    pooled_projections=pooled_projections,
+                    encoder_hidden_states=encoder_hidden_states,
+                    txt_ids=txt_ids,
+                    img_ids=img_ids,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                    return_dict=return_dict,
+                )
+
+                # merge samples
+                if i == 0:
+                    control_block_samples = block_samples
+                    control_single_block_samples = single_block_samples
+                else:
+                    control_block_samples = [
+                        control_block_sample + block_sample
+                        for control_block_sample, block_sample in zip(control_block_samples, block_samples)
+                    ]
+
+                    control_single_block_samples = [
+                        control_single_block_sample + block_sample
+                        for control_single_block_sample, block_sample in zip(
+                            control_single_block_samples, single_block_samples
+                        )
+                    ]
+
+        return control_block_samples, control_single_block_samples

src/diffusers/models/controlnet_sparsectrl.py

@@ -691,7 +691,6 @@ class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
 
         emb = self.time_embedding(t_emb, timestep_cond)
         emb = emb.repeat_interleave(sample_num_frames, dim=0)
-        encoder_hidden_states = encoder_hidden_states.repeat_interleave(sample_num_frames, dim=0)
 
         # 2. pre-process
         batch_size, channels, num_frames, height, width = sample.shape

src/diffusers/models/embeddings.py

@@ -374,6 +374,74 @@ class CogVideoXPatchEmbed(nn.Module):
         return embeds
 
 
+def get_3d_rotary_pos_embed(
+    embed_dim, crops_coords, grid_size, temporal_size, theta: int = 10000, use_real: bool = True
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    """
+    RoPE for video tokens with 3D structure.
+
+    Args:
+    embed_dim: (`int`):
+        The embedding dimension size, corresponding to hidden_size_head.
+    crops_coords (`Tuple[int]`):
+        The top-left and bottom-right coordinates of the crop.
+    grid_size (`Tuple[int]`):
+        The grid size of the spatial positional embedding (height, width).
+    temporal_size (`int`):
+        The size of the temporal dimension.
+    theta (`float`):
+        Scaling factor for frequency computation.
+
+    Returns:
+        `torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
+    """
+    if use_real is not True:
+        raise ValueError(" `use_real = False` is not currently supported for get_3d_rotary_pos_embed")
+    start, stop = crops_coords
+    grid_size_h, grid_size_w = grid_size
+    grid_h = np.linspace(start[0], stop[0], grid_size_h, endpoint=False, dtype=np.float32)
+    grid_w = np.linspace(start[1], stop[1], grid_size_w, endpoint=False, dtype=np.float32)
+    grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
+
+    # Compute dimensions for each axis
+    dim_t = embed_dim // 4
+    dim_h = embed_dim // 8 * 3
+    dim_w = embed_dim // 8 * 3
+
+    # Temporal frequencies
+    freqs_t = get_1d_rotary_pos_embed(dim_t, grid_t, use_real=True)
+    # Spatial frequencies for height and width
+    freqs_h = get_1d_rotary_pos_embed(dim_h, grid_h, use_real=True)
+    freqs_w = get_1d_rotary_pos_embed(dim_w, grid_w, use_real=True)
+
+    # BroadCast and concatenate temporal and spaial frequencie (height and width) into a 3d tensor
+    def combine_time_height_width(freqs_t, freqs_h, freqs_w):
+        freqs_t = freqs_t[:, None, None, :].expand(
+            -1, grid_size_h, grid_size_w, -1
+        )  # temporal_size, grid_size_h, grid_size_w, dim_t
+        freqs_h = freqs_h[None, :, None, :].expand(
+            temporal_size, -1, grid_size_w, -1
+        )  # temporal_size, grid_size_h, grid_size_2, dim_h
+        freqs_w = freqs_w[None, None, :, :].expand(
+            temporal_size, grid_size_h, -1, -1
+        )  # temporal_size, grid_size_h, grid_size_2, dim_w
+
+        freqs = torch.cat(
+            [freqs_t, freqs_h, freqs_w], dim=-1
+        )  # temporal_size, grid_size_h, grid_size_w, (dim_t + dim_h + dim_w)
+        freqs = freqs.view(
+            temporal_size * grid_size_h * grid_size_w, -1
+        )  # (temporal_size * grid_size_h * grid_size_w), (dim_t + dim_h + dim_w)
+        return freqs
+
+    t_cos, t_sin = freqs_t  # both t_cos and t_sin has shape: temporal_size, dim_t
+    h_cos, h_sin = freqs_h  # both h_cos and h_sin has shape: grid_size_h, dim_h
+    w_cos, w_sin = freqs_w  # both w_cos and w_sin has shape: grid_size_w, dim_w
+    cos = combine_time_height_width(t_cos, h_cos, w_cos)
+    sin = combine_time_height_width(t_sin, h_sin, w_sin)
+    return cos, sin
+
+
 def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
     """
     RoPE for image tokens with 2d structure.
@@ -446,6 +514,7 @@ def get_1d_rotary_pos_embed(
     linear_factor=1.0,
     ntk_factor=1.0,
     repeat_interleave_real=True,
+    freqs_dtype=torch.float32,  #  torch.float32, torch.float64 (flux)
 ):
     """
     Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
@@ -468,26 +537,34 @@ def get_1d_rotary_pos_embed(
         repeat_interleave_real (`bool`, *optional*, defaults to `True`):
             If `True` and `use_real`, real part and imaginary part are each interleaved with themselves to reach `dim`.
             Otherwise, they are concateanted with themselves.
+        freqs_dtype (`torch.float32` or `torch.float64`, *optional*, defaults to `torch.float32`):
+            the dtype of the frequency tensor.
     Returns:
         `torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
     """
     assert dim % 2 == 0
 
     if isinstance(pos, int):
-        pos = np.arange(pos)
+        pos = torch.arange(pos)
+    if isinstance(pos, np.ndarray):
+        pos = torch.from_numpy(pos)  # type: ignore  # [S]
+
     theta = theta * ntk_factor
-    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) / linear_factor  # [D/2]
-    t = torch.from_numpy(pos).to(freqs.device)  # type: ignore  # [S]
-    freqs = torch.outer(t, freqs).float()  # type: ignore   # [S, D/2]
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype)[: (dim // 2)] / dim)) / linear_factor  # [D/2]
+    freqs = freqs.to(pos.device)
+    freqs = torch.outer(pos, freqs)  # type: ignore   # [S, D/2]
     if use_real and repeat_interleave_real:
-        freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
-        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
+        # flux, hunyuan-dit, cogvideox
+        freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float()  # [S, D]
+        freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float()  # [S, D]
         return freqs_cos, freqs_sin
     elif use_real:
-        freqs_cos = torch.cat([freqs.cos(), freqs.cos()], dim=-1)  # [S, D]
-        freqs_sin = torch.cat([freqs.sin(), freqs.sin()], dim=-1)  # [S, D]
+        # stable audio
+        freqs_cos = torch.cat([freqs.cos(), freqs.cos()], dim=-1).float()  # [S, D]
+        freqs_sin = torch.cat([freqs.sin(), freqs.sin()], dim=-1).float()  # [S, D]
         return freqs_cos, freqs_sin
     else:
+        # lumina
         freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
         return freqs_cis
 
@@ -519,11 +596,11 @@ def apply_rotary_emb(
         cos, sin = cos.to(x.device), sin.to(x.device)
 
         if use_real_unbind_dim == -1:
-            # Use for example in Lumina
+            # Used for flux, cogvideox, hunyuan-dit
             x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
             x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
         elif use_real_unbind_dim == -2:
-            # Use for example in Stable Audio
+            # Used for Stable Audio
             x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
             x_rotated = torch.cat([-x_imag, x_real], dim=-1)
         else:
@@ -533,6 +610,7 @@ def apply_rotary_emb(
 
         return out
     else:
+        # used for lumina
         x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
         freqs_cis = freqs_cis.unsqueeze(2)
         x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
@@ -540,6 +618,31 @@ def apply_rotary_emb(
         return x_out.type_as(x)
 
 
+class FluxPosEmbed(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.squeeze().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], 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 TimestepEmbedding(nn.Module):
     def __init__(
         self,

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
 import torch
 from torch import nn
@@ -22,6 +22,7 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import is_torch_version, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import Attention, FeedForward
+from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
 from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -97,6 +98,7 @@ class CogVideoXBlock(nn.Module):
             eps=1e-6,
             bias=attention_bias,
             out_bias=attention_out_bias,
+            processor=CogVideoXAttnProcessor2_0(),
         )
 
         # 2. Feed Forward
@@ -116,24 +118,24 @@ class CogVideoXBlock(nn.Module):
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
         temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # norm & modulate
         norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
             hidden_states, encoder_hidden_states, temb
         )
 
         # attention
-        text_length = norm_encoder_hidden_states.size(1)
-
-        # CogVideoX uses concatenated text + video embeddings with self-attention instead of using
-        # them in cross-attention individually
-        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
-        attn_output = self.attn1(
+        attn_hidden_states, attn_encoder_hidden_states = self.attn1(
             hidden_states=norm_hidden_states,
-            encoder_hidden_states=None,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
         )
 
-        hidden_states = hidden_states + gate_msa * attn_output[:, text_length:]
-        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_output[:, :text_length]
+        hidden_states = hidden_states + gate_msa * attn_hidden_states
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
 
         # norm & modulate
         norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
@@ -144,8 +146,9 @@ class CogVideoXBlock(nn.Module):
         norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
         ff_output = self.ff(norm_hidden_states)
 
-        hidden_states = hidden_states + gate_ff * ff_output[:, text_length:]
-        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_length]
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
         return hidden_states, encoder_hidden_states
 
 
@@ -231,6 +234,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         norm_eps: float = 1e-5,
         spatial_interpolation_scale: float = 1.875,
         temporal_interpolation_scale: float = 1.0,
+        use_rotary_positional_embeddings: bool = False,
     ):
         super().__init__()
         inner_dim = num_attention_heads * attention_head_dim
@@ -295,12 +299,113 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
     def _set_gradient_checkpointing(self, module, value=False):
         self.gradient_checkpointing = value
 
+    @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.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_0
+    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.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        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(FusedCogVideoXAttnProcessor2_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.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def forward(
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
         timestep: Union[int, float, torch.LongTensor],
         timestep_cond: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
         return_dict: bool = True,
     ):
         batch_size, num_frames, channels, height, width = hidden_states.shape
@@ -319,14 +424,16 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
 
         # 3. Position embedding
-        seq_length = height * width * num_frames // (self.config.patch_size**2)
+        text_seq_length = encoder_hidden_states.shape[1]
+        if not self.config.use_rotary_positional_embeddings:
+            seq_length = height * width * num_frames // (self.config.patch_size**2)
 
-        pos_embeds = self.pos_embedding[:, : self.config.max_text_seq_length + seq_length]
-        hidden_states = hidden_states + pos_embeds
-        hidden_states = self.embedding_dropout(hidden_states)
+            pos_embeds = self.pos_embedding[:, : text_seq_length + seq_length]
+            hidden_states = hidden_states + pos_embeds
+            hidden_states = self.embedding_dropout(hidden_states)
 
-        encoder_hidden_states = hidden_states[:, : self.config.max_text_seq_length]
-        hidden_states = hidden_states[:, self.config.max_text_seq_length :]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
 
         # 4. Transformer blocks
         for i, block in enumerate(self.transformer_blocks):
@@ -344,6 +451,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
                     hidden_states,
                     encoder_hidden_states,
                     emb,
+                    image_rotary_emb,
                     **ckpt_kwargs,
                 )
             else:
@@ -351,9 +459,17 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
                     hidden_states=hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
                     temb=emb,
+                    image_rotary_emb=image_rotary_emb,
                 )
 
-        hidden_states = self.norm_final(hidden_states)
+        if not self.config.use_rotary_positional_embeddings:
+            # CogVideoX-2B
+            hidden_states = self.norm_final(hidden_states)
+        else:
+            # CogVideoX-5B
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            hidden_states = self.norm_final(hidden_states)
+            hidden_states = hidden_states[:, text_seq_length:]
 
         # 5. Final block
         hidden_states = self.norm_out(hidden_states, temb=emb)

src/diffusers/models/transformers/transformer_flux.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -23,52 +23,23 @@ import torch.nn.functional as F
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...models.attention import FeedForward
-from ...models.attention_processor import Attention, FluxAttnProcessor2_0, FluxSingleAttnProcessor2_0
+from ...models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    FluxAttnProcessor2_0,
+    FusedFluxAttnProcessor2_0,
+)
 from ...models.modeling_utils import ModelMixin
 from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
 from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
+from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-# YiYi to-do: refactor rope related functions/classes
-def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
-    assert dim % 2 == 0, "The dimension must be even."
-
-    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
-    omega = 1.0 / (theta**scale)
-
-    batch_size, seq_length = pos.shape
-    out = torch.einsum("...n,d->...nd", pos, omega)
-    cos_out = torch.cos(out)
-    sin_out = torch.sin(out)
-
-    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
-    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
-    return out.float()
-
-
-# YiYi to-do: refactor rope related functions/classes
-class EmbedND(nn.Module):
-    def __init__(self, dim: int, theta: int, axes_dim: List[int]):
-        super().__init__()
-        self.dim = dim
-        self.theta = theta
-        self.axes_dim = axes_dim
-
-    def forward(self, ids: torch.Tensor) -> torch.Tensor:
-        n_axes = ids.shape[-1]
-        emb = torch.cat(
-            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
-            dim=-3,
-        )
-        return emb.unsqueeze(1)
-
-
 @maybe_allow_in_graph
 class FluxSingleTransformerBlock(nn.Module):
     r"""
@@ -93,7 +64,7 @@ class FluxSingleTransformerBlock(nn.Module):
         self.act_mlp = nn.GELU(approximate="tanh")
         self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
 
-        processor = FluxSingleAttnProcessor2_0()
+        processor = FluxAttnProcessor2_0()
         self.attn = Attention(
             query_dim=dim,
             cross_attention_dim=None,
@@ -265,13 +236,14 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         joint_attention_dim: int = 4096,
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
-        axes_dims_rope: List[int] = [16, 56, 56],
+        axes_dims_rope: Tuple[int] = (16, 56, 56),
     ):
         super().__init__()
         self.out_channels = in_channels
         self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
 
-        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
+        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
+
         text_time_guidance_cls = (
             CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
         )
@@ -309,6 +281,106 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         self.gradient_checkpointing = False
 
+    @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.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
+    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.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        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(FusedFluxAttnProcessor2_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.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def _set_gradient_checkpointing(self, module, value=False):
         if hasattr(module, "gradient_checkpointing"):
             module.gradient_checkpointing = value
@@ -381,8 +453,19 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
-        txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
-        ids = torch.cat((txt_ids, img_ids), dim=1)
+        if txt_ids.ndim == 3:
+            logger.warning(
+                "Passing `txt_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            txt_ids = txt_ids[0]
+        if img_ids.ndim == 3:
+            logger.warning(
+                "Passing `img_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            img_ids = img_ids[0]
+        ids = torch.cat((txt_ids, img_ids), dim=0)
         image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):

src/diffusers/models/unets/unet_motion_model.py

@@ -116,7 +116,7 @@ class AnimateDiffTransformer3D(nn.Module):
 
         self.in_channels = in_channels
 
-        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        self.norm = nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
         self.proj_in = nn.Linear(in_channels, inner_dim)
 
         # 3. Define transformers blocks
@@ -2178,7 +2178,6 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
 
         emb = emb if aug_emb is None else emb + aug_emb
         emb = emb.repeat_interleave(repeats=num_frames, dim=0)
-        encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
 
         if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
             if "image_embeds" not in added_cond_kwargs:

src/diffusers/pipelines/__init__.py

@@ -154,6 +154,7 @@ else:
             "StableDiffusionControlNetPAGPipeline",
             "StableDiffusionXLPAGPipeline",
             "StableDiffusionXLPAGInpaintPipeline",
+            "StableDiffusionXLControlNetPAGImg2ImgPipeline",
             "StableDiffusionXLControlNetPAGPipeline",
             "StableDiffusionXLPAGImg2ImgPipeline",
             "PixArtSigmaPAGPipeline",
@@ -547,6 +548,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusion3PAGPipeline,
             StableDiffusionControlNetPAGPipeline,
             StableDiffusionPAGPipeline,
+            StableDiffusionXLControlNetPAGImg2ImgPipeline,
             StableDiffusionXLControlNetPAGPipeline,
             StableDiffusionXLPAGImg2ImgPipeline,
             StableDiffusionXLPAGInpaintPipeline,

src/diffusers/pipelines/animatediff/pipeline_animatediff.py

@@ -432,7 +432,6 @@ class AnimateDiffPipeline(
             extra_step_kwargs["generator"] = generator
         return extra_step_kwargs
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
     def check_inputs(
         self,
         prompt,
@@ -470,8 +469,8 @@ class AnimateDiffPipeline(
             raise ValueError(
                 "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
             )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt is not None and not isinstance(prompt, (str, list, dict)):
+            raise ValueError(f"`prompt` has to be of type `str`, `list` or `dict` but is {type(prompt)=}")
 
         if negative_prompt is not None and negative_prompt_embeds is not None:
             raise ValueError(
@@ -557,11 +556,15 @@ class AnimateDiffPipeline(
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
         self,
-        prompt: Union[str, List[str]] = None,
+        prompt: Optional[Union[str, List[str]]] = None,
         num_frames: Optional[int] = 16,
         height: Optional[int] = None,
         width: Optional[int] = None,
@@ -701,9 +704,10 @@ class AnimateDiffPipeline(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
+        if prompt is not None and isinstance(prompt, (str, dict)):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
             batch_size = len(prompt)
@@ -716,22 +720,39 @@ class AnimateDiffPipeline(
         text_encoder_lora_scale = (
             self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
         )
-        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
-            prompt,
-            device,
-            num_videos_per_prompt,
-            self.do_classifier_free_guidance,
-            negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=text_encoder_lora_scale,
-            clip_skip=self.clip_skip,
-        )
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and text embeddings into a single batch
-        # to avoid doing two forward passes
-        if self.do_classifier_free_guidance:
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+        if self.free_noise_enabled:
+            prompt_embeds, negative_prompt_embeds = self._encode_prompt_free_noise(
+                prompt=prompt,
+                num_frames=num_frames,
+                device=device,
+                num_videos_per_prompt=num_videos_per_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                negative_prompt=negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
+        else:
+            prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+                prompt,
+                device,
+                num_videos_per_prompt,
+                self.do_classifier_free_guidance,
+                negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            if self.do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+            prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
 
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
@@ -783,6 +804,9 @@ class AnimateDiffPipeline(
             # 8. Denoising loop
             with self.progress_bar(total=self._num_timesteps) as progress_bar:
                 for i, t in enumerate(timesteps):
+                    if self.interrupt:
+                        continue
+
                     # expand the latents if we are doing classifier free guidance
                     latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

src/diffusers/pipelines/animatediff/pipeline_animatediff_controlnet.py

@@ -505,8 +505,8 @@ class AnimateDiffControlNetPipeline(
             raise ValueError(
                 "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
             )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt is not None and not isinstance(prompt, (str, list, dict)):
+            raise ValueError(f"`prompt` has to be of type `str`, `list` or `dict` but is {type(prompt)}")
 
         if negative_prompt is not None and negative_prompt_embeds is not None:
             raise ValueError(
@@ -699,6 +699,10 @@ class AnimateDiffControlNetPipeline(
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     def __call__(
         self,
@@ -858,9 +862,10 @@ class AnimateDiffControlNetPipeline(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
+        if prompt is not None and isinstance(prompt, (str, dict)):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
             batch_size = len(prompt)
@@ -883,22 +888,39 @@ class AnimateDiffControlNetPipeline(
         text_encoder_lora_scale = (
             cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
         )
-        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
-            prompt,
-            device,
-            num_videos_per_prompt,
-            self.do_classifier_free_guidance,
-            negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=text_encoder_lora_scale,
-            clip_skip=self.clip_skip,
-        )
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and text embeddings into a single batch
-        # to avoid doing two forward passes
-        if self.do_classifier_free_guidance:
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+        if self.free_noise_enabled:
+            prompt_embeds, negative_prompt_embeds = self._encode_prompt_free_noise(
+                prompt=prompt,
+                num_frames=num_frames,
+                device=device,
+                num_videos_per_prompt=num_videos_per_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                negative_prompt=negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
+        else:
+            prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+                prompt,
+                device,
+                num_videos_per_prompt,
+                self.do_classifier_free_guidance,
+                negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            if self.do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+            prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
 
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
@@ -990,6 +1012,9 @@ class AnimateDiffControlNetPipeline(
             # 8. Denoising loop
             with self.progress_bar(total=self._num_timesteps) as progress_bar:
                 for i, t in enumerate(timesteps):
+                    if self.interrupt:
+                        continue
+
                     # expand the latents if we are doing classifier free guidance
                     latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
@@ -1002,7 +1027,6 @@ class AnimateDiffControlNetPipeline(
                     else:
                         control_model_input = latent_model_input
                         controlnet_prompt_embeds = prompt_embeds
-                    controlnet_prompt_embeds = controlnet_prompt_embeds.repeat_interleave(num_frames, dim=0)
 
                     if isinstance(controlnet_keep[i], list):
                         cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]

src/diffusers/pipelines/animatediff/pipeline_animatediff_sdxl.py

@@ -1143,6 +1143,8 @@ class AnimateDiffSDXLPipeline(
             add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
             add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
 
+        prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
+
         prompt_embeds = prompt_embeds.to(device)
         add_text_embeds = add_text_embeds.to(device)
         add_time_ids = add_time_ids.to(device).repeat(batch_size * num_videos_per_prompt, 1)

src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py

@@ -878,6 +878,8 @@ class AnimateDiffSparseControlNetPipeline(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+        prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
+
         # 4. Prepare IP-Adapter embeddings
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(

src/diffusers/pipelines/animatediff/pipeline_animatediff_video2video.py

@@ -246,7 +246,6 @@ class AnimateDiffVideoToVideoPipeline(
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor)
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
     def encode_prompt(
         self,
         prompt,
@@ -299,7 +298,7 @@ class AnimateDiffVideoToVideoPipeline(
             else:
                 scale_lora_layers(self.text_encoder, lora_scale)
 
-        if prompt is not None and isinstance(prompt, str):
+        if prompt is not None and isinstance(prompt, (str, dict)):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
             batch_size = len(prompt)
@@ -582,8 +581,8 @@ class AnimateDiffVideoToVideoPipeline(
             raise ValueError(
                 "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
             )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt is not None and not isinstance(prompt, (str, list, dict)):
+            raise ValueError(f"`prompt` has to be of type `str`, `list` or `dict` but is {type(prompt)}")
 
         if negative_prompt is not None and negative_prompt_embeds is not None:
             raise ValueError(
@@ -628,23 +627,20 @@ class AnimateDiffVideoToVideoPipeline(
 
     def prepare_latents(
         self,
-        video,
-        height,
-        width,
-        num_channels_latents,
-        batch_size,
-        timestep,
-        dtype,
-        device,
-        generator,
-        latents=None,
+        video: Optional[torch.Tensor] = None,
+        height: int = 64,
+        width: int = 64,
+        num_channels_latents: int = 4,
+        batch_size: int = 1,
+        timestep: Optional[int] = None,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
         decode_chunk_size: int = 16,
-    ):
-        if latents is None:
-            num_frames = video.shape[1]
-        else:
-            num_frames = latents.shape[2]
-
+        add_noise: bool = False,
+    ) -> torch.Tensor:
+        num_frames = video.shape[1] if latents is None else latents.shape[2]
         shape = (
             batch_size,
             num_channels_latents,
@@ -708,8 +704,13 @@ class AnimateDiffVideoToVideoPipeline(
             if shape != latents.shape:
                 # [B, C, F, H, W]
                 raise ValueError(f"`latents` expected to have {shape=}, but found {latents.shape=}")
+
             latents = latents.to(device, dtype=dtype)
 
+            if add_noise:
+                noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+                latents = self.scheduler.add_noise(latents, noise, timestep)
+
         return latents
 
     @property
@@ -735,6 +736,10 @@ class AnimateDiffVideoToVideoPipeline(
     def num_timesteps(self):
         return self._num_timesteps
 
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     def __call__(
         self,
@@ -743,6 +748,7 @@ class AnimateDiffVideoToVideoPipeline(
         height: Optional[int] = None,
         width: Optional[int] = None,
         num_inference_steps: int = 50,
+        enforce_inference_steps: bool = False,
         timesteps: Optional[List[int]] = None,
         sigmas: Optional[List[float]] = None,
         guidance_scale: float = 7.5,
@@ -874,9 +880,10 @@ class AnimateDiffVideoToVideoPipeline(
         self._guidance_scale = guidance_scale
         self._clip_skip = clip_skip
         self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
 
         # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
+        if prompt is not None and isinstance(prompt, (str, dict)):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
             batch_size = len(prompt)
@@ -884,29 +891,85 @@ class AnimateDiffVideoToVideoPipeline(
             batch_size = prompt_embeds.shape[0]
 
         device = self._execution_device
+        dtype = self.dtype
 
-        # 3. Encode input prompt
+        # 3. Prepare timesteps
+        if not enforce_inference_steps:
+            timesteps, num_inference_steps = retrieve_timesteps(
+                self.scheduler, num_inference_steps, device, timesteps, sigmas
+            )
+            timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
+            latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+        else:
+            denoising_inference_steps = int(num_inference_steps / strength)
+            timesteps, denoising_inference_steps = retrieve_timesteps(
+                self.scheduler, denoising_inference_steps, device, timesteps, sigmas
+            )
+            timesteps = timesteps[-num_inference_steps:]
+            latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+
+        # 4. Prepare latent variables
+        if latents is None:
+            video = self.video_processor.preprocess_video(video, height=height, width=width)
+            # Move the number of frames before the number of channels.
+            video = video.permute(0, 2, 1, 3, 4)
+            video = video.to(device=device, dtype=dtype)
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            video=video,
+            height=height,
+            width=width,
+            num_channels_latents=num_channels_latents,
+            batch_size=batch_size * num_videos_per_prompt,
+            timestep=latent_timestep,
+            dtype=dtype,
+            device=device,
+            generator=generator,
+            latents=latents,
+            decode_chunk_size=decode_chunk_size,
+            add_noise=enforce_inference_steps,
+        )
+
+        # 5. Encode input prompt
         text_encoder_lora_scale = (
             self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
         )
-        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
-            prompt,
-            device,
-            num_videos_per_prompt,
-            self.do_classifier_free_guidance,
-            negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=text_encoder_lora_scale,
-            clip_skip=self.clip_skip,
-        )
+        num_frames = latents.shape[2]
+        if self.free_noise_enabled:
+            prompt_embeds, negative_prompt_embeds = self._encode_prompt_free_noise(
+                prompt=prompt,
+                num_frames=num_frames,
+                device=device,
+                num_videos_per_prompt=num_videos_per_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                negative_prompt=negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
+        else:
+            prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+                prompt,
+                device,
+                num_videos_per_prompt,
+                self.do_classifier_free_guidance,
+                negative_prompt,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                lora_scale=text_encoder_lora_scale,
+                clip_skip=self.clip_skip,
+            )
 
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and text embeddings into a single batch
-        # to avoid doing two forward passes
-        if self.do_classifier_free_guidance:
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            if self.do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+            prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
+
+        # 6. Prepare IP-Adapter embeddings
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
                 ip_adapter_image,
@@ -916,38 +979,10 @@ class AnimateDiffVideoToVideoPipeline(
                 self.do_classifier_free_guidance,
             )
 
-        # 4. Prepare timesteps
-        timesteps, num_inference_steps = retrieve_timesteps(
-            self.scheduler, num_inference_steps, device, timesteps, sigmas
-        )
-        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
-        latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
-
-        # 5. Prepare latent variables
-        if latents is None:
-            video = self.video_processor.preprocess_video(video, height=height, width=width)
-            # Move the number of frames before the number of channels.
-            video = video.permute(0, 2, 1, 3, 4)
-            video = video.to(device=device, dtype=prompt_embeds.dtype)
-        num_channels_latents = self.unet.config.in_channels
-        latents = self.prepare_latents(
-            video=video,
-            height=height,
-            width=width,
-            num_channels_latents=num_channels_latents,
-            batch_size=batch_size * num_videos_per_prompt,
-            timestep=latent_timestep,
-            dtype=prompt_embeds.dtype,
-            device=device,
-            generator=generator,
-            latents=latents,
-            decode_chunk_size=decode_chunk_size,
-        )
-
-        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
-        # 7. Add image embeds for IP-Adapter
+        # 8. Add image embeds for IP-Adapter
         added_cond_kwargs = (
             {"image_embeds": image_embeds}
             if ip_adapter_image is not None or ip_adapter_image_embeds is not None
@@ -967,9 +1002,12 @@ class AnimateDiffVideoToVideoPipeline(
             self._num_timesteps = len(timesteps)
             num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
 
-            # 8. Denoising loop
+            # 9. Denoising loop
             with self.progress_bar(total=self._num_timesteps) as progress_bar:
                 for i, t in enumerate(timesteps):
+                    if self.interrupt:
+                        continue
+
                     # expand the latents if we are doing classifier free guidance
                     latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
@@ -1005,14 +1043,14 @@ class AnimateDiffVideoToVideoPipeline(
                     if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                         progress_bar.update()
 
-        # 9. Post-processing
+        # 10. Post-processing
         if output_type == "latent":
             video = latents
         else:
             video_tensor = self.decode_latents(latents, decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
 
-        # 10. Offload all models
+        # 11. Offload all models
         self.maybe_free_model_hooks()
 
         if not return_dict:

src/diffusers/pipelines/auto_pipeline.py

@@ -56,6 +56,7 @@ from .pag import (
     StableDiffusion3PAGPipeline,
     StableDiffusionControlNetPAGPipeline,
     StableDiffusionPAGPipeline,
+    StableDiffusionXLControlNetPAGImg2ImgPipeline,
     StableDiffusionXLControlNetPAGPipeline,
     StableDiffusionXLPAGImg2ImgPipeline,
     StableDiffusionXLPAGInpaintPipeline,
@@ -123,6 +124,7 @@ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
         ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
         ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
+        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),
         ("lcm", LatentConsistencyModelImg2ImgPipeline),
     ]
 )

src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -23,6 +23,7 @@ from transformers import T5EncoderModel, T5Tokenizer
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
 from ...utils import BaseOutput, logging, replace_example_docstring
@@ -40,6 +41,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import CogVideoXPipeline
         >>> from diffusers.utils import export_to_video
 
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
         >>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
         >>> prompt = (
         ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
@@ -55,6 +57,25 @@ EXAMPLE_DOC_STRING = """
 """
 
 
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
 def retrieve_timesteps(
     scheduler,
@@ -409,6 +430,45 @@ class CogVideoXPipeline(DiffusionPipeline):
                     f" {negative_prompt_embeds.shape}."
                 )
 
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        grid_crops_coords = get_resize_crop_region_for_grid(
+            (grid_height, grid_width), base_size_width, base_size_height
+        )
+        freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+            embed_dim=self.transformer.config.attention_head_dim,
+            crops_coords=grid_crops_coords,
+            grid_size=(grid_height, grid_width),
+            temporal_size=num_frames,
+        )
+
+        freqs_cos = freqs_cos.to(device=device)
+        freqs_sin = freqs_sin.to(device=device)
+        return freqs_cos, freqs_sin
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -599,7 +659,14 @@ class CogVideoXPipeline(DiffusionPipeline):
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
-        # 7. Denoising loop
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
 
         with self.progress_bar(total=num_inference_steps) as progress_bar:
@@ -620,6 +687,7 @@ class CogVideoXPipeline(DiffusionPipeline):
                     hidden_states=latent_model_input,
                     encoder_hidden_states=prompt_embeds,
                     timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
                     return_dict=False,
                 )[0]
                 noise_pred = noise_pred.float()

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -1538,7 +1538,6 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
                     if isinstance(controlnet_cond_scale, list):
                         controlnet_cond_scale = controlnet_cond_scale[0]
                     cond_scale = controlnet_cond_scale * controlnet_keep[i]
-
                 down_block_res_samples, mid_block_res_sample = self.controlnet(
                     control_model_input,
                     t,

src/diffusers/pipelines/flux/pipeline_flux.py

@@ -20,7 +20,7 @@ import torch
 from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
 
 from ...image_processor import VaeImageProcessor
-from ...loaders import FluxLoraLoaderMixin
+from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin
 from ...models.autoencoders import AutoencoderKL
 from ...models.transformers import FluxTransformer2DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
@@ -137,7 +137,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
+class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
     r"""
     The Flux pipeline for text-to-image generation.
 
@@ -280,7 +280,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
 
         # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt)
         prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)
 
         return prompt_embeds
@@ -331,10 +331,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 scale_lora_layers(self.text_encoder_2, lora_scale)
 
         prompt = [prompt] if isinstance(prompt, str) else prompt
-        if prompt is not None:
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
 
         if prompt_embeds is None:
             prompt_2 = prompt_2 or prompt
@@ -364,8 +360,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 unscale_lora_layers(self.text_encoder_2, lora_scale)
 
         dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
-        text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
-        text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
 
         return prompt_embeds, pooled_prompt_embeds, text_ids
 
@@ -425,9 +420,8 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
 
         latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
 
-        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
         latent_image_ids = latent_image_ids.reshape(
-            batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
+            latent_image_id_height * latent_image_id_width, latent_image_id_channels
         )
 
         return latent_image_ids.to(device=device, dtype=dtype)
@@ -542,7 +536,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         width: Optional[int] = None,
         num_inference_steps: int = 28,
         timesteps: List[int] = None,
-        guidance_scale: float = 7.0,
+        guidance_scale: float = 3.5,
         num_images_per_prompt: Optional[int] = 1,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
@@ -724,7 +718,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
 
                 noise_pred = self.transformer(
                     hidden_states=latents,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                     timestep=timestep / 1000,
                     guidance=guidance,
                     pooled_projections=pooled_prompt_embeds,

src/diffusers/pipelines/flux/pipeline_flux_controlnet.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import inspect
-from typing import Any, Callable, Dict, List, Optional, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -25,9 +25,9 @@ from transformers import (
 )
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FluxLoraLoaderMixin
+from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin
 from ...models.autoencoders import AutoencoderKL
-from ...models.controlnet_flux import FluxControlNetModel
+from ...models.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel
 from ...models.transformers import FluxTransformer2DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
 from ...utils import (
@@ -61,7 +61,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import FluxControlNetPipeline
         >>> from diffusers import FluxControlNetModel
 
-        >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny-alpha"
+        >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny"
         >>> controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
         >>> pipe = FluxControlNetPipeline.from_pretrained(
         ...     base_model, controlnet=controlnet, torch_dtype=torch.bfloat16
@@ -155,7 +155,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
+class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
     r"""
     The Flux pipeline for text-to-image generation.
 
@@ -195,7 +195,9 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         text_encoder_2: T5EncoderModel,
         tokenizer_2: T5TokenizerFast,
         transformer: FluxTransformer2DModel,
-        controlnet: FluxControlNetModel,
+        controlnet: Union[
+            FluxControlNetModel, List[FluxControlNetModel], Tuple[FluxControlNetModel], FluxMultiControlNetModel
+        ],
     ):
         super().__init__()
 
@@ -300,7 +302,7 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
 
         # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt)
         prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)
 
         return prompt_embeds
@@ -354,10 +356,6 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 scale_lora_layers(self.text_encoder_2, lora_scale)
 
         prompt = [prompt] if isinstance(prompt, str) else prompt
-        if prompt is not None:
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
 
         if prompt_embeds is None:
             prompt_2 = prompt_2 or prompt
@@ -387,8 +385,7 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 unscale_lora_layers(self.text_encoder_2, lora_scale)
 
         dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
-        text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
-        text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
 
         return prompt_embeds, pooled_prompt_embeds, text_ids
 
@@ -449,9 +446,8 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
 
         latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
 
-        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
         latent_image_ids = latent_image_ids.reshape(
-            batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
+            latent_image_id_height * latent_image_id_width, latent_image_id_channels
         )
 
         return latent_image_ids.to(device=device, dtype=dtype)
@@ -577,6 +573,7 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         timesteps: List[int] = None,
         guidance_scale: float = 7.0,
         control_image: PipelineImageInput = None,
+        control_mode: Optional[Union[int, List[int]]] = None,
         controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
         num_images_per_prompt: Optional[int] = 1,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
@@ -617,6 +614,20 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
+            control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+                    `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted
+                as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or
+                width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`,
+                images must be passed as a list such that each element of the list can be correctly batched for input
+                to a single ControlNet.
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
+            control_mode (`int` or `List[int]`,, *optional*, defaults to None):
+                The control mode when applying ControlNet-Union.
             num_images_per_prompt (`int`, *optional*, defaults to 1):
                 The number of images to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -736,6 +747,55 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 width_control_image,
             )
 
+            # set control mode
+            if control_mode is not None:
+                control_mode = torch.tensor(control_mode).to(device, dtype=torch.long)
+                control_mode = control_mode.reshape([-1, 1])
+
+        elif isinstance(self.controlnet, FluxMultiControlNetModel):
+            control_images = []
+
+            for control_image_ in control_image:
+                control_image_ = self.prepare_image(
+                    image=control_image_,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=dtype,
+                )
+                height, width = control_image_.shape[-2:]
+
+                # vae encode
+                control_image_ = self.vae.encode(control_image_).latent_dist.sample()
+                control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+
+                # pack
+                height_control_image, width_control_image = control_image_.shape[2:]
+                control_image_ = self._pack_latents(
+                    control_image_,
+                    batch_size * num_images_per_prompt,
+                    num_channels_latents,
+                    height_control_image,
+                    width_control_image,
+                )
+
+                control_images.append(control_image_)
+
+            control_image = control_images
+
+            # set control mode
+            control_mode_ = []
+            if isinstance(control_mode, list):
+                for cmode in control_mode:
+                    if cmode is None:
+                        control_mode_.append(-1)
+                    else:
+                        control_mode_.append(cmode)
+            control_mode = torch.tensor(control_mode_).to(device, dtype=torch.long)
+            control_mode = control_mode.reshape([-1, 1])
+
         # 4. Prepare latent variables
         num_channels_latents = self.transformer.config.in_channels // 4
         latents, latent_image_ids = self.prepare_latents(
@@ -791,6 +851,7 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 controlnet_block_samples, controlnet_single_block_samples = self.controlnet(
                     hidden_states=latents,
                     controlnet_cond=control_image,
+                    controlnet_mode=control_mode,
                     conditioning_scale=controlnet_conditioning_scale,
                     timestep=timestep / 1000,
                     guidance=guidance,
@@ -804,7 +865,6 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
 
                 noise_pred = self.transformer(
                     hidden_states=latents,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                     timestep=timestep / 1000,
                     guidance=guidance,
                     pooled_projections=pooled_prompt_embeds,

src/diffusers/pipelines/free_noise_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional, Union
+from typing import Callable, Dict, Optional, Union
 
 import torch
 
@@ -22,6 +22,7 @@ from ..models.unets.unet_motion_model import (
     DownBlockMotion,
     UpBlockMotion,
 )
+from ..pipelines.pipeline_utils import DiffusionPipeline
 from ..utils import logging
 from ..utils.torch_utils import randn_tensor
 
@@ -98,6 +99,142 @@ class AnimateDiffFreeNoiseMixin:
                         free_noise_transfomer_block.state_dict(), strict=True
                     )
 
+    def _check_inputs_free_noise(
+        self,
+        prompt,
+        negative_prompt,
+        prompt_embeds,
+        negative_prompt_embeds,
+        num_frames,
+    ) -> None:
+        if not isinstance(prompt, (str, dict)):
+            raise ValueError(f"Expected `prompt` to have type `str` or `dict` but found {type(prompt)=}")
+
+        if negative_prompt is not None:
+            if not isinstance(negative_prompt, (str, dict)):
+                raise ValueError(
+                    f"Expected `negative_prompt` to have type `str` or `dict` but found {type(negative_prompt)=}"
+                )
+
+        if prompt_embeds is not None or negative_prompt_embeds is not None:
+            raise ValueError("`prompt_embeds` and `negative_prompt_embeds` is not supported in FreeNoise yet.")
+
+        frame_indices = [isinstance(x, int) for x in prompt.keys()]
+        frame_prompts = [isinstance(x, str) for x in prompt.values()]
+        min_frame = min(list(prompt.keys()))
+        max_frame = max(list(prompt.keys()))
+
+        if not all(frame_indices):
+            raise ValueError("Expected integer keys in `prompt` dict for FreeNoise.")
+        if not all(frame_prompts):
+            raise ValueError("Expected str values in `prompt` dict for FreeNoise.")
+        if min_frame != 0:
+            raise ValueError("The minimum frame index in `prompt` dict must be 0 as a starting prompt is necessary.")
+        if max_frame >= num_frames:
+            raise ValueError(
+                f"The maximum frame index in `prompt` dict must be lesser than {num_frames=} and follow 0-based indexing."
+            )
+
+    def _encode_prompt_free_noise(
+        self,
+        prompt: Union[str, Dict[int, str]],
+        num_frames: int,
+        device: torch.device,
+        num_videos_per_prompt: int,
+        do_classifier_free_guidance: bool,
+        negative_prompt: Optional[Union[str, Dict[int, str]]] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ) -> torch.Tensor:
+        if negative_prompt is None:
+            negative_prompt = ""
+
+        # Ensure that we have a dictionary of prompts
+        if isinstance(prompt, str):
+            prompt = {0: prompt}
+        if isinstance(negative_prompt, str):
+            negative_prompt = {0: negative_prompt}
+
+        self._check_inputs_free_noise(prompt, negative_prompt, prompt_embeds, negative_prompt_embeds, num_frames)
+
+        # Sort the prompts based on frame indices
+        prompt = dict(sorted(prompt.items()))
+        negative_prompt = dict(sorted(negative_prompt.items()))
+
+        # Ensure that we have a prompt for the last frame index
+        prompt[num_frames - 1] = prompt[list(prompt.keys())[-1]]
+        negative_prompt[num_frames - 1] = negative_prompt[list(negative_prompt.keys())[-1]]
+
+        frame_indices = list(prompt.keys())
+        frame_prompts = list(prompt.values())
+        frame_negative_indices = list(negative_prompt.keys())
+        frame_negative_prompts = list(negative_prompt.values())
+
+        # Generate and interpolate positive prompts
+        prompt_embeds, _ = self.encode_prompt(
+            prompt=frame_prompts,
+            device=device,
+            num_images_per_prompt=num_videos_per_prompt,
+            do_classifier_free_guidance=False,
+            negative_prompt=None,
+            prompt_embeds=None,
+            negative_prompt_embeds=None,
+            lora_scale=lora_scale,
+            clip_skip=clip_skip,
+        )
+
+        shape = (num_frames, *prompt_embeds.shape[1:])
+        prompt_interpolation_embeds = prompt_embeds.new_zeros(shape)
+
+        for i in range(len(frame_indices) - 1):
+            start_frame = frame_indices[i]
+            end_frame = frame_indices[i + 1]
+            start_tensor = prompt_embeds[i].unsqueeze(0)
+            end_tensor = prompt_embeds[i + 1].unsqueeze(0)
+
+            prompt_interpolation_embeds[start_frame : end_frame + 1] = self._free_noise_prompt_interpolation_callback(
+                start_frame, end_frame, start_tensor, end_tensor
+            )
+
+        # Generate and interpolate negative prompts
+        negative_prompt_embeds = None
+        negative_prompt_interpolation_embeds = None
+
+        if do_classifier_free_guidance:
+            _, negative_prompt_embeds = self.encode_prompt(
+                prompt=[""] * len(frame_negative_prompts),
+                device=device,
+                num_images_per_prompt=num_videos_per_prompt,
+                do_classifier_free_guidance=True,
+                negative_prompt=frame_negative_prompts,
+                prompt_embeds=None,
+                negative_prompt_embeds=None,
+                lora_scale=lora_scale,
+                clip_skip=clip_skip,
+            )
+
+            negative_prompt_interpolation_embeds = negative_prompt_embeds.new_zeros(shape)
+
+            for i in range(len(frame_negative_indices) - 1):
+                start_frame = frame_negative_indices[i]
+                end_frame = frame_negative_indices[i + 1]
+                start_tensor = negative_prompt_embeds[i].unsqueeze(0)
+                end_tensor = negative_prompt_embeds[i + 1].unsqueeze(0)
+
+                negative_prompt_interpolation_embeds[
+                    start_frame : end_frame + 1
+                ] = self._free_noise_prompt_interpolation_callback(start_frame, end_frame, start_tensor, end_tensor)
+
+        prompt_embeds = prompt_interpolation_embeds
+        negative_prompt_embeds = negative_prompt_interpolation_embeds
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        return prompt_embeds, negative_prompt_embeds
+
     def _prepare_latents_free_noise(
         self,
         batch_size: int,
@@ -172,12 +309,29 @@ class AnimateDiffFreeNoiseMixin:
         latents = latents[:, :, :num_frames]
         return latents
 
+    def _lerp(
+        self, start_index: int, end_index: int, start_tensor: torch.Tensor, end_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        num_indices = end_index - start_index + 1
+        interpolated_tensors = []
+
+        for i in range(num_indices):
+            alpha = i / (num_indices - 1)
+            interpolated_tensor = (1 - alpha) * start_tensor + alpha * end_tensor
+            interpolated_tensors.append(interpolated_tensor)
+
+        interpolated_tensors = torch.cat(interpolated_tensors)
+        return interpolated_tensors
+
     def enable_free_noise(
         self,
         context_length: Optional[int] = 16,
         context_stride: int = 4,
         weighting_scheme: str = "pyramid",
         noise_type: str = "shuffle_context",
+        prompt_interpolation_callback: Optional[
+            Callable[[DiffusionPipeline, int, int, torch.Tensor, torch.Tensor], torch.Tensor]
+        ] = None,
     ) -> None:
         r"""
         Enable long video generation using FreeNoise.
@@ -195,13 +349,27 @@ class AnimateDiffFreeNoiseMixin:
             weighting_scheme (`str`, defaults to `pyramid`):
                 Weighting scheme for averaging latents after accumulation in FreeNoise blocks. The following weighting
                 schemes are supported currently:
+                    - "flat"
+                       Performs weighting averaging with a flat weight pattern: [1, 1, 1, 1, 1].
                     - "pyramid"
-                        Peforms weighted averaging with a pyramid like weight pattern: [1, 2, 3, 2, 1].
+                        Performs weighted averaging with a pyramid like weight pattern: [1, 2, 3, 2, 1].
+                    - "delayed_reverse_sawtooth"
+                        Performs weighted averaging with low weights for earlier frames and high-to-low weights for
+                        later frames: [0.01, 0.01, 3, 2, 1].
             noise_type (`str`, defaults to "shuffle_context"):
-                TODO
+                Must be one of ["shuffle_context", "repeat_context", "random"].
+                    - "shuffle_context"
+                        Shuffles a fixed batch of `context_length` latents to create a final latent of size
+                        `num_frames`. This is usually the best setting for most generation scenarious. However, there
+                        might be visible repetition noticeable in the kinds of motion/animation generated.
+                    - "repeated_context"
+                        Repeats a fixed batch of `context_length` latents to create a final latent of size
+                        `num_frames`.
+                    - "random"
+                        The final latents are random without any repetition.
         """
 
-        allowed_weighting_scheme = ["pyramid"]
+        allowed_weighting_scheme = ["flat", "pyramid", "delayed_reverse_sawtooth"]
         allowed_noise_type = ["shuffle_context", "repeat_context", "random"]
 
         if context_length > self.motion_adapter.config.motion_max_seq_length:
@@ -219,14 +387,25 @@ class AnimateDiffFreeNoiseMixin:
         self._free_noise_context_stride = context_stride
         self._free_noise_weighting_scheme = weighting_scheme
         self._free_noise_noise_type = noise_type
+        self._free_noise_prompt_interpolation_callback = prompt_interpolation_callback or self._lerp
+
+        if hasattr(self.unet.mid_block, "motion_modules"):
+            blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
+        else:
+            blocks = [*self.unet.down_blocks, *self.unet.up_blocks]
 
-        blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
         for block in blocks:
             self._enable_free_noise_in_block(block)
 
     def disable_free_noise(self) -> None:
+        r"""Disable the FreeNoise sampling mechanism."""
         self._free_noise_context_length = None
 
+        if hasattr(self.unet.mid_block, "motion_modules"):
+            blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
+        else:
+            blocks = [*self.unet.down_blocks, *self.unet.up_blocks]
+
         blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks]
         for block in blocks:
             self._disable_free_noise_in_block(block)

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py

@@ -547,7 +547,7 @@ class KandinskyV22Img2ImgCombinedPipeline(DiffusionPipeline):
         negative_image_embeds = prior_outputs[1]
 
         prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
-        image = [image] if isinstance(prompt, PIL.Image.Image) else image
+        image = [image] if isinstance(image, PIL.Image.Image) else image
 
         if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
             prompt = (image_embeds.shape[0] // len(prompt)) * prompt
@@ -813,7 +813,7 @@ class KandinskyV22InpaintCombinedPipeline(DiffusionPipeline):
         negative_image_embeds = prior_outputs[1]
 
         prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
-        image = [image] if isinstance(prompt, PIL.Image.Image) else image
+        image = [image] if isinstance(image, PIL.Image.Image) else image
         mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image
 
         if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:

src/diffusers/pipelines/pag/__init__.py

@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_pag_controlnet_sd"] = ["StableDiffusionControlNetPAGPipeline"]
     _import_structure["pipeline_pag_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPAGPipeline"]
+    _import_structure["pipeline_pag_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetPAGImg2ImgPipeline"]
     _import_structure["pipeline_pag_hunyuandit"] = ["HunyuanDiTPAGPipeline"]
     _import_structure["pipeline_pag_kolors"] = ["KolorsPAGPipeline"]
     _import_structure["pipeline_pag_pixart_sigma"] = ["PixArtSigmaPAGPipeline"]
@@ -44,6 +45,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipeline_pag_controlnet_sd import StableDiffusionControlNetPAGPipeline
         from .pipeline_pag_controlnet_sd_xl import StableDiffusionXLControlNetPAGPipeline
+        from .pipeline_pag_controlnet_sd_xl_img2img import StableDiffusionXLControlNetPAGImg2ImgPipeline
         from .pipeline_pag_hunyuandit import HunyuanDiTPAGPipeline
         from .pipeline_pag_kolors import KolorsPAGPipeline
         from .pipeline_pag_pixart_sigma import PixArtSigmaPAGPipeline

src/diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py

@@ -734,6 +734,8 @@ class AnimateDiffPAGPipeline(
         elif self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+        prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
+
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
                 ip_adapter_image,
@@ -805,7 +807,9 @@ class AnimateDiffPAGPipeline(
             with self.progress_bar(total=self._num_timesteps) as progress_bar:
                 for i, t in enumerate(timesteps):
                     # expand the latents if we are doing classifier free guidance
-                    latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0]))
+                    latent_model_input = torch.cat(
+                        [latents] * (prompt_embeds.shape[0] // num_frames // latents.shape[0])
+                    )
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
                     # predict the noise residual

src/diffusers/pipelines/pia/pipeline_pia.py

@@ -824,6 +824,8 @@ class PIAPipeline(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+        prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0)
+
         if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
             image_embeds = self.prepare_ip_adapter_image_embeds(
                 ip_adapter_image,

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -22,7 +22,7 @@ from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
 import torch
-from huggingface_hub import model_info
+from huggingface_hub import ModelCard, model_info
 from huggingface_hub.utils import validate_hf_hub_args
 from packaging import version
 
@@ -33,6 +33,7 @@ from ..utils import (
     ONNX_WEIGHTS_NAME,
     SAFETENSORS_WEIGHTS_NAME,
     WEIGHTS_NAME,
+    deprecate,
     get_class_from_dynamic_module,
     is_accelerate_available,
     is_peft_available,
@@ -89,7 +90,7 @@ for library in LOADABLE_CLASSES:
     ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
 
 
-def is_safetensors_compatible(filenames, passed_components=None) -> bool:
+def is_safetensors_compatible(filenames, passed_components=None, folder_names=None) -> bool:
     """
     Checking for safetensors compatibility:
     - The model is safetensors compatible only if there is a safetensors file for each model component present in
@@ -101,6 +102,8 @@ def is_safetensors_compatible(filenames, passed_components=None) -> bool:
       extension is replaced with ".safetensors"
     """
     passed_components = passed_components or []
+    if folder_names is not None:
+        filenames = {f for f in filenames if os.path.split(f)[0] in folder_names}
 
     # extract all components of the pipeline and their associated files
     components = {}
@@ -744,3 +747,92 @@ def _fetch_class_library_tuple(module):
     class_name = not_compiled_module.__class__.__name__
 
     return (library, class_name)
+
+
+def _identify_model_variants(folder: str, variant: str, config: dict) -> dict:
+    model_variants = {}
+    if variant is not None:
+        for sub_folder in os.listdir(folder):
+            folder_path = os.path.join(folder, sub_folder)
+            is_folder = os.path.isdir(folder_path) and sub_folder in config
+            variant_exists = is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
+            if variant_exists:
+                model_variants[sub_folder] = variant
+    return model_variants
+
+
+def _resolve_custom_pipeline_and_cls(folder, config, custom_pipeline):
+    custom_class_name = None
+    if os.path.isfile(os.path.join(folder, f"{custom_pipeline}.py")):
+        custom_pipeline = os.path.join(folder, f"{custom_pipeline}.py")
+    elif isinstance(config["_class_name"], (list, tuple)) and os.path.isfile(
+        os.path.join(folder, f"{config['_class_name'][0]}.py")
+    ):
+        custom_pipeline = os.path.join(folder, f"{config['_class_name'][0]}.py")
+        custom_class_name = config["_class_name"][1]
+
+    return custom_pipeline, custom_class_name
+
+
+def _maybe_raise_warning_for_inpainting(pipeline_class, pretrained_model_name_or_path: str, config: dict):
+    if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
+        version.parse(config["_diffusers_version"]).base_version
+    ) <= version.parse("0.5.1"):
+        from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
+
+        pipeline_class = StableDiffusionInpaintPipelineLegacy
+
+        deprecation_message = (
+            "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
+            f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
+            " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
+            " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
+            f" checkpoint {pretrained_model_name_or_path} to the format of"
+            " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
+            " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
+        )
+        deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
+
+
+def _update_init_kwargs_with_connected_pipeline(
+    init_kwargs: dict, passed_pipe_kwargs: dict, passed_class_objs: dict, folder: str, **pipeline_loading_kwargs
+) -> dict:
+    from .pipeline_utils import DiffusionPipeline
+
+    modelcard = ModelCard.load(os.path.join(folder, "README.md"))
+    connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
+
+    # We don't scheduler argument to match the existing logic:
+    # https://github.com/huggingface/diffusers/blob/867e0c919e1aa7ef8b03c8eb1460f4f875a683ae/src/diffusers/pipelines/pipeline_utils.py#L906C13-L925C14
+    pipeline_loading_kwargs_cp = pipeline_loading_kwargs.copy()
+    if pipeline_loading_kwargs_cp is not None and len(pipeline_loading_kwargs_cp) >= 1:
+        for k in pipeline_loading_kwargs:
+            if "scheduler" in k:
+                _ = pipeline_loading_kwargs_cp.pop(k)
+
+    def get_connected_passed_kwargs(prefix):
+        connected_passed_class_obj = {
+            k.replace(f"{prefix}_", ""): w for k, w in passed_class_objs.items() if k.split("_")[0] == prefix
+        }
+        connected_passed_pipe_kwargs = {
+            k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix
+        }
+
+        connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs}
+        return connected_passed_kwargs
+
+    connected_pipes = {
+        prefix: DiffusionPipeline.from_pretrained(
+            repo_id, **pipeline_loading_kwargs_cp, **get_connected_passed_kwargs(prefix)
+        )
+        for prefix, repo_id in connected_pipes.items()
+        if repo_id is not None
+    }
+
+    for prefix, connected_pipe in connected_pipes.items():
+        # add connected pipes to `init_kwargs` with <prefix>_<component_name>, e.g. "prior_text_encoder"
+        init_kwargs.update(
+            {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
+        )
+
+    return init_kwargs

src/diffusers/pipelines/pipeline_utils.py

@@ -75,7 +75,11 @@ from .pipeline_loading_utils import (
     _get_custom_pipeline_class,
     _get_final_device_map,
     _get_pipeline_class,
+    _identify_model_variants,
+    _maybe_raise_warning_for_inpainting,
+    _resolve_custom_pipeline_and_cls,
     _unwrap_model,
+    _update_init_kwargs_with_connected_pipeline,
     is_safetensors_compatible,
     load_sub_model,
     maybe_raise_or_warn,
@@ -622,6 +626,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         >>> pipeline.scheduler = scheduler
         ```
         """
+        # Copy the kwargs to re-use during loading connected pipeline.
+        kwargs_copied = kwargs.copy()
+
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
@@ -722,33 +729,19 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         config_dict.pop("_ignore_files", None)
 
         # 2. Define which model components should load variants
-        # We retrieve the information by matching whether variant
-        # model checkpoints exist in the subfolders
-        model_variants = {}
-        if variant is not None:
-            for folder in os.listdir(cached_folder):
-                folder_path = os.path.join(cached_folder, folder)
-                is_folder = os.path.isdir(folder_path) and folder in config_dict
-                variant_exists = is_folder and any(
-                    p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)
-                )
-                if variant_exists:
-                    model_variants[folder] = variant
+        # We retrieve the information by matching whether variant model checkpoints exist in the subfolders.
+        # Example: `diffusion_pytorch_model.safetensors` -> `diffusion_pytorch_model.fp16.safetensors`
+        # with variant being `"fp16"`.
+        model_variants = _identify_model_variants(folder=cached_folder, variant=variant, config=config_dict)
 
         # 3. Load the pipeline class, if using custom module then load it from the hub
         # if we load from explicit class, let's use it
-        custom_class_name = None
-        if os.path.isfile(os.path.join(cached_folder, f"{custom_pipeline}.py")):
-            custom_pipeline = os.path.join(cached_folder, f"{custom_pipeline}.py")
-        elif isinstance(config_dict["_class_name"], (list, tuple)) and os.path.isfile(
-            os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
-        ):
-            custom_pipeline = os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
-            custom_class_name = config_dict["_class_name"][1]
-
+        custom_pipeline, custom_class_name = _resolve_custom_pipeline_and_cls(
+            folder=cached_folder, config=config_dict, custom_pipeline=custom_pipeline
+        )
         pipeline_class = _get_pipeline_class(
             cls,
-            config_dict,
+            config=config_dict,
             load_connected_pipeline=load_connected_pipeline,
             custom_pipeline=custom_pipeline,
             class_name=custom_class_name,
@@ -760,23 +753,13 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise NotImplementedError("`device_map` is not yet supported for connected pipelines.")
 
         # DEPRECATED: To be removed in 1.0.0
-        if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
-            version.parse(config_dict["_diffusers_version"]).base_version
-        ) <= version.parse("0.5.1"):
-            from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
-
-            pipeline_class = StableDiffusionInpaintPipelineLegacy
-
-            deprecation_message = (
-                "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
-                f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
-                " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
-                " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
-                f" checkpoint {pretrained_model_name_or_path} to the format of"
-                " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
-                " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
-            )
-            deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
+        # we are deprecating the `StableDiffusionInpaintPipelineLegacy` pipeline which gets loaded
+        # when a user requests for a `StableDiffusionInpaintPipeline` with `diffusers` version being <= 0.5.1.
+        _maybe_raise_warning_for_inpainting(
+            pipeline_class=pipeline_class,
+            pretrained_model_name_or_path=pretrained_model_name_or_path,
+            config=config_dict,
+        )
 
         # 4. Define expected modules given pipeline signature
         # and define non-None initialized modules (=`init_kwargs`)
@@ -787,7 +770,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
         passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
         passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
-
         init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
 
         # define init kwargs and make sure that optional component modules are filtered out
@@ -847,6 +829,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         # 7. Load each module in the pipeline
         current_device_map = None
         for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):
+            # 7.1 device_map shenanigans
             if final_device_map is not None and len(final_device_map) > 0:
                 component_device = final_device_map.get(name, None)
                 if component_device is not None:
@@ -854,15 +837,15 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 else:
                     current_device_map = None
 
-            # 7.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
+            # 7.2 - now that JAX/Flax is an official framework of the library, we might load from Flax names
             class_name = class_name[4:] if class_name.startswith("Flax") else class_name
 
-            # 7.2 Define all importable classes
+            # 7.3 Define all importable classes
             is_pipeline_module = hasattr(pipelines, library_name)
             importable_classes = ALL_IMPORTABLE_CLASSES
             loaded_sub_model = None
 
-            # 7.3 Use passed sub model or load class_name from library_name
+            # 7.4 Use passed sub model or load class_name from library_name
             if name in passed_class_obj:
                 # if the model is in a pipeline module, then we load it from the pipeline
                 # check that passed_class_obj has correct parent class
@@ -900,56 +883,17 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
 
             init_kwargs[name] = loaded_sub_model  # UNet(...), # DiffusionSchedule(...)
 
+        # 8. Handle connected pipelines.
         if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")):
-            modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
-            connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
-            load_kwargs = {
-                "cache_dir": cache_dir,
-                "force_download": force_download,
-                "proxies": proxies,
-                "local_files_only": local_files_only,
-                "token": token,
-                "revision": revision,
-                "torch_dtype": torch_dtype,
-                "custom_pipeline": custom_pipeline,
-                "custom_revision": custom_revision,
-                "provider": provider,
-                "sess_options": sess_options,
-                "device_map": device_map,
-                "max_memory": max_memory,
-                "offload_folder": offload_folder,
-                "offload_state_dict": offload_state_dict,
-                "low_cpu_mem_usage": low_cpu_mem_usage,
-                "variant": variant,
-                "use_safetensors": use_safetensors,
-            }
+            init_kwargs = _update_init_kwargs_with_connected_pipeline(
+                init_kwargs=init_kwargs,
+                passed_pipe_kwargs=passed_pipe_kwargs,
+                passed_class_objs=passed_class_obj,
+                folder=cached_folder,
+                **kwargs_copied,
+            )
 
-            def get_connected_passed_kwargs(prefix):
-                connected_passed_class_obj = {
-                    k.replace(f"{prefix}_", ""): w for k, w in passed_class_obj.items() if k.split("_")[0] == prefix
-                }
-                connected_passed_pipe_kwargs = {
-                    k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix
-                }
-
-                connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs}
-                return connected_passed_kwargs
-
-            connected_pipes = {
-                prefix: DiffusionPipeline.from_pretrained(
-                    repo_id, **load_kwargs.copy(), **get_connected_passed_kwargs(prefix)
-                )
-                for prefix, repo_id in connected_pipes.items()
-                if repo_id is not None
-            }
-
-            for prefix, connected_pipe in connected_pipes.items():
-                # add connected pipes to `init_kwargs` with <prefix>_<component_name>, e.g. "prior_text_encoder"
-                init_kwargs.update(
-                    {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
-                )
-
-        # 8. Potentially add passed objects if expected
+        # 9. Potentially add passed objects if expected
         missing_modules = set(expected_modules) - set(init_kwargs.keys())
         passed_modules = list(passed_class_obj.keys())
         optional_modules = pipeline_class._optional_components
@@ -1416,14 +1360,18 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if (
                 use_safetensors
                 and not allow_pickle
-                and not is_safetensors_compatible(model_filenames, passed_components=passed_components)
+                and not is_safetensors_compatible(
+                    model_filenames, passed_components=passed_components, folder_names=model_folder_names
+                )
             ):
                 raise EnvironmentError(
                     f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})"
                 )
             if from_flax:
                 ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
-            elif use_safetensors and is_safetensors_compatible(model_filenames, passed_components=passed_components):
+            elif use_safetensors and is_safetensors_compatible(
+                model_filenames, passed_components=passed_components, folder_names=model_folder_names
+            ):
                 ignore_patterns = ["*.bin", "*.msgpack"]
 
                 use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx

src/diffusers/training_utils.py

@@ -418,11 +418,11 @@ class EMAModel:
         one_minus_decay = 1 - decay
 
         context_manager = contextlib.nullcontext
-        if is_transformers_available() and transformers.deepspeed.is_deepspeed_zero3_enabled():
+        if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
             import deepspeed
 
         if self.foreach:
-            if is_transformers_available() and transformers.deepspeed.is_deepspeed_zero3_enabled():
+            if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
                 context_manager = deepspeed.zero.GatheredParameters(parameters, modifier_rank=None)
 
             with context_manager():
@@ -444,7 +444,7 @@ class EMAModel:
 
         else:
             for s_param, param in zip(self.shadow_params, parameters):
-                if is_transformers_available() and transformers.deepspeed.is_deepspeed_zero3_enabled():
+                if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
                     context_manager = deepspeed.zero.GatheredParameters(param, modifier_rank=None)
 
                 with context_manager():

src/diffusers/utils/dummy_pt_objects.py

@@ -197,6 +197,21 @@ class FluxControlNetModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class FluxMultiControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class FluxTransformer2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1637,6 +1637,21 @@ class StableDiffusionXLControlNetInpaintPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusionXLControlNetPAGImg2ImgPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableDiffusionXLControlNetPAGPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/lora/test_lora_layers_flux.py

@@ -12,19 +12,26 @@
 # 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 os
 import sys
+import tempfile
 import unittest
 
+import numpy as np
+import safetensors.torch
 import torch
 from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import FlowMatchEulerDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
-from diffusers.utils.testing_utils import floats_tensor, require_peft_backend
+from diffusers.utils.testing_utils import floats_tensor, is_peft_available, require_peft_backend, torch_device
 
 
+if is_peft_available():
+    from peft.utils import get_peft_model_state_dict
+
 sys.path.append(".")
 
-from utils import PeftLoraLoaderMixinTests  # noqa: E402
+from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set  # noqa: E402
 
 
 @require_peft_backend
@@ -90,3 +97,51 @@ class FluxLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
             pipeline_inputs.update({"generator": generator})
 
         return noise, input_ids, pipeline_inputs
+
+    def test_with_alpha_in_state_dict(self):
+        components, _, denoiser_lora_config = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+        output_no_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertTrue(output_no_lora.shape == self.output_shape)
+
+        pipe.transformer.add_adapter(denoiser_lora_config)
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
+
+        images_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
+            self.pipeline_class.save_lora_weights(tmpdirname, transformer_lora_layers=denoiser_state_dict)
+
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
+            pipe.unload_lora_weights()
+            pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
+
+            # modify the state dict to have alpha values following
+            # https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA/blob/main/jon_snow.safetensors
+            state_dict_with_alpha = safetensors.torch.load_file(
+                os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")
+            )
+            alpha_dict = {}
+            for k, v in state_dict_with_alpha.items():
+                # only do for `transformer` and for the k projections -- should be enough to test.
+                if "transformer" in k and "to_k" in k and "lora_A" in k:
+                    alpha_dict[f"{k}.alpha"] = float(torch.randint(10, 100, size=()))
+            state_dict_with_alpha.update(alpha_dict)
+
+        images_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0)).images
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+
+        pipe.unload_lora_weights()
+        pipe.load_lora_weights(state_dict_with_alpha)
+        images_lora_with_alpha = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+        self.assertTrue(
+            np.allclose(images_lora, images_lora_from_pretrained, atol=1e-3, rtol=1e-3),
+            "Loading from saved checkpoints should give same results.",
+        )
+        self.assertFalse(np.allclose(images_lora_with_alpha, images_lora, atol=1e-3, rtol=1e-3))

tests/models/test_modeling_common.py

@@ -417,6 +417,9 @@ class ModelTesterMixin:
 
     @require_torch_gpu
     def test_set_attn_processor_for_determinism(self):
+        if self.uses_custom_attn_processor:
+            return
+
         torch.use_deterministic_algorithms(False)
         if self.forward_requires_fresh_args:
             model = self.model_class(**self.init_dict)
@@ -976,7 +979,6 @@ class ModelTesterMixin:
             self.assertTrue(actual_num_shards == expected_num_shards)
 
             new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto")
-            new_model = new_model.to(torch_device)
 
             torch.manual_seed(0)
             if "generator" in inputs_dict:

tests/models/transformers/test_models_transformer_flux.py

@@ -32,6 +32,9 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
     # We override the items here because the transformer under consideration is small.
     model_split_percents = [0.7, 0.6, 0.6]
 
+    # Skip setting testing with default: AttnProcessor
+    uses_custom_attn_processor = True
+
     @property
     def dummy_input(self):
         batch_size = 1
@@ -44,8 +47,8 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
         hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
         encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
         pooled_prompt_embeds = torch.randn((batch_size, embedding_dim)).to(torch_device)
-        text_ids = torch.randn((batch_size, sequence_length, num_image_channels)).to(torch_device)
-        image_ids = torch.randn((batch_size, height * width, num_image_channels)).to(torch_device)
+        text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
+        image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
         timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
 
         return {
@@ -80,3 +83,31 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
 
         inputs_dict = self.dummy_input
         return init_dict, inputs_dict
+
+    def test_deprecated_inputs_img_txt_ids_3d(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            output_1 = model(**inputs_dict).to_tuple()[0]
+
+        # update inputs_dict with txt_ids and img_ids as 3d tensors (deprecated)
+        text_ids_3d = inputs_dict["txt_ids"].unsqueeze(0)
+        image_ids_3d = inputs_dict["img_ids"].unsqueeze(0)
+
+        assert text_ids_3d.ndim == 3, "text_ids_3d should be a 3d tensor"
+        assert image_ids_3d.ndim == 3, "img_ids_3d should be a 3d tensor"
+
+        inputs_dict["txt_ids"] = text_ids_3d
+        inputs_dict["img_ids"] = image_ids_3d
+
+        with torch.no_grad():
+            output_2 = model(**inputs_dict).to_tuple()[0]
+
+        self.assertEqual(output_1.shape, output_2.shape)
+        self.assertTrue(
+            torch.allclose(output_1, output_2, atol=1e-5),
+            msg="output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) are not equal as them as 2d inputs",
+        )

tests/models/unets/test_models_unet_motion.py

@@ -51,7 +51,7 @@ class UNetMotionModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase)
 
         noise = floats_tensor((batch_size, num_channels, num_frames) + sizes).to(torch_device)
         time_step = torch.tensor([10]).to(torch_device)
-        encoder_hidden_states = floats_tensor((batch_size, 4, 16)).to(torch_device)
+        encoder_hidden_states = floats_tensor((batch_size * num_frames, 4, 16)).to(torch_device)
 
         return {"sample": noise, "timestep": time_step, "encoder_hidden_states": encoder_hidden_states}
 

tests/pipelines/animatediff/test_animatediff.py

@@ -460,6 +460,29 @@ class AnimateDiffPipelineFastTests(
                     "Disabling of FreeNoise should lead to results similar to the default pipeline results",
                 )
 
+    def test_free_noise_multi_prompt(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        context_length = 8
+        context_stride = 4
+        pipe.enable_free_noise(context_length, context_stride)
+
+        # Make sure that pipeline works when prompt indices are within num_frames bounds
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf"}
+        inputs["num_frames"] = 16
+        pipe(**inputs).frames[0]
+
+        with self.assertRaises(ValueError):
+            # Ensure that prompt indices are within bounds
+            inputs = self.get_dummy_inputs(torch_device)
+            inputs["num_frames"] = 16
+            inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf", 42: "Error on a leaf"}
+            pipe(**inputs).frames[0]
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",

tests/pipelines/animatediff/test_animatediff_controlnet.py

@@ -20,6 +20,7 @@ from diffusers import (
 )
 from diffusers.models.attention import FreeNoiseTransformerBlock
 from diffusers.utils import logging
+from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
@@ -329,6 +330,13 @@ class AnimateDiffControlNetPipelineFastTests(
         inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
         pipe(**inputs)
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
+
     def test_free_init(self):
         components = self.get_dummy_components()
         pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
@@ -468,6 +476,27 @@ class AnimateDiffControlNetPipelineFastTests(
                     "Disabling of FreeNoise should lead to results similar to the default pipeline results",
                 )
 
+    def test_free_noise_multi_prompt(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        context_length = 8
+        context_stride = 4
+        pipe.enable_free_noise(context_length, context_stride)
+
+        # Make sure that pipeline works when prompt indices are within num_frames bounds
+        inputs = self.get_dummy_inputs(torch_device, num_frames=16)
+        inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf"}
+        pipe(**inputs).frames[0]
+
+        with self.assertRaises(ValueError):
+            # Ensure that prompt indices are within bounds
+            inputs = self.get_dummy_inputs(torch_device, num_frames=16)
+            inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf", 42: "Error on a leaf"}
+            pipe(**inputs).frames[0]
+
     def test_vae_slicing(self, video_count=2):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/animatediff/test_animatediff_sparsectrl.py

@@ -19,6 +19,7 @@ from diffusers import (
     UNetMotionModel,
 )
 from diffusers.utils import logging
+from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import torch_device
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
@@ -393,6 +394,13 @@ class AnimateDiffSparseControlNetPipelineFastTests(
         inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
         pipe(**inputs)
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
+
     def test_free_init(self):
         components = self.get_dummy_components()
         pipe: AnimateDiffSparseControlNetPipeline = self.pipeline_class(**components)

tests/pipelines/animatediff/test_animatediff_video2video.py

@@ -491,3 +491,28 @@ class AnimateDiffVideoToVideoPipelineFastTests(
                     1e-4,
                     "Disabling of FreeNoise should lead to results similar to the default pipeline results",
                 )
+
+    def test_free_noise_multi_prompt(self):
+        components = self.get_dummy_components()
+        pipe: AnimateDiffVideoToVideoPipeline = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.to(torch_device)
+
+        context_length = 8
+        context_stride = 4
+        pipe.enable_free_noise(context_length, context_stride)
+
+        # Make sure that pipeline works when prompt indices are within num_frames bounds
+        inputs = self.get_dummy_inputs(torch_device, num_frames=16)
+        inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf"}
+        inputs["num_inference_steps"] = 2
+        inputs["strength"] = 0.5
+        pipe(**inputs).frames[0]
+
+        with self.assertRaises(ValueError):
+            # Ensure that prompt indices are within bounds
+            inputs = self.get_dummy_inputs(torch_device, num_frames=16)
+            inputs["num_inference_steps"] = 2
+            inputs["strength"] = 0.5
+            inputs["prompt"] = {0: "Caterpillar on a leaf", 10: "Butterfly on a leaf", 42: "Error on a leaf"}
+            pipe(**inputs).frames[0]

tests/pipelines/cogvideox/test_cogvideox.py

@@ -30,7 +30,12 @@ from diffusers.utils.testing_utils import (
 )
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import PipelineTesterMixin, to_np
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+    to_np,
+)
 
 
 enable_full_determinism()
@@ -275,6 +280,48 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             "VAE tiling should not affect the inference results",
         )
 
+    @unittest.skip("xformers attention processor does not exist for CogVideoX")
+    def test_xformers_attention_forwardGenerator_pass(self):
+        pass
+
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames  # [B, F, C, H, W]
+        original_image_slice = frames[0, -2:, -1, -3:, -3:]
+
+        pipe.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames
+        image_slice_fused = frames[0, -2:, -1, -3:, -3:]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames
+        image_slice_disabled = frames[0, -2:, -1, -3:, -3:]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
 
 @slow
 @require_torch_gpu

tests/pipelines/flux/test_pipeline_flux.py

@@ -13,15 +13,21 @@ from diffusers.utils.testing_utils import (
     torch_device,
 )
 
-from ..test_pipelines_common import PipelineTesterMixin
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+)
 
 
-@unittest.skipIf(torch_device == "mps", "Flux has a float64 operation which is not supported in MPS.")
 class FluxPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
     pipeline_class = FluxPipeline
     params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"])
     batch_params = frozenset(["prompt"])
 
+    # there is no xformers processor for Flux
+    test_xformers_attention = False
+
     def get_dummy_components(self):
         torch.manual_seed(0)
         transformer = FluxTransformer2DModel(
@@ -143,6 +149,46 @@ class FluxPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         max_diff = np.abs(output_with_prompt - output_with_embeds).max()
         assert max_diff < 1e-4
 
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        original_image_slice = image[0, -3:, -3:, -1]
+
+        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
+        # to the pipeline level.
+        pipe.transformer.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_fused = image[0, -3:, -3:, -1]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice_disabled = image[0, -3:, -3:, -1]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
 
 @slow
 @require_torch_gpu

tests/pipelines/latte/test_latte.py

@@ -28,6 +28,7 @@ from diffusers import (
     LattePipeline,
     LatteTransformer3DModel,
 )
+from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
     numpy_cosine_similarity_distance,
@@ -256,6 +257,13 @@ class LattePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
         self.assertLess(max_diff, 1.0)
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
+
 
 @slow
 @require_torch_gpu

tests/pipelines/pag/test_pag_controlnet_sdxl_img2img.py

@@ -0,0 +1,271 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import random
+import unittest
+
+import numpy as np
+import torch
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from diffusers import (
+    AutoencoderKL,
+    ControlNetModel,
+    EulerDiscreteScheduler,
+    StableDiffusionXLControlNetImg2ImgPipeline,
+    StableDiffusionXLControlNetPAGImg2ImgPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor
+
+from ..pipeline_params import (
+    IMAGE_TO_IMAGE_IMAGE_PARAMS,
+    TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
+    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
+)
+from ..test_pipelines_common import (
+    IPAdapterTesterMixin,
+    PipelineFromPipeTesterMixin,
+    PipelineLatentTesterMixin,
+    PipelineTesterMixin,
+    SDXLOptionalComponentsTesterMixin,
+)
+
+
+enable_full_determinism()
+
+
+class StableDiffusionXLControlNetPAGImg2ImgPipelineFastTests(
+    IPAdapterTesterMixin,
+    PipelineLatentTesterMixin,
+    PipelineTesterMixin,
+    PipelineFromPipeTesterMixin,
+    SDXLOptionalComponentsTesterMixin,
+    unittest.TestCase,
+):
+    pipeline_class = StableDiffusionXLControlNetPAGImg2ImgPipeline
+    params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS.union({"pag_scale", "pag_adaptive_scale"})
+    batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS
+    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
+    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union(
+        {"add_text_embeds", "add_time_ids", "add_neg_time_ids"}
+    )
+
+    # Copied from tests.pipelines.controlnet.test_controlnet_sdxl_img2img.ControlNetPipelineSDXLImg2ImgFastTests.get_dummy_components
+    def get_dummy_components(self, skip_first_text_encoder=False):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            # SD2-specific config below
+            attention_head_dim=(2, 4),
+            use_linear_projection=True,
+            addition_embed_type="text_time",
+            addition_time_embed_dim=8,
+            transformer_layers_per_block=(1, 2),
+            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
+            cross_attention_dim=64 if not skip_first_text_encoder else 32,
+        )
+        torch.manual_seed(0)
+        controlnet = ControlNetModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            in_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            conditioning_embedding_out_channels=(16, 32),
+            # SD2-specific config below
+            attention_head_dim=(2, 4),
+            use_linear_projection=True,
+            addition_embed_type="text_time",
+            addition_time_embed_dim=8,
+            transformer_layers_per_block=(1, 2),
+            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
+            cross_attention_dim=64,
+        )
+        torch.manual_seed(0)
+        scheduler = EulerDiscreteScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            steps_offset=1,
+            beta_schedule="scaled_linear",
+            timestep_spacing="leading",
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            # SD2-specific config below
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
+        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        components = {
+            "unet": unet,
+            "controlnet": controlnet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder if not skip_first_text_encoder else None,
+            "tokenizer": tokenizer if not skip_first_text_encoder else None,
+            "text_encoder_2": text_encoder_2,
+            "tokenizer_2": tokenizer_2,
+            "image_encoder": None,
+            "feature_extractor": None,
+        }
+        return components
+
+    # based on tests.pipelines.controlnet.test_controlnet_sdxl_img2img.ControlNetPipelineSDXLImg2ImgFastTests.get_dummy_inputs
+    # add `pag_scale` to the inputs
+    def get_dummy_inputs(self, device, seed=0):
+        controlnet_embedder_scale_factor = 2
+        image = floats_tensor(
+            (1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
+            rng=random.Random(seed),
+        ).to(device)
+
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "pag_scale": 3.0,
+            "output_type": "np",
+            "image": image,
+            "control_image": image,
+        }
+
+        return inputs
+
+    def test_pag_disable_enable(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline
+        pipe_sd = StableDiffusionXLControlNetImg2ImgPipeline(**components)
+        pipe_sd = pipe_sd.to(device)
+        pipe_sd.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        del inputs["pag_scale"]
+        assert (
+            "pag_scale" not in inspect.signature(pipe_sd.__call__).parameters
+        ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}."
+        out = pipe_sd(**inputs).images[0, -3:, -3:, -1]
+
+        # pag disabled with pag_scale=0.0
+        pipe_pag = self.pipeline_class(**components)
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["pag_scale"] = 0.0
+        out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        # pag enable
+        pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        out_pag_enabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
+        assert np.abs(out.flatten() - out_pag_enabled.flatten()).max() > 1e-3
+
+    def test_save_load_optional_components(self):
+        pass
+
+    def test_pag_cfg(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe_pag(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (
+            1,
+            64,
+            64,
+            3,
+        ), f"the shape of the output image should be (1, 64, 64, 3) but got {image.shape}"
+        expected_slice = np.array(
+            [0.5562928, 0.44882968, 0.4588066, 0.63200223, 0.5694165, 0.4955688, 0.6126959, 0.57588536, 0.43827885]
+        )
+
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        assert max_diff < 1e-3, f"output is different from expected, {image_slice.flatten()}"
+
+    def test_pag_uncond(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["guidance_scale"] = 0.0
+        image = pipe_pag(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (
+            1,
+            64,
+            64,
+            3,
+        ), f"the shape of the output image should be (1, 64, 64, 3) but got {image.shape}"
+        expected_slice = np.array(
+            [0.5543988, 0.45614323, 0.4665692, 0.6202247, 0.5598917, 0.49621183, 0.6084159, 0.5722314, 0.43945464]
+        )
+
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        assert max_diff < 1e-3, f"output is different from expected, {image_slice.flatten()}"

tests/pipelines/pag/test_pag_sd3.py

@@ -37,6 +37,7 @@ class StableDiffusion3PAGPipelineFastTests(unittest.TestCase, PipelineTesterMixi
         ]
     )
     batch_params = frozenset(["prompt", "negative_prompt"])
+    test_xformers_attention = False
 
     def get_dummy_components(self):
         torch.manual_seed(0)

tests/pipelines/stable_audio/test_stable_audio.py

@@ -68,6 +68,8 @@ class StableAudioPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             "callback_steps",
         ]
     )
+    # There is not xformers version of the StableAudioPipeline custom attention processor
+    test_xformers_attention = False
 
     def get_dummy_components(self):
         torch.manual_seed(0)

tests/pipelines/test_pipeline_utils.py

@@ -116,6 +116,30 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
         ]
         self.assertFalse(is_safetensors_compatible(filenames))
 
+    def test_transformer_model_is_compatible_variant_extra_folder(self):
+        filenames = [
+            "safety_checker/pytorch_model.fp16.bin",
+            "safety_checker/model.fp16.safetensors",
+            "vae/diffusion_pytorch_model.fp16.bin",
+            "vae/diffusion_pytorch_model.fp16.safetensors",
+            "text_encoder/pytorch_model.fp16.bin",
+            "unet/diffusion_pytorch_model.fp16.bin",
+            "unet/diffusion_pytorch_model.fp16.safetensors",
+        ]
+        self.assertTrue(is_safetensors_compatible(filenames, folder_names={"vae", "unet"}))
+
+    def test_transformer_model_is_not_compatible_variant_extra_folder(self):
+        filenames = [
+            "safety_checker/pytorch_model.fp16.bin",
+            "safety_checker/model.fp16.safetensors",
+            "vae/diffusion_pytorch_model.fp16.bin",
+            "vae/diffusion_pytorch_model.fp16.safetensors",
+            "text_encoder/pytorch_model.fp16.bin",
+            "unet/diffusion_pytorch_model.fp16.bin",
+            "unet/diffusion_pytorch_model.fp16.safetensors",
+        ]
+        self.assertFalse(is_safetensors_compatible(filenames, folder_names={"text_encoder"}))
+
     def test_transformers_is_compatible_sharded(self):
         filenames = [
             "text_encoder/pytorch_model.bin",