update

* update

* Update docs/source/en/optimization/memory.md

* Apply suggestions from code review

Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

* apply review suggestions

* update

---------

Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

.github/workflows/benchmark.yml

@@ -38,6 +38,7 @@ jobs:
           python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           python -m uv pip install -e [quality,test]
           python -m uv pip install pandas peft
+          python -m uv pip uninstall transformers && python -m uv pip install transformers==4.48.0
       - name: Environment
         run: |
           python utils/print_env.py

.github/workflows/nightly_tests.yml

@@ -414,12 +414,16 @@ jobs:
         config:
           - backend: "bitsandbytes"
             test_location: "bnb"
+            additional_deps: ["peft"]
           - backend: "gguf"
             test_location: "gguf"
+            additional_deps: []
           - backend: "torchao"
             test_location: "torchao"
+            additional_deps: []
           - backend: "optimum_quanto"
             test_location: "quanto"
+            additional_deps: []
     runs-on:
       group: aws-g6e-xlarge-plus
     container:
@@ -437,6 +441,9 @@ jobs:
           python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
           python -m uv pip install -e [quality,test]
           python -m uv pip install -U ${{ matrix.config.backend }}
+          if [ "${{ join(matrix.config.additional_deps, ' ') }}" != "" ]; then
+              python -m uv pip install ${{ join(matrix.config.additional_deps, ' ') }}
+          fi
           python -m uv pip install pytest-reportlog
       - name: Environment
         run: |

.github/workflows/pr_tests_gpu.yml

@@ -28,7 +28,51 @@ env:
   PIPELINE_USAGE_CUTOFF: 1000000000 # set high cutoff so that only always-test pipelines run
 
 jobs:
+  check_code_quality:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check quality
+        run: make quality
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Quality check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make style && make quality'" >> $GITHUB_STEP_SUMMARY
+
+  check_repository_consistency:
+    needs: check_code_quality
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[quality]
+      - name: Check repo consistency
+        run: |
+          python utils/check_copies.py
+          python utils/check_dummies.py
+          python utils/check_support_list.py
+          make deps_table_check_updated
+      - name: Check if failure
+        if: ${{ failure() }}
+        run: |
+          echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
+  
   setup_torch_cuda_pipeline_matrix:
+    needs: [check_code_quality, check_repository_consistency]
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
     runs-on:
       group: aws-general-8-plus
@@ -133,6 +177,7 @@ jobs:
 
   torch_cuda_tests:
     name: Torch CUDA Tests
+    needs: [check_code_quality, check_repository_consistency]
     runs-on:
       group: aws-g4dn-2xlarge
     container:
@@ -201,7 +246,7 @@ jobs:
 
   run_examples_tests:
     name: Examples PyTorch CUDA tests on Ubuntu
-        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+    needs: [check_code_quality, check_repository_consistency]
     runs-on:
       group: aws-g4dn-2xlarge
 
@@ -220,6 +265,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
         python -m uv pip install -e [quality,test,training]
 
     - name: Environment

docs/source/en/_toctree.yml

@@ -81,6 +81,8 @@
     title: Overview
   - local: hybrid_inference/vae_decode
     title: VAE Decode
+  - local: hybrid_inference/vae_encode
+    title: VAE Encode
   - local: hybrid_inference/api_reference
     title: API Reference
   title: Hybrid Inference
@@ -494,6 +496,8 @@
       title: PixArt-Σ
     - local: api/pipelines/sana
       title: Sana
+    - local: api/pipelines/sana_sprint
+      title: Sana Sprint
     - local: api/pipelines/self_attention_guidance
       title: Self-Attention Guidance
     - local: api/pipelines/semantic_stable_diffusion

docs/source/en/api/cache.md

@@ -38,6 +38,33 @@ config = PyramidAttentionBroadcastConfig(
 pipe.transformer.enable_cache(config)
 ```
 
+## Faster Cache
+
+[FasterCache](https://huggingface.co/papers/2410.19355) from Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, Kwan-Yee K. Wong.
+
+FasterCache is a method that speeds up inference in diffusion transformers by:
+- Reusing attention states between successive inference steps, due to high similarity between them
+- Skipping unconditional branch prediction used in classifier-free guidance by revealing redundancies between unconditional and conditional branch outputs for the same timestep, and therefore approximating the unconditional branch output using the conditional branch output
+
+```python
+import torch
+from diffusers import CogVideoXPipeline, FasterCacheConfig
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+config = FasterCacheConfig(
+    spatial_attention_block_skip_range=2,
+    spatial_attention_timestep_skip_range=(-1, 681),
+    current_timestep_callback=lambda: pipe.current_timestep,
+    attention_weight_callback=lambda _: 0.3,
+    unconditional_batch_skip_range=5,
+    unconditional_batch_timestep_skip_range=(-1, 781),
+    tensor_format="BFCHW",
+)
+pipe.transformer.enable_cache(config)
+```
+
 ### CacheMixin
 
 [[autodoc]] CacheMixin
@@ -47,3 +74,9 @@ pipe.transformer.enable_cache(config)
 [[autodoc]] PyramidAttentionBroadcastConfig
 
 [[autodoc]] apply_pyramid_attention_broadcast
+
+### FasterCacheConfig
+
+[[autodoc]] FasterCacheConfig
+
+[[autodoc]] apply_faster_cache

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

@@ -50,7 +50,8 @@ The following models are available for the image-to-video pipeline:
 | Model name | Description |
 |:---|:---|
 | [`Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
-| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |
+| [`hunyuanvideo-community/HunyuanVideo-I2V-33ch`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 33-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20). |
+| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 16-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |
 
 ## Quantization
 

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

@@ -196,6 +196,12 @@ export_to_video(video, "ship.mp4", fps=24)
   - all
   - __call__
 
+## LTXConditionPipeline
+
+[[autodoc]] LTXConditionPipeline
+  - all
+  - __call__
+
 ## LTXPipelineOutput
 
 [[autodoc]] pipelines.ltx.pipeline_output.LTXPipelineOutput

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

@@ -58,10 +58,10 @@ Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fa
 First, load the pipeline:
 
 ```python
-from diffusers import LuminaText2ImgPipeline
+from diffusers import LuminaPipeline
 import torch
 
-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
 	"Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16
 ).to("cuda")
 ```
@@ -86,11 +86,11 @@ image = pipeline(prompt="Upper body of a young woman in a Victorian-era outfit w
 
 Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
 
-Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaText2ImgPipeline`] for inference with bitsandbytes.
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaPipeline`] for inference with bitsandbytes.
 
 ```py
 import torch
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaText2ImgPipeline
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, Transformer2DModel, LuminaPipeline
 from transformers import BitsAndBytesConfig as BitsAndBytesConfig, T5EncoderModel
 
 quant_config = BitsAndBytesConfig(load_in_8bit=True)
@@ -109,7 +109,7 @@ transformer_8bit = Transformer2DModel.from_pretrained(
     torch_dtype=torch.float16,
 )
 
-pipeline = LuminaText2ImgPipeline.from_pretrained(
+pipeline = LuminaPipeline.from_pretrained(
     "Alpha-VLLM/Lumina-Next-SFT-diffusers",
     text_encoder=text_encoder_8bit,
     transformer=transformer_8bit,
@@ -122,9 +122,9 @@ image = pipeline(prompt).images[0]
 image.save("lumina.png")
 ```
 
-## LuminaText2ImgPipeline
+## LuminaPipeline
 
-[[autodoc]] LuminaText2ImgPipeline
+[[autodoc]] LuminaPipeline
 	- all
 	- __call__
 

docs/source/en/api/pipelines/lumina2.md

@@ -36,14 +36,14 @@ Single file loading for Lumina Image 2.0 is available for the `Lumina2Transforme
 
 ```python
 import torch
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline
 
 ckpt_path = "https://huggingface.co/Alpha-VLLM/Lumina-Image-2.0/blob/main/consolidated.00-of-01.pth"
 transformer = Lumina2Transformer2DModel.from_single_file(
     ckpt_path, torch_dtype=torch.bfloat16
 )
 
-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
     "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -60,7 +60,7 @@ image.save("lumina-single-file.png")
 GGUF Quantized checkpoints for the `Lumina2Transformer2DModel` can be loaded via `from_single_file` with the `GGUFQuantizationConfig` 
 
 ```python
-from diffusers import Lumina2Transformer2DModel, Lumina2Text2ImgPipeline, GGUFQuantizationConfig 
+from diffusers import Lumina2Transformer2DModel, Lumina2Pipeline, GGUFQuantizationConfig 
 
 ckpt_path = "https://huggingface.co/calcuis/lumina-gguf/blob/main/lumina2-q4_0.gguf"
 transformer = Lumina2Transformer2DModel.from_single_file(
@@ -69,7 +69,7 @@ transformer = Lumina2Transformer2DModel.from_single_file(
     torch_dtype=torch.bfloat16,
 )
 
-pipe = Lumina2Text2ImgPipeline.from_pretrained(
+pipe = Lumina2Pipeline.from_pretrained(
     "Alpha-VLLM/Lumina-Image-2.0", transformer=transformer, torch_dtype=torch.bfloat16
 )
 pipe.enable_model_cpu_offload()
@@ -80,8 +80,8 @@ image = pipe(
 image.save("lumina-gguf.png")
 ```
 
-## Lumina2Text2ImgPipeline
+## Lumina2Pipeline
 
-[[autodoc]] Lumina2Text2ImgPipeline
+[[autodoc]] Lumina2Pipeline
   - all
   - __call__

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

@@ -0,0 +1,100 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# SanaSprintPipeline
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+[SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation](https://huggingface.co/papers/2503.09641) from NVIDIA, MIT HAN Lab, and Hugging Face by Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Enze Xie, Song Han
+
+The abstract from the paper is:
+
+*This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.*
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj), [shuchen Xue](https://github.com/scxue) and [Enze Xie](https://github.com/xieenze). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model/).
+
+Available models:
+
+|                                                                    Model                                                                    | Recommended dtype |
+|:-------------------------------------------------------------------------------------------------------------------------------------------:|:-----------------:|
+| [`Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers) | `torch.bfloat16`  |
+| [`Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers) | `torch.bfloat16`  |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-sprint-67d6810d65235085b3b17c76) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder must stay in `torch.bfloat16` and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaSprintPipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaTransformer2DModel, SanaSprintPipeline
+from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+text_encoder_8bit = AutoModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="text_encoder",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = SanaTransformer2DModel.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.bfloat16,
+)
+
+pipeline = SanaSprintPipeline.from_pretrained(
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
+    text_encoder=text_encoder_8bit,
+    transformer=transformer_8bit,
+    torch_dtype=torch.bfloat16,
+    device_map="balanced",
+)
+
+prompt = "a tiny astronaut hatching from an egg on the moon"
+image = pipeline(prompt).images[0]
+image.save("sana.png")
+```
+
+## Setting `max_timesteps`
+
+Users can tweak the `max_timesteps` value for experimenting with the visual quality of the generated outputs. The default `max_timesteps` value was obtained with an inference-time search process. For more details about it, check out the paper.
+
+## SanaSprintPipeline
+
+[[autodoc]] SanaSprintPipeline
+  - all
+  - __call__
+
+
+## SanaPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput

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

@@ -14,6 +14,10 @@
 
 # Wan
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
 
 <!-- TODO(aryan): update abstract once paper is out -->

docs/source/en/hybrid_inference/api_reference.md

@@ -3,3 +3,7 @@
 ## Remote Decode
 
 [[autodoc]] utils.remote_utils.remote_decode
+
+## Remote Encode
+
+[[autodoc]] utils.remote_utils.remote_encode

docs/source/en/hybrid_inference/overview.md

@@ -36,7 +36,7 @@ Hybrid Inference offers a fast and simple way to offload local generation requir
 ## Available Models
 
 * **VAE Decode 🖼️:** Quickly decode latent representations into high-quality images without compromising performance or workflow speed.
-* **VAE Encode 🔢 (coming soon):** Efficiently encode images into latent representations for generation and training.
+* **VAE Encode 🔢:** Efficiently encode images into latent representations for generation and training.
 * **Text Encoders 📃 (coming soon):** Compute text embeddings for your prompts quickly and accurately, ensuring a smooth and high-quality workflow.
 
 ---
@@ -46,9 +46,15 @@ Hybrid Inference offers a fast and simple way to offload local generation requir
 * **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
 * **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
 
+## Changelog
+
+- March 10 2025: Added VAE encode
+- March 2 2025: Initial release with VAE decoding
+
 ## Contents
 
-The documentation is organized into two sections:
+The documentation is organized into three sections:
 
 * **VAE Decode** Learn the basics of how to use VAE Decode with Hybrid Inference.
+* **VAE Encode** Learn the basics of how to use VAE Encode with Hybrid Inference.
 * **API Reference** Dive into task-specific settings and parameters.

docs/source/en/hybrid_inference/vae_encode.md

@@ -0,0 +1,183 @@
+# Getting Started: VAE Encode with Hybrid Inference
+
+VAE encode is used for training, image-to-image and image-to-video - turning into images or videos into latent representations.
+
+## Memory
+
+These tables demonstrate the VRAM requirements for VAE encode with SD v1 and SD XL on different GPUs.
+
+For the majority of these GPUs the memory usage % dictates other models (text encoders, UNet/Transformer) must be offloaded, or tiled encoding has to be used which increases time taken and impacts quality.
+
+<details><summary>SD v1.5</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory (%) |   Tiled Time (secs) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|-------------:|--------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.015 |      3.51901 |               0.015 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.004 |      1.3154  |               0.005 |            1.3154  |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.402 |     47.1852  |               0.496 |            3.51901 |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.078 |     12.2658  |               0.094 |            3.51901 |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.023 |      5.30105 |               0.023 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.006 |      1.98152 |               0.006 |            1.98152 |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            0.574 |     71.08    |               0.656 |            5.30105 |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.111 |     18.4772  |               0.14  |            5.30105 |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.032 |      3.52782 |               0.032 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.01  |      1.31869 |               0.009 |            1.31869 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            0.742 |     47.3033  |               0.954 |            3.52782 |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.136 |     12.2965  |               0.207 |            3.52782 |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.036 |      8.51761 |               0.036 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.01  |      3.18387 |               0.01  |            3.18387 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |            0.863 |     86.7424  |               1.191 |            8.51761 |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.157 |     29.6888  |               0.227 |            8.51761 |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.051 |     10.6941  |               0.051 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.015 |      3.99743 |               0.015 |            3.99743 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |            1.217 |     96.054   |               1.482 |           10.6941  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.223 |     37.2751  |               0.327 |           10.6941  |
+
+
+</details>
+
+<details><summary>SDXL</summary>
+
+| GPU                           | Resolution   |   Time (seconds) |   Memory Consumed (%) |   Tiled Time (seconds) |   Tiled Memory (%) |
+|:------------------------------|:-------------|-----------------:|----------------------:|-----------------------:|-------------------:|
+| NVIDIA GeForce RTX 4090       | 512x512      |            0.029 |               4.95707 |                  0.029 |            4.95707 |
+| NVIDIA GeForce RTX 4090       | 256x256      |            0.007 |               2.29666 |                  0.007 |            2.29666 |
+| NVIDIA GeForce RTX 4090       | 2048x2048    |            0.873 |              66.3452  |                  0.863 |           15.5649  |
+| NVIDIA GeForce RTX 4090       | 1024x1024    |            0.142 |              15.5479  |                  0.143 |           15.5479  |
+| NVIDIA GeForce RTX 4080 SUPER | 512x512      |            0.044 |               7.46735 |                  0.044 |            7.46735 |
+| NVIDIA GeForce RTX 4080 SUPER | 256x256      |            0.01  |               3.4597  |                  0.01  |            3.4597  |
+| NVIDIA GeForce RTX 4080 SUPER | 2048x2048    |            1.317 |              87.1615  |                  1.291 |           23.447   |
+| NVIDIA GeForce RTX 4080 SUPER | 1024x1024    |            0.213 |              23.4215  |                  0.214 |           23.4215  |
+| NVIDIA GeForce RTX 3090       | 512x512      |            0.058 |               5.65638 |                  0.058 |            5.65638 |
+| NVIDIA GeForce RTX 3090       | 256x256      |            0.016 |               2.45081 |                  0.016 |            2.45081 |
+| NVIDIA GeForce RTX 3090       | 2048x2048    |            1.755 |              77.8239  |                  1.614 |           18.4193  |
+| NVIDIA GeForce RTX 3090       | 1024x1024    |            0.265 |              18.4023  |                  0.265 |           18.4023  |
+| NVIDIA GeForce RTX 3080       | 512x512      |            0.064 |              13.6568  |                  0.064 |           13.6568  |
+| NVIDIA GeForce RTX 3080       | 256x256      |            0.018 |               5.91728 |                  0.018 |            5.91728 |
+| NVIDIA GeForce RTX 3080       | 2048x2048    |          OOM     |             OOM       |                  1.866 |           44.4717  |
+| NVIDIA GeForce RTX 3080       | 1024x1024    |            0.302 |              44.4308  |                  0.302 |           44.4308  |
+| NVIDIA GeForce RTX 3070       | 512x512      |            0.093 |              17.1465  |                  0.093 |           17.1465  |
+| NVIDIA GeForce RTX 3070       | 256x256      |            0.025 |               7.42931 |                  0.026 |            7.42931 |
+| NVIDIA GeForce RTX 3070       | 2048x2048    |          OOM     |             OOM       |                  2.674 |           55.8355  |
+| NVIDIA GeForce RTX 3070       | 1024x1024    |            0.443 |              55.7841  |                  0.443 |           55.7841  |
+
+</details>
+
+## Available VAEs
+
+|   | **Endpoint** | **Model** |
+|:-:|:-----------:|:--------:|
+| **Stable Diffusion v1** | [https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud](https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud) | [`stabilityai/sd-vae-ft-mse`](https://hf.co/stabilityai/sd-vae-ft-mse) |
+| **Stable Diffusion XL** | [https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud](https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud) | [`madebyollin/sdxl-vae-fp16-fix`](https://hf.co/madebyollin/sdxl-vae-fp16-fix) |
+| **Flux** | [https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud](https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud) | [`black-forest-labs/FLUX.1-schnell`](https://hf.co/black-forest-labs/FLUX.1-schnell) |
+
+
+> [!TIP]
+> Model support can be requested [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+## Code
+
+> [!TIP]
+> Install `diffusers` from `main` to run the code: `pip install git+https://github.com/huggingface/diffusers@main`
+
+
+A helper method simplifies interacting with Hybrid Inference.
+
+```python
+from diffusers.utils.remote_utils import remote_encode
+```
+
+### Basic example
+
+Let's encode an image, then decode it to demonstrate.
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"/>
+</figure>
+
+<details><summary>Code</summary>
+
+```python
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode
+
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg?download=true")
+
+latent = remote_encode(
+    endpoint="https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/",
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+
+decoded = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/decoded.png"/>
+</figure>
+
+
+### Generation
+
+Now let's look at a generation example, we'll encode the image, generate then remotely decode too!
+
+<details><summary>Code</summary>
+
+```python
+import torch
+from diffusers import StableDiffusionImg2ImgPipeline
+from diffusers.utils import load_image
+from diffusers.utils.remote_utils import remote_decode, remote_encode
+
+pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=None,
+).to("cuda")
+
+init_image = load_image(
+    "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+)
+init_image = init_image.resize((768, 512))
+
+init_latent = remote_encode(
+    endpoint="https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/",
+    image=init_image,
+    scaling_factor=0.18215,
+)
+
+prompt = "A fantasy landscape, trending on artstation"
+latent = pipe(
+    prompt=prompt,
+    image=init_latent,
+    strength=0.75,
+    output_type="latent",
+).images
+
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.18215,
+)
+image.save("fantasy_landscape.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/fantasy_landscape.png"/>
+</figure>
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.

docs/source/en/optimization/memory.md

@@ -198,6 +198,18 @@ export_to_video(video, "output.mp4", fps=8)
 
 Group offloading (for CUDA devices with support for asynchronous data transfer streams) overlaps data transfer and computation to reduce the overall execution time compared to sequential offloading. This is enabled using layer prefetching with CUDA streams. The next layer to be executed is loaded onto the accelerator device while the current layer is being executed - this increases the memory requirements slightly. Group offloading also supports leaf-level offloading (equivalent to sequential CPU offloading) but can be made much faster when using streams.
 
+<Tip>
+
+- Group offloading may not work with all models out-of-the-box. If the forward implementations of the model contain weight-dependent device-casting of inputs, it may clash with the offloading mechanism's handling of device-casting.
+- The `offload_type` parameter can be set to either `block_level` or `leaf_level`. `block_level` offloads groups of `torch::nn::ModuleList` or `torch::nn:Sequential` modules based on a configurable attribute `num_blocks_per_group`. For example, if you set `num_blocks_per_group=2` on a standard transformer model containing 40 layers, it will onload/offload 2 layers at a time for a total of 20 onload/offloads. This drastically reduces the VRAM requirements. `leaf_level` offloads individual layers at the lowest level, which is equivalent to sequential offloading. However, unlike sequential offloading, group offloading can be made much faster when using streams, with minimal compromise to end-to-end generation time.
+- The `use_stream` parameter can be used with CUDA devices to enable prefetching layers for onload. It defaults to `False`. Layer prefetching allows overlapping computation and data transfer of model weights, which drastically reduces the overall execution time compared to other offloading methods. However, it can increase the CPU RAM usage significantly. Ensure that available CPU RAM that is at least twice the size of the model when setting `use_stream=True`. You can find more information about CUDA streams [here](https://pytorch.org/docs/stable/generated/torch.cuda.Stream.html)
+- If specifying `use_stream=True` on VAEs with tiling enabled, make sure to do a dummy forward pass (possibly with dummy inputs) before the actual inference to avoid device-mismatch errors. This may not work on all implementations. Please open an issue if you encounter any problems.
+- The parameter `low_cpu_mem_usage` can be set to `True` to reduce CPU memory usage when using streams for group offloading. This is useful when the CPU memory is the bottleneck, but it may counteract the benefits of using streams and increase the overall execution time. The CPU memory savings come from creating pinned-tensors on-the-fly instead of pre-pinning them. This parameter is better suited for using `leaf_level` offloading.
+
+For more information about available parameters and an explanation of how group offloading works, refer to [`~hooks.group_offloading.apply_group_offloading`].
+
+</Tip>
+
 ## FP8 layerwise weight-casting
 
 PyTorch supports `torch.float8_e4m3fn` and `torch.float8_e5m2` as weight storage dtypes, but they can't be used for computation in many different tensor operations due to unimplemented kernel support. However, you can use these dtypes to store model weights in fp8 precision and upcast them on-the-fly when the layers are used in the forward pass. This is known as layerwise weight-casting.
@@ -235,6 +247,14 @@ In the above example, layerwise casting is enabled on the transformer component
 
 However, you gain more control and flexibility by directly utilizing the [`~hooks.layerwise_casting.apply_layerwise_casting`] function instead of [`~ModelMixin.enable_layerwise_casting`].
 
+<Tip>
+
+- Layerwise casting may not work with all models out-of-the-box. Sometimes, the forward implementations of the model might contain internal typecasting of weight values. Such implementations are not supported due to the currently simplistic implementation of layerwise casting, which assumes that the forward pass is independent of the weight precision and that the input dtypes are always in `compute_dtype`. An example of an incompatible implementation can be found [here](https://github.com/huggingface/transformers/blob/7f5077e53682ca855afc826162b204ebf809f1f9/src/transformers/models/t5/modeling_t5.py#L294-L299).
+- Layerwise casting may fail on custom modeling implementations that make use of [PEFT](https://github.com/huggingface/peft) layers. Some minimal checks to handle this case is implemented but is not extensively tested or guaranteed to work in all cases.
+- It can be also be applied partially to specific layers of a model. Partially applying layerwise casting can either be done manually by calling the `apply_layerwise_casting` function on specific internal modules, or by specifying the `skip_modules_pattern` and `skip_modules_classes` parameters for a root module. These parameters are particularly useful for layers such as normalization and modulation.
+
+</Tip>
+
 ## Channels-last memory format
 
 The channels-last memory format is an alternative way of ordering NCHW tensors in memory to preserve dimension ordering. Channels-last tensors are ordered in such a way that the channels become the densest dimension (storing images pixel-per-pixel). Since not all operators currently support the channels-last format, it may result in worst performance but you should still try and see if it works for your model.

docs/source/ko/training/controlnet.md

@@ -66,12 +66,6 @@ from accelerate.utils import write_basic_config
 write_basic_config()
 ```
 
-## 원을 채우는 데이터셋
-
-원본 데이터셋은 ControlNet [repo](https://huggingface.co/lllyasviel/ControlNet/blob/main/training/fill50k.zip)에 올라와있지만, 우리는 [여기](https://huggingface.co/datasets/fusing/fill50k)에 새롭게 다시 올려서 🤗 Datasets 과 호환가능합니다. 그래서 학습 스크립트 상에서 데이터 불러오기를 다룰 수 있습니다.
-
-우리의 학습 예시는 원래 ControlNet의 학습에 쓰였던 [`stable-diffusion-v1-5/stable-diffusion-v1-5`](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5)을 사용합니다. 그렇지만 ControlNet은 대응되는 어느 Stable Diffusion 모델([`CompVis/stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4)) 혹은 [`stabilityai/stable-diffusion-2-1`](https://huggingface.co/stabilityai/stable-diffusion-2-1)의 증가를 위해 학습될 수 있습니다.
-
 자체 데이터셋을 사용하기 위해서는 [학습을 위한 데이터셋 생성하기](create_dataset) 가이드를 확인하세요.
 
 ## 학습

examples/advanced_diffusion_training/README_flux.md

@@ -79,13 +79,13 @@ This command will prompt you for a token. Copy-paste yours from your [settings/t
 ### Target Modules
 When LoRA was first adapted from language models to diffusion models, it was applied to the cross-attention layers in the Unet that relate the image representations with the prompts that describe them. 
 More recently, SOTA text-to-image diffusion models replaced the Unet with a diffusion Transformer(DiT). With this change, we may also want to explore 
-applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma seperated string
+applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma separated string
 the exact modules for LoRA training. Here are some examples of target modules you can provide: 
 - for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0"`
 - to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2"`
 - to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2,norm1_context.linear, norm1.linear,norm.linear,proj_mlp,proj_out"`
 > [!NOTE]
-> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma seperated string:
+> `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma separated string:
 > **single DiT blocks**: to target the ith single transformer block, add the prefix `single_transformer_blocks.i`, e.g. - `single_transformer_blocks.i.attn.to_k`
 > **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k` 
 > [!NOTE]

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -378,7 +378,7 @@ def parse_args(input_args=None):
         default=None,
         help="the concept to use to initialize the new inserted tokens when training with "
         "--train_text_encoder_ti = True. By default, new tokens (<si><si+1>) are initialized with random value. "
-        "Alternatively, you could specify a different word/words whos value will be used as the starting point for the new inserted tokens. "
+        "Alternatively, you could specify a different word/words whose value will be used as the starting point for the new inserted tokens. "
         "--num_new_tokens_per_abstraction is ignored when initializer_concept is provided",
     )
     parser.add_argument(
@@ -662,7 +662,7 @@ def parse_args(input_args=None):
         type=str,
         default=None,
         help=(
-            "The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. "
+            "The transformer modules to apply LoRA training on. Please specify the layers in a comma separated. "
             'E.g. - "to_k,to_q,to_v,to_out.0" will result in lora training of attention layers only. For more examples refer to https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/README_flux.md'
         ),
     )

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -662,7 +662,7 @@ def parse_args(input_args=None):
         action="store_true",
         default=False,
         help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
             "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
         ),
     )

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -773,7 +773,7 @@ def parse_args(input_args=None):
         action="store_true",
         default=False,
         help=(
-            "Wether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
+            "Whether to train a DoRA as proposed in- DoRA: Weight-Decomposed Low-Rank Adaptation https://arxiv.org/abs/2402.09353. "
             "Note: to use DoRA you need to install peft from main, `pip install git+https://github.com/huggingface/peft.git`"
         ),
     )
@@ -1875,7 +1875,7 @@ def main(args):
     # pack the statically computed variables appropriately here. This is so that we don't
     # have to pass them to the dataloader.
 
-    # if --train_text_encoder_ti we need add_special_tokens to be True fo textual inversion
+    # if --train_text_encoder_ti we need add_special_tokens to be True for textual inversion
     add_special_tokens = True if args.train_text_encoder_ti else False
 
     if not train_dataset.custom_instance_prompts:

examples/cogview4-control/README.md

@@ -0,0 +1,201 @@
+# Training CogView4 Control
+
+This (experimental) example shows how to train Control LoRAs with [CogView4](https://huggingface.co/THUDM/CogView4-6B) by conditioning it with additional structural controls (like depth maps, poses, etc.). We provide a script for full fine-tuning, too, refer to [this section](#full-fine-tuning). To know more about CogView4 Control family, refer to the following resources:
+
+To incorporate additional condition latents, we expand the input features of CogView-4 from 64 to 128. The first 64 channels correspond to the original input latents to be denoised, while the latter 64 channels correspond to control latents. This expansion happens on the `patch_embed` layer, where the combined latents are projected to the expected feature dimension of rest of the network. Inference is performed using the `CogView4ControlPipeline`.
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [CogView4 Hugging Face page](https://huggingface.co/THUDM/CogView4-6B), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+The example command below shows how to launch fine-tuning for pose conditions. The dataset ([`raulc0399/open_pose_controlnet`](https://huggingface.co/datasets/raulc0399/open_pose_controlnet)) being used here already has the pose conditions of the original images, so we don't have to compute them.
+
+```bash
+accelerate launch train_control_lora_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control-lora" \
+  --mixed_precision="bf16" \
+  --train_batch_size=1 \
+  --rank=64 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=5000 \
+  --validation_image="openpose.png" \
+  --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+`openpose.png` comes from [here](https://huggingface.co/Adapter/t2iadapter/resolve/main/openpose.png).
+
+You need to install `diffusers` from the branch of [this PR](https://github.com/huggingface/diffusers/pull/9999). When it's merged, you should install `diffusers` from the `main`.
+
+The training script exposes additional CLI args that might be useful to experiment with:
+
+* `use_lora_bias`: When set, additionally trains the biases of the `lora_B` layer. 
+* `train_norm_layers`: When set, additionally trains the normalization scales. Takes care of saving and loading.
+* `lora_layers`: Specify the layers you want to apply LoRA to. If you specify "all-linear", all the linear layers will be LoRA-attached.
+
+### Training with DeepSpeed
+
+It's possible to train with [DeepSpeed](https://github.com/microsoft/DeepSpeed), specifically leveraging the Zero2 system optimization. To use it, save the following config to an YAML file (feel free to modify as needed):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+And then while launching training, pass the config file:
+
+```bash
+accelerate launch --config_file=CONFIG_FILE.yaml ...
+```
+
+### Inference
+
+The pose images in our dataset were computed using the [`controlnet_aux`](https://github.com/huggingface/controlnet_aux) library. Let's install it first:
+
+```bash
+pip install controlnet_aux
+```
+
+And then we are ready:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+pipe = CogView4ControlPipeline.from_pretrained("THUDM/CogView4-6B", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("...") # change this.
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  joint_attention_kwargs={"scale": 0.9},
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Full fine-tuning
+
+We provide a non-LoRA version of the training script `train_control_cogview4.py`. Here is an example command:
+
+```bash
+accelerate launch --config_file=accelerate_ds2.yaml train_control_cogview4.py \
+  --pretrained_model_name_or_path="THUDM/CogView4-6B" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control" \
+  --mixed_precision="bf16" \
+  --train_batch_size=2 \
+  --dataloader_num_workers=4 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --proportion_empty_prompts=0.2 \
+  --learning_rate=5e-5 \
+  --adam_weight_decay=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="cosine" \
+  --lr_warmup_steps=1000 \
+  --checkpointing_steps=1000 \
+  --max_train_steps=10000 \
+  --validation_steps=200 \
+  --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
+  --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+Change the `validation_image` and `validation_prompt` as needed.
+
+For inference, this time, we will run:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import CogView4ControlPipeline, CogView4Transformer2DModel
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+transformer = CogView4Transformer2DModel.from_pretrained("...") # change this.
+pipe = CogView4ControlPipeline.from_pretrained(
+  "THUDM/CogView4-6B",  transformer=transformer, torch_dtype=torch.bfloat16
+).to("cuda")
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  control_image=image,
+  num_inference_steps=50,
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Things to note
+
+* The scripts provided in this directory are experimental and educational. This means we may have to tweak things around to get good results on a given condition. We believe this is best done with the community 🤗
+* The scripts are not memory-optimized but we offload the VAE and the text encoders to CPU when they are not used if `--offload` is specified. 
+* We can extract LoRAs from the fully fine-tuned model. While we currently don't provide any utilities for that, users are welcome to refer to [this script](https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/master/control_lora_create.py) that provides a similar functionality. 

examples/cogview4-control/requirements.txt

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

examples/community/README.md

@@ -24,12 +24,12 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Long Prompt Weighting Stable Diffusion                                                                                                | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt.                                                                                                                                                                                                                                                                                                                                                                                                  | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion)         | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/long_prompt_weighting_stable_diffusion.ipynb)                                                                                        |                           [SkyTNT](https://github.com/SkyTNT) |
 | Speech to Image                                                                                                                       | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images                                                                                                                                                                                                                                                                                                                                                                                                            | [Speech to Image](#speech-to-image)                                                       |[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/speech_to_image.ipynb)                                                                                                                                                                                                   |             [Mikail Duzenli](https://github.com/MikailINTech)
 | Wild Card Stable Diffusion                                                                                                            | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values                                                                                                                                                                                                                                                                            | [Wildcard Stable Diffusion](#wildcard-stable-diffusion)                                   | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/wildcard_stable_diffusion.ipynb)                                                                                                                                                                                 |              [Shyam Sudhakaran](https://github.com/shyamsn97) |
-| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | -                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
+| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts.                                                                                                                                                                                                                                                                                                            | [Composable Stable Diffusion](#composable-stable-diffusion)                               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/composable_stable_diffusion.ipynb)                                                                                                                                                                                                                 |                      [Mark Rich](https://github.com/MarkRich) |
 | Seed Resizing Stable Diffusion                                                                                                        | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation.                                                                                                                                                                                                                                                                                                                                                                                       | [Seed Resizing](#seed-resizing)                                                           | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/seed_resizing.ipynb)                                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Imagic Stable Diffusion                                                                                                               | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image                                                                                                                                                                                                                                                                                                                                                                                                                   | [Imagic Stable Diffusion](#imagic-stable-diffusion)                                       | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/imagic_stable_diffusion.ipynb)                                                                                                                                                                                                  |                      [Mark Rich](https://github.com/MarkRich) |
 | Multilingual Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in 50 different languages.                                                                                                                                                                                                                                                                                                                                                                                                                               | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/multilingual_stable_diffusion.ipynb)                                                                                                                                                                             |          [Juan Carlos Piñeros](https://github.com/juancopi81) |
 | GlueGen Stable Diffusion                                                                                                         | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter.                                                                                                                                                                                                                                                                                                                                                                                                                               | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline)                  | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/gluegen_stable_diffusion.ipynb)                                                                                                                                                                                                                  |          [Phạm Hồng Vinh](https://github.com/rootonchair) |
-| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | -                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
+| Image to Image Inpainting Stable Diffusion                                                                                            | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting                                                                                                                                                                                                                                                                                                                                                                                                            | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/image_to_image_inpainting_stable_diffusion.ipynb)                                                                                                                                                                                                                  |                    [Alex McKinney](https://github.com/vvvm23) |
 | Text Based Inpainting Stable Diffusion                                                                                                | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting                                                                                                                                                                                                                                                                                                                                                                                              | [Text Based Inpainting Stable Diffusion](#text-based-inpainting-stable-diffusion)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/text_based_inpainting_stable_dffusion.ipynb)                                                                                                                                                                                                    |                   [Dhruv Karan](https://github.com/unography) |
 | Bit Diffusion                                                                                                                         | Diffusion on discrete data                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | [Bit Diffusion](#bit-diffusion)                                                           | -  |                       [Stuti R.](https://github.com/kingstut) |
 | K-Diffusion Stable Diffusion                                                                                                          | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py)                                                                                                                                                                                                                                                                                                                                                                  | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion)                   | -  |    [Patrick von Platen](https://github.com/patrickvonplaten/) |
@@ -41,7 +41,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | UnCLIP Image Interpolation Pipeline                                                                                                   | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings                                                                                                                                                                                                                                                                                                                                                                | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline)               | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/unclip_image_interpolation.ipynb)| [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
 | DDIM Noise Comparative Analysis Pipeline                                                                                              | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227))                                                                                                                                                                                                                                                                                                                             | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline)     | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/ddim_noise_comparative_analysis.ipynb)|              [Aengus (Duc-Anh)](https://github.com/aengusng8) |
 | CLIP Guided Img2Img Stable Diffusion Pipeline                                                                                         | Doing CLIP guidance for image to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                  | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion)             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/clip_guided_img2img_stable_diffusion.ipynb) |               [Nipun Jindal](https://github.com/nipunjindal/) |
-| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
+| TensorRT Stable Diffusion Text to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline)      | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/tensorrt_text2image_stable_diffusion_pipeline.ipynb) |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
 | EDICT Image Editing Pipeline                                                                                                          | Diffusion pipeline for text-guided image editing                                                                                                                                                                                                                                                                                                                                                                                                                                                         | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline)                             | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/edict_image_pipeline.ipynb) |                    [Joqsan Azocar](https://github.com/Joqsan) |
 | Stable Diffusion RePaint                                                                                                              | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting.                                                                                                                                                                                                                                                                                                                                                                                                               | [Stable Diffusion RePaint](#stable-diffusion-repaint )|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_repaint.ipynb)|                  [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
 | TensorRT Stable Diffusion Image to Image Pipeline                                                                                                    | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT                                                                                                                                                                                                                                                                                                                                                                                                                                      | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline)      | - |              [Asfiya Baig](https://github.com/asfiyab-nvidia) |
@@ -58,7 +58,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_fabric.ipynb)| [Shauray Singh](https://shauray8.github.io/about_shauray/) |
 | sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
 | sketch inpaint xl - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion XL Pipeline](#stable-diffusion-xl-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
-| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) |
+| prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/prompt_2_prompt_pipeline.ipynb) | [Umer H. Adil](https://twitter.com/UmerHAdil) |
 |   Latent Consistency Pipeline                                                                                                    | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Latent Consistency Pipeline](#latent-consistency-pipeline)      | - |              [Simian Luo](https://github.com/luosiallen) |
 |   Latent Consistency Img2img Pipeline                                                                                                    | Img2img pipeline for Latent Consistency Models                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline)      | - |              [Logan Zoellner](https://github.com/nagolinc) |
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
@@ -954,6 +954,7 @@ for i in range(args.num_images):
     images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
 grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
 tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
+print("Image saved successfully!")
 ```
 
 ### Imagic Stable Diffusion
@@ -1269,28 +1270,39 @@ The aim is to overlay two images, then mask out the boundary between `image` and
 For example, this could be used to place a logo on a shirt and make it blend seamlessly.
 
 ```python
-import PIL
 import torch
-
+import requests
+from PIL import Image
+from io import BytesIO
 from diffusers import DiffusionPipeline
 
-image_path = "./path-to-image.png"
-inner_image_path = "./path-to-inner-image.png"
-mask_path = "./path-to-mask.png"
+image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+inner_image_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
 
-init_image = PIL.Image.open(image_path).convert("RGB").resize((512, 512))
-inner_image = PIL.Image.open(inner_image_path).convert("RGBA").resize((512, 512))
-mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
+def load_image(url, mode="RGB"):
+    response = requests.get(url)
+    if response.status_code == 200:
+        return Image.open(BytesIO(response.content)).convert(mode).resize((512, 512))
+    else:
+        raise FileNotFoundError(f"Could not retrieve image from {url}")
+
+
+init_image = load_image(image_url, mode="RGB")
+inner_image = load_image(inner_image_url, mode="RGBA")
+mask_image = load_image(mask_url, mode="RGB")
 
 pipe = DiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-inpainting",
+    "stable-diffusion-v1-5/stable-diffusion-inpainting",
     custom_pipeline="img2img_inpainting",
     torch_dtype=torch.float16
 )
 pipe = pipe.to("cuda")
 
-prompt = "Your prompt here!"
+prompt = "a mecha robot sitting on a bench"
 image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
+
+image.save("output.png")
 ```
 
 ![2 by 2 grid demonstrating image to image inpainting.](https://user-images.githubusercontent.com/44398246/203506577-ec303be4-887e-4ebd-a773-c83fcb3dd01a.png)
@@ -3252,14 +3264,19 @@ Here's a full example for `ReplaceEdit``:
 
 ```python
 import torch
-import numpy as np
-import matplotlib.pyplot as plt
 from diffusers import DiffusionPipeline
+import numpy as np
+from PIL import Image
 
-pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="pipeline_prompt2prompt").to("cuda")
+pipe = DiffusionPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4", 
+    custom_pipeline="pipeline_prompt2prompt"
+).to("cuda")
 
-prompts = ["A turtle playing with a ball",
-           "A monkey playing with a ball"]
+prompts = [
+    "A turtle playing with a ball",
+    "A monkey playing with a ball"
+]
 
 cross_attention_kwargs = {
     "edit_type": "replace",
@@ -3267,7 +3284,15 @@ cross_attention_kwargs = {
     "self_replace_steps": 0.4
 }
 
-outputs = pipe(prompt=prompts, height=512, width=512, num_inference_steps=50, cross_attention_kwargs=cross_attention_kwargs)
+outputs = pipe(
+    prompt=prompts,
+    height=512,
+    width=512,
+    num_inference_steps=50,
+    cross_attention_kwargs=cross_attention_kwargs
+)
+
+outputs.images[0].save("output_image_0.png")
 ```
 
 And abbreviated examples for the other edits:

examples/community/mixture_tiling_sdxl.py

@@ -1,4 +1,4 @@
-# Copyright 2025 The HuggingFace Team. All rights reserved.
+# Copyright 2025 The DEVAIEXP Team and The HuggingFace Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -1070,32 +1070,32 @@ class StableDiffusionXLTilingPipeline(
                     text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
                 else:
                     text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
-                    add_time_ids = self._get_add_time_ids(
-                        original_size,
-                        crops_coords_top_left[row][col],
-                        target_size,
+                add_time_ids = self._get_add_time_ids(
+                    original_size,
+                    crops_coords_top_left[row][col],
+                    target_size,
+                    dtype=prompt_embeds.dtype,
+                    text_encoder_projection_dim=text_encoder_projection_dim,
+                )
+                if negative_original_size is not None and negative_target_size is not None:
+                    negative_add_time_ids = self._get_add_time_ids(
+                        negative_original_size,
+                        negative_crops_coords_top_left[row][col],
+                        negative_target_size,
                         dtype=prompt_embeds.dtype,
                         text_encoder_projection_dim=text_encoder_projection_dim,
                     )
-                    if negative_original_size is not None and negative_target_size is not None:
-                        negative_add_time_ids = self._get_add_time_ids(
-                            negative_original_size,
-                            negative_crops_coords_top_left[row][col],
-                            negative_target_size,
-                            dtype=prompt_embeds.dtype,
-                            text_encoder_projection_dim=text_encoder_projection_dim,
-                        )
-                    else:
-                        negative_add_time_ids = add_time_ids
+                else:
+                    negative_add_time_ids = add_time_ids
 
-                    if self.do_classifier_free_guidance:
-                        prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-                        add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
-                        add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+                if self.do_classifier_free_guidance:
+                    prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+                    add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+                    add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
 
-                    prompt_embeds = prompt_embeds.to(device)
-                    add_text_embeds = add_text_embeds.to(device)
-                    add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+                prompt_embeds = prompt_embeds.to(device)
+                add_text_embeds = add_text_embeds.to(device)
+                add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
                 addition_embed_type_row.append((prompt_embeds, add_text_embeds, add_time_ids))
             embeddings_and_added_time.append(addition_embed_type_row)
 

examples/controlnet/train_controlnet.py

@@ -152,9 +152,7 @@ def log_validation(
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

examples/controlnet/train_controlnet_flux.py

@@ -166,9 +166,7 @@ def log_validation(
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

examples/controlnet/train_controlnet_sd3.py

@@ -1283,8 +1283,8 @@ def main(args):
                 noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
 
                 # Get the text embedding for conditioning
-                prompt_embeds = batch["prompt_embeds"]
-                pooled_prompt_embeds = batch["pooled_prompt_embeds"]
+                prompt_embeds = batch["prompt_embeds"].to(dtype=weight_dtype)
+                pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(dtype=weight_dtype)
 
                 # controlnet(s) inference
                 controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)

examples/controlnet/train_controlnet_sdxl.py

@@ -157,9 +157,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step,
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

examples/research_projects/anytext/README.md

@@ -1,20 +1,27 @@
-# AnyTextPipeline Pipeline
+# AnyTextPipeline
 
 Project page: https://aigcdesigngroup.github.io/homepage_anytext
 
 "AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy."
 
-Each text line that needs to be generated should be enclosed in double quotes. For any usage questions, please refer to the [paper](https://arxiv.org/abs/2311.03054).
+> **Note:** Each text line that needs to be generated should be enclosed in double quotes.
 
+For any usage questions, please refer to the [paper](https://arxiv.org/abs/2311.03054).
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/gist/tolgacangoz/b87ec9d2f265b448dd947c9d4a0da389/anytext.ipynb)
 
 ```py
+# This example requires the `anytext_controlnet.py` file:
+# !git clone --depth 1 https://github.com/huggingface/diffusers.git
+# %cd diffusers/examples/research_projects/anytext
+# Let's choose a font file shared by an HF staff:
+# !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
+
 import torch
 from diffusers import DiffusionPipeline
 from anytext_controlnet import AnyTextControlNetModel
 from diffusers.utils import load_image
 
-# I chose a font file shared by an HF staff:
-# !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
 
 anytext_controlnet = AnyTextControlNetModel.from_pretrained("tolgacangoz/anytext-controlnet", torch_dtype=torch.float16,
                                                             variant="fp16",)
@@ -26,6 +33,7 @@ pipe = DiffusionPipeline.from_pretrained("tolgacangoz/anytext", font_path="arial
 # generate image
 prompt = 'photo of caramel macchiato coffee on the table, top-down perspective, with "Any" "Text" written on it using cream'
 draw_pos = load_image("https://raw.githubusercontent.com/tyxsspa/AnyText/refs/heads/main/example_images/gen9.png")
+# There are two modes: "generate" and "edit". "edit" mode requires `ori_image` parameter for the image to be edited.
 image = pipe(prompt, num_inference_steps=20, mode="generate", draw_pos=draw_pos,
              ).images[0]
 image

examples/research_projects/anytext/anytext.py

@@ -146,14 +146,17 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 EXAMPLE_DOC_STRING = """
     Examples:
         ```py
+        >>> # This example requires the `anytext_controlnet.py` file:
+        >>> # !git clone --depth 1 https://github.com/huggingface/diffusers.git
+        >>> # %cd diffusers/examples/research_projects/anytext
+        >>> # Let's choose a font file shared by an HF staff:
+        >>> # !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
+
         >>> import torch
         >>> from diffusers import DiffusionPipeline
         >>> from anytext_controlnet import AnyTextControlNetModel
         >>> from diffusers.utils import load_image
 
-        >>> # I chose a font file shared by an HF staff:
-        >>> !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
-
         >>> anytext_controlnet = AnyTextControlNetModel.from_pretrained("tolgacangoz/anytext-controlnet", torch_dtype=torch.float16,
         ...                                                             variant="fp16",)
         >>> pipe = DiffusionPipeline.from_pretrained("tolgacangoz/anytext", font_path="arial-unicode-ms.ttf",
@@ -165,6 +168,7 @@ EXAMPLE_DOC_STRING = """
         >>> # generate image
         >>> prompt = 'photo of caramel macchiato coffee on the table, top-down perspective, with "Any" "Text" written on it using cream'
         >>> draw_pos = load_image("https://raw.githubusercontent.com/tyxsspa/AnyText/refs/heads/main/example_images/gen9.png")
+        >>> # There are two modes: "generate" and "edit". "edit" mode requires `ori_image` parameter for the image to be edited.
         >>> image = pipe(prompt, num_inference_steps=20, mode="generate", draw_pos=draw_pos,
         ...              ).images[0]
         >>> image
@@ -257,11 +261,11 @@ class EmbeddingManager(ModelMixin, ConfigMixin):
             idx = tokenized_text[i] == self.placeholder_token.to(device)
             if sum(idx) > 0:
                 if i >= len(self.text_embs_all):
-                    print("truncation for log images...")
+                    logger.warning("truncation for log images...")
                     break
                 text_emb = torch.cat(self.text_embs_all[i], dim=0)
                 if sum(idx) != len(text_emb):
-                    print("truncation for long caption...")
+                    logger.warning("truncation for long caption...")
                 text_emb = text_emb.to(embedded_text.device)
                 embedded_text[i][idx] = text_emb[: sum(idx)]
         return embedded_text
@@ -1058,6 +1062,8 @@ class AuxiliaryLatentModule(ModelMixin, ConfigMixin):
                     raise ValueError(f"Can't read ori_image image from {ori_image}!")
             elif isinstance(ori_image, torch.Tensor):
                 ori_image = ori_image.cpu().numpy()
+            elif isinstance(ori_image, PIL.Image.Image):
+                ori_image = np.array(ori_image.convert("RGB"))
             else:
                 if not isinstance(ori_image, np.ndarray):
                     raise ValueError(f"Unknown format of ori_image: {type(ori_image)}")

examples/research_projects/autoencoderkl/train_autoencoderkl.py

@@ -627,6 +627,7 @@ def main(args):
         ema_vae = EMAModel(vae.parameters(), model_cls=AutoencoderKL, model_config=vae.config)
     perceptual_loss = lpips.LPIPS(net="vgg").eval()
     discriminator = NLayerDiscriminator(input_nc=3, n_layers=3, use_actnorm=False).apply(weights_init)
+    discriminator = torch.nn.SyncBatchNorm.convert_sync_batchnorm(discriminator)
 
     # Taken from [Sayak Paul's Diffusers PR #6511](https://github.com/huggingface/diffusers/pull/6511/files)
     def unwrap_model(model):
@@ -951,13 +952,20 @@ def main(args):
                     logits_fake = discriminator(reconstructions)
                     disc_loss = hinge_d_loss if args.disc_loss == "hinge" else vanilla_d_loss
                     disc_factor = args.disc_factor if global_step >= args.disc_start else 0.0
-                    disc_loss = disc_factor * disc_loss(logits_real, logits_fake)
+                    d_loss = disc_factor * disc_loss(logits_real, logits_fake)
                     logs = {
-                        "disc_loss": disc_loss.detach().mean().item(),
+                        "disc_loss": d_loss.detach().mean().item(),
                         "logits_real": logits_real.detach().mean().item(),
                         "logits_fake": logits_fake.detach().mean().item(),
                         "disc_lr": disc_lr_scheduler.get_last_lr()[0],
                     }
+                    accelerator.backward(d_loss)
+                    if accelerator.sync_gradients:
+                        params_to_clip = discriminator.parameters()
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    disc_optimizer.step()
+                    disc_lr_scheduler.step()
+                    disc_optimizer.zero_grad(set_to_none=args.set_grads_to_none)
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)

examples/research_projects/controlnet/train_controlnet_webdataset.py

@@ -381,9 +381,7 @@ def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step)
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

examples/research_projects/pixart/train_pixart_controlnet_hf.py

@@ -164,9 +164,7 @@ def log_validation(
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

examples/research_projects/pytorch_xla/inference/flux/README.md

@@ -1,8 +1,6 @@
 # Generating images using Flux and PyTorch/XLA
 
-The `flux_inference` script shows how to do image generation using Flux on TPU devices using PyTorch/XLA. It uses the pallas kernel for flash attention for faster generation.
-
-It has been tested on [Trillium](https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus) TPU versions. No other TPU types have been tested.
+The `flux_inference` script shows how to do image generation using Flux on TPU devices using PyTorch/XLA. It uses the pallas kernel for flash attention for faster generation using custom flash block sizes for better performance on [Trillium](https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus) TPU versions. No other TPU types have been tested.
 
 ## Create TPU
 
@@ -23,20 +21,23 @@ Verify that PyTorch and PyTorch/XLA were installed correctly:
 python3 -c "import torch; import torch_xla;"
 ```
 
-Install dependencies
+Clone the diffusers repo and install dependencies
 
 ```bash
+git clone https://github.com/huggingface/diffusers.git
+cd diffusers
 pip install transformers accelerate sentencepiece structlog
-pushd ../../..
 pip install .
-popd
+cd examples/research_projects/pytorch_xla/inference/flux/
 ```
 
 ## Run the inference job
 
 ### Authenticate
 
-Run the following command to authenticate your token in order to download Flux weights.
+**Gated Model**
+
+As the model is gated, before using it with diffusers you first need to go to the [FLUX.1 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.1-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
 
 ```bash
 huggingface-cli login
@@ -50,51 +51,116 @@ python flux_inference.py
 
 The script loads the text encoders onto the CPU and the Flux transformer and VAE models onto the TPU. The first time the script runs, the compilation time is longer, while the cache stores the compiled programs. On subsequent runs, compilation is much faster and the subsequent passes being the fastest. 
 
-On a Trillium v6e-4, you should expect ~9 sec / 4 images or 2.25 sec / image (as devices run generation in parallel):
+On a Trillium v6e-4, you should expect ~6 sec / 4 images or 1.5 sec / image (as devices run generation in parallel):
 
 ```bash
 WARNING:root:libtpu.so and TPU device found. Setting PJRT_DEVICE=TPU.
-Loading checkpoint shards: 100%|███████████████████████████████| 2/2 [00:00<00:00,  7.01it/s]
-Loading pipeline components...:  40%|██████████▍               | 2/5 [00:00<00:00,  3.78it/s]You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
-Loading pipeline components...: 100%|██████████████████████████| 5/5 [00:00<00:00,  6.72it/s]
-2025-01-10 00:51:25 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:25 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:26 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-2025-01-10 00:51:26 [info     ] loading flux from black-forest-labs/FLUX.1-dev
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  4.29it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.26it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.27it/s]
-Loading pipeline components...: 100%|██████████████████████████| 3/3 [00:00<00:00,  3.25it/s]
-2025-01-10 00:51:34 [info     ] starting compilation run...   
-2025-01-10 00:51:35 [info     ] starting compilation run...   
-2025-01-10 00:51:37 [info     ] starting compilation run...   
-2025-01-10 00:51:37 [info     ] starting compilation run...   
-2025-01-10 00:52:52 [info     ] compilation took 78.5155531649998 sec.
-2025-01-10 00:52:53 [info     ] starting inference run...     
-2025-01-10 00:52:57 [info     ] compilation took 79.52986721400157 sec.
-2025-01-10 00:52:57 [info     ] compilation took 81.91776501700042 sec.
-2025-01-10 00:52:57 [info     ] compilation took 80.24951512600092 sec.
-2025-01-10 00:52:57 [info     ] starting inference run...     
-2025-01-10 00:52:57 [info     ] starting inference run...     
-2025-01-10 00:52:58 [info     ] starting inference run...     
-2025-01-10 00:53:22 [info     ] inference time: 25.112665320000815
-2025-01-10 00:53:30 [info     ] inference time: 7.7019307739992655
-2025-01-10 00:53:38 [info     ] inference time: 7.693858365000779
-2025-01-10 00:53:46 [info     ] inference time: 7.690621814001133
-2025-01-10 00:53:53 [info     ] inference time: 7.679490454000188
-2025-01-10 00:54:01 [info     ] inference time: 7.68949568500102
-2025-01-10 00:54:09 [info     ] inference time: 7.686633744000574
-2025-01-10 00:54:16 [info     ] inference time: 7.696786873999372
-2025-01-10 00:54:24 [info     ] inference time: 7.691988694999964
-2025-01-10 00:54:32 [info     ] inference time: 7.700649563999832
-2025-01-10 00:54:39 [info     ] inference time: 7.684993574001055
-2025-01-10 00:54:47 [info     ] inference time: 7.68343457499941
-2025-01-10 00:54:55 [info     ] inference time: 7.667921153999487
-2025-01-10 00:55:02 [info     ] inference time: 7.683585194001353
-2025-01-10 00:55:06 [info     ] avg. inference over 15 iterations took 8.61202360273334 sec.
-2025-01-10 00:55:07 [info     ] avg. inference over 15 iterations took 8.952725123600006 sec.
-2025-01-10 00:55:10 [info     ] inference time: 7.673799695001435
-2025-01-10 00:55:10 [info     ] avg. inference over 15 iterations took 8.849190365400379 sec.
-2025-01-10 00:55:10 [info     ] saved metric information as /tmp/metrics_report.txt
-2025-01-10 00:55:12 [info     ] avg. inference over 15 iterations took 8.940161458400205 sec.
+Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00,  7.06it/s]
+Loading pipeline components...:  60%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████                                                                                                | 3/5 [00:00<00:00,  6.80it/s]You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00,  6.28it/s]
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:53 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:54 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+2025-03-14 21:17:54 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.66it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  4.48it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.32it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.69it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.74it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.10it/s]
+2025-03-14 21:17:56 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...:   0%|                                                                                                                                                                                                                                                        | 0/3 [00:00<?, ?it/s]2025-03-14 21:17:56 [info     ] loading flux from black-forest-labs/FLUX.1-dev
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00,  3.55it/s]
+Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00,  1.46it/s]
+2025-03-14 21:18:34 [info     ] starting compilation run...   
+2025-03-14 21:18:37 [info     ] starting compilation run...   
+2025-03-14 21:18:38 [info     ] starting compilation run...   
+2025-03-14 21:18:39 [info     ] starting compilation run...   
+2025-03-14 21:18:41 [info     ] starting compilation run...   
+2025-03-14 21:18:41 [info     ] starting compilation run...   
+2025-03-14 21:18:42 [info     ] starting compilation run...   
+2025-03-14 21:18:43 [info     ] starting compilation run...   
+ 82%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                | 23/28 [13:35<03:04, 36.80s/it]2025-03-14 21:33:42.057559: W torch_xla/csrc/runtime/pjrt_computation_client.cc:667] Failed to deserialize executable: INTERNAL: TfrtTpuExecutable proto deserialization failed while parsing core program!
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:27<00:00, 35.28s/it]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:27<00:00, 35.26s/it]
+2025-03-14 21:36:38 [info     ] compilation took 1079.3314765350078 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:12<00:00, 34.73s/it]
+2025-03-14 21:36:38 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:12<00:00, 34.73s/it]
+2025-03-14 21:36:38 [info     ] compilation took 1081.89390801001 sec.
+2025-03-14 21:36:39 [info     ] starting inference run...     
+2025-03-14 21:36:39 [info     ] compilation took 1077.1543154849933 sec.
+2025-03-14 21:36:39 [info     ] compilation took 1075.7239800530078 sec.
+2025-03-14 21:36:39 [info     ] starting inference run...     
+2025-03-14 21:36:40 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:22<00:00, 35.10s/it]
+2025-03-14 21:36:50 [info     ] compilation took 1088.1632604240003 sec.
+2025-03-14 21:36:50 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:28<00:00, 35.32s/it]
+2025-03-14 21:36:55 [info     ] compilation took 1096.8027802760043 sec.
+2025-03-14 21:36:56 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:59<00:00, 36.40s/it]
+2025-03-14 21:37:08 [info     ] compilation took 1113.8591305939917 sec.
+2025-03-14 21:37:08 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [16:55<00:00, 36.26s/it]
+2025-03-14 21:37:22 [info     ] compilation took 1120.5590810020076 sec.
+2025-03-14 21:37:22 [info     ] starting inference run...     
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.00it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:09<00:00,  2.98it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.08it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:09<00:00,  2.82it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:08<00:00,  3.34it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.22it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.09it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:11<00:00,  2.41it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:06<00:00,  4.50it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.10it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.27it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  4.80it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.39it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:05<00:00,  5.39it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.67it/s]
+ 29%|█████████████████████████████████████████████████████████████████████████████▍                                                                                                                                                                                                 | 8/28 [00:01<00:03,  6.08it/s]/home/jfacevedo_google_com/diffusers/src/diffusers/image_processor.py:147: RuntimeWarning: invalid value encountered in cast
+  images = (images * 255).round().astype("uint8")
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.82it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.93it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.98it/s]
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.03it/s]2025-03-14 21:38:32 [info     ] inference time: 5.962021178987925
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.89it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.09it/s]
+2025-03-14 21:38:32 [info     ] avg. inference over 5 iterations took 7.2685392687970305 sec.
+2025-03-14 21:38:32 [info     ] avg. inference over 5 iterations took 7.402720856998348 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.01it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.89it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.96it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.06it/s]
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.01it/s]2025-03-14 21:38:38 [info     ] inference time: 5.950578948002658
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.87it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.09it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.00it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.86it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.99it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.05it/s]
+2025-03-14 21:38:43 [info     ] avg. inference over 5 iterations took 6.763298449796276 sec.
+ 71%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊                                                                             | 20/28 [00:03<00:01,  6.04it/s]2025-03-14 21:38:44 [info     ] inference time: 5.949129879008979
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.92it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.10it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+ 39%|██████████████████████████████████████████████████████████████████████████████████████████████████████████                                                                                                                                                                    | 11/28 [00:01<00:02,  5.98it/s]2025-03-14 21:38:46 [info     ] avg. inference over 5 iterations took 7.221068455604836 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.96it/s]
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.08it/s]
+ 93%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▋                   | 26/28 [00:04<00:00,  5.92it/s]2025-03-14 21:38:50 [info     ] inference time: 5.954778069004533
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.90it/s]
+ 11%|█████████████████████████████                                                                                                                                                                                                                                                  | 3/28 [00:00<00:04,  6.03it/s]2025-03-14 21:38:50 [info     ] avg. inference over 5 iterations took 6.05970350120042 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  6.02it/s]
+ 32%|███████████████████████████████████████████████████████████████████████████████████████                                                                                                                                                                                        | 9/28 [00:01<00:03,  5.99it/s]2025-03-14 21:38:51 [info     ] avg. inference over 5 iterations took 6.018543455796316 sec.
+ 54%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▋                                                                                                                             | 15/28 [00:02<00:02,  6.00it/s]2025-03-14 21:38:52 [info     ] avg. inference over 5 iterations took 5.9609976705978625 sec.
+100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:04<00:00,  5.97it/s]
+2025-03-14 21:38:56 [info     ] inference time: 5.944058528999449
+2025-03-14 21:38:56 [info     ] avg. inference over 5 iterations took 5.952113320800708 sec.
+2025-03-14 21:38:56 [info     ] saved metric information as /tmp/metrics_report.txt
 ```

examples/research_projects/pytorch_xla/inference/flux/flux_inference.py

@@ -9,6 +9,7 @@ import torch_xla.debug.metrics as met
 import torch_xla.debug.profiler as xp
 import torch_xla.distributed.xla_multiprocessing as xmp
 import torch_xla.runtime as xr
+from torch_xla.experimental.custom_kernel import FlashAttention
 
 from diffusers import FluxPipeline
 
@@ -36,6 +37,19 @@ def _main(index, args, text_pipe, ckpt_id):
         ckpt_id, text_encoder=None, tokenizer=None, text_encoder_2=None, tokenizer_2=None, torch_dtype=torch.bfloat16
     ).to(device0)
     flux_pipe.transformer.enable_xla_flash_attention(partition_spec=("data", None, None, None), is_flux=True)
+    FlashAttention.DEFAULT_BLOCK_SIZES = {
+        "block_q": 1536,
+        "block_k_major": 1536,
+        "block_k": 1536,
+        "block_b": 1536,
+        "block_q_major_dkv": 1536,
+        "block_k_major_dkv": 1536,
+        "block_q_dkv": 1536,
+        "block_k_dkv": 1536,
+        "block_q_dq": 1536,
+        "block_k_dq": 1536,
+        "block_k_major_dq": 1536,
+    }
 
     prompt = "photograph of an electronics chip in the shape of a race car with trillium written on its side"
     width = args.width
@@ -69,14 +83,14 @@ def _main(index, args, text_pipe, ckpt_id):
     xm.set_rng_state(seed=unique_seed, device=device0)
     times = []
     logger.info("starting inference run...")
+    with torch.no_grad():
+        prompt_embeds, pooled_prompt_embeds, text_ids = text_pipe.encode_prompt(
+            prompt=prompt, prompt_2=None, max_sequence_length=512
+        )
+    prompt_embeds = prompt_embeds.to(device0)
+    pooled_prompt_embeds = pooled_prompt_embeds.to(device0)
     for _ in range(args.itters):
         ts = perf_counter()
-        with torch.no_grad():
-            prompt_embeds, pooled_prompt_embeds, text_ids = text_pipe.encode_prompt(
-                prompt=prompt, prompt_2=None, max_sequence_length=512
-            )
-        prompt_embeds = prompt_embeds.to(device0)
-        pooled_prompt_embeds = pooled_prompt_embeds.to(device0)
 
         if args.profile:
             xp.trace_detached(f"localhost:{profiler_port}", str(profile_path), duration_ms=profile_duration)
@@ -92,7 +106,7 @@ def _main(index, args, text_pipe, ckpt_id):
         if index == 0:
             logger.info(f"inference time: {inference_time}")
         times.append(inference_time)
-    logger.info(f"avg. inference over {args.itters} iterations took {sum(times)/len(times)} sec.")
+    logger.info(f"avg. inference over {args.itters} iterations took {sum(times) / len(times)} sec.")
     image.save(f"/tmp/inference_out-{index}.png")
     if index == 0:
         metrics_report = met.metrics_report()

examples/t2i_adapter/train_t2i_adapter_sdxl.py

@@ -141,9 +141,7 @@ def log_validation(vae, unet, adapter, args, accelerator, weight_dtype, step):
                 validation_prompt = log["validation_prompt"]
                 validation_image = log["validation_image"]
 
-                formatted_images = []
-
-                formatted_images.append(np.asarray(validation_image))
+                formatted_images = [np.asarray(validation_image)]
 
                 for image in images:
                     formatted_images.append(np.asarray(image))

scripts/convert_cogview4_to_diffusers.py

@@ -53,8 +53,18 @@ args = parser.parse_args()
 # this is specific to `AdaLayerNormContinuous`:
 # diffusers implementation split the linear projection into the scale, shift while CogView4 split it tino shift, scale
 def swap_scale_shift(weight, dim):
-    shift, scale = weight.chunk(2, dim=0)
-    new_weight = torch.cat([scale, shift], dim=0)
+    """
+    Swap the scale and shift components in the weight tensor.
+
+    Args:
+        weight (torch.Tensor): The original weight tensor.
+        dim (int): The dimension along which to split.
+
+    Returns:
+        torch.Tensor: The modified weight tensor with scale and shift swapped.
+    """
+    shift, scale = weight.chunk(2, dim=dim)
+    new_weight = torch.cat([scale, shift], dim=dim)
     return new_weight
 
 
@@ -200,6 +210,7 @@ def main(args):
             "norm_num_groups": 32,
             "sample_size": 1024,
             "scaling_factor": 1.0,
+            "shift_factor": 0.0,
             "force_upcast": True,
             "use_quant_conv": False,
             "use_post_quant_conv": False,

scripts/convert_cogview4_to_diffusers_megatron.py

@@ -25,9 +25,15 @@ import argparse
 
 import torch
 from tqdm import tqdm
-from transformers import GlmForCausalLM, PreTrainedTokenizerFast
+from transformers import GlmModel, PreTrainedTokenizerFast
 
-from diffusers import AutoencoderKL, CogView4Pipeline, CogView4Transformer2DModel, FlowMatchEulerDiscreteScheduler
+from diffusers import (
+    AutoencoderKL,
+    CogView4ControlPipeline,
+    CogView4Pipeline,
+    CogView4Transformer2DModel,
+    FlowMatchEulerDiscreteScheduler,
+)
 from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
 
 
@@ -112,6 +118,12 @@ parser.add_argument(
     default=128,
     help="Maximum size for positional embeddings.",
 )
+parser.add_argument(
+    "--control",
+    action="store_true",
+    default=False,
+    help="Whether to use control model.",
+)
 
 args = parser.parse_args()
 
@@ -150,13 +162,15 @@ def convert_megatron_transformer_checkpoint_to_diffusers(
     Returns:
         dict: The converted state dictionary compatible with Diffusers.
     """
-    ckpt = torch.load(ckpt_path, map_location="cpu")
+    ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
     mega = ckpt["model"]
 
     new_state_dict = {}
 
     # Patch Embedding
-    new_state_dict["patch_embed.proj.weight"] = mega["encoder_expand_linear.weight"].reshape(hidden_size, 64)
+    new_state_dict["patch_embed.proj.weight"] = mega["encoder_expand_linear.weight"].reshape(
+        hidden_size, 128 if args.control else 64
+    )
     new_state_dict["patch_embed.proj.bias"] = mega["encoder_expand_linear.bias"]
     new_state_dict["patch_embed.text_proj.weight"] = mega["text_projector.weight"]
     new_state_dict["patch_embed.text_proj.bias"] = mega["text_projector.bias"]
@@ -189,14 +203,8 @@ def convert_megatron_transformer_checkpoint_to_diffusers(
         block_prefix = f"transformer_blocks.{i}."
 
         # AdaLayerNorm
-        new_state_dict[block_prefix + "norm1.linear.weight"] = swap_scale_shift(
-            mega[f"decoder.layers.{i}.adaln.weight"], dim=0
-        )
-        new_state_dict[block_prefix + "norm1.linear.bias"] = swap_scale_shift(
-            mega[f"decoder.layers.{i}.adaln.bias"], dim=0
-        )
-
-        # QKV
+        new_state_dict[block_prefix + "norm1.linear.weight"] = mega[f"decoder.layers.{i}.adaln.weight"]
+        new_state_dict[block_prefix + "norm1.linear.bias"] = mega[f"decoder.layers.{i}.adaln.bias"]
         qkv_weight = mega[f"decoder.layers.{i}.self_attention.linear_qkv.weight"]
         qkv_bias = mega[f"decoder.layers.{i}.self_attention.linear_qkv.bias"]
 
@@ -221,7 +229,7 @@ def convert_megatron_transformer_checkpoint_to_diffusers(
         # Attention Output
         new_state_dict[block_prefix + "attn1.to_out.0.weight"] = mega[
             f"decoder.layers.{i}.self_attention.linear_proj.weight"
-        ].T
+        ]
         new_state_dict[block_prefix + "attn1.to_out.0.bias"] = mega[
             f"decoder.layers.{i}.self_attention.linear_proj.bias"
         ]
@@ -252,7 +260,7 @@ def convert_cogview4_vae_checkpoint_to_diffusers(ckpt_path, vae_config):
     Returns:
         dict: The converted VAE state dictionary compatible with Diffusers.
     """
-    original_state_dict = torch.load(ckpt_path, map_location="cpu")["state_dict"]
+    original_state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=False)["state_dict"]
     return convert_ldm_vae_checkpoint(original_state_dict, vae_config)
 
 
@@ -286,7 +294,7 @@ def main(args):
         )
         transformer = CogView4Transformer2DModel(
             patch_size=2,
-            in_channels=16,
+            in_channels=32 if args.control else 16,
             num_layers=args.num_layers,
             attention_head_dim=args.attention_head_dim,
             num_attention_heads=args.num_heads,
@@ -317,6 +325,7 @@ def main(args):
             "norm_num_groups": 32,
             "sample_size": 1024,
             "scaling_factor": 1.0,
+            "shift_factor": 0.0,
             "force_upcast": True,
             "use_quant_conv": False,
             "use_post_quant_conv": False,
@@ -331,7 +340,7 @@ def main(args):
     # Load the text encoder and tokenizer
     text_encoder_id = "THUDM/glm-4-9b-hf"
     tokenizer = PreTrainedTokenizerFast.from_pretrained(text_encoder_id)
-    text_encoder = GlmForCausalLM.from_pretrained(
+    text_encoder = GlmModel.from_pretrained(
         text_encoder_id,
         cache_dir=args.text_encoder_cache_dir,
         torch_dtype=torch.bfloat16 if args.dtype == "bf16" else torch.float32,
@@ -345,13 +354,22 @@ def main(args):
     )
 
     # Create the pipeline
-    pipe = CogView4Pipeline(
-        tokenizer=tokenizer,
-        text_encoder=text_encoder,
-        vae=vae,
-        transformer=transformer,
-        scheduler=scheduler,
-    )
+    if args.control:
+        pipe = CogView4ControlPipeline(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            vae=vae,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+    else:
+        pipe = CogView4Pipeline(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            vae=vae,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
 
     # Save the converted pipeline
     pipe.save_pretrained(

scripts/convert_hunyuan_video_to_diffusers.py

@@ -160,8 +160,9 @@ TRANSFORMER_CONFIGS = {
         "pooled_projection_dim": 768,
         "rope_theta": 256.0,
         "rope_axes_dim": (16, 56, 56),
+        "image_condition_type": None,
     },
-    "HYVideo-T/2-I2V": {
+    "HYVideo-T/2-I2V-33ch": {
         "in_channels": 16 * 2 + 1,
         "out_channels": 16,
         "num_attention_heads": 24,
@@ -178,6 +179,26 @@ TRANSFORMER_CONFIGS = {
         "pooled_projection_dim": 768,
         "rope_theta": 256.0,
         "rope_axes_dim": (16, 56, 56),
+        "image_condition_type": "latent_concat",
+    },
+    "HYVideo-T/2-I2V-16ch": {
+        "in_channels": 16,
+        "out_channels": 16,
+        "num_attention_heads": 24,
+        "attention_head_dim": 128,
+        "num_layers": 20,
+        "num_single_layers": 40,
+        "num_refiner_layers": 2,
+        "mlp_ratio": 4.0,
+        "patch_size": 2,
+        "patch_size_t": 1,
+        "qk_norm": "rms_norm",
+        "guidance_embeds": True,
+        "text_embed_dim": 4096,
+        "pooled_projection_dim": 768,
+        "rope_theta": 256.0,
+        "rope_axes_dim": (16, 56, 56),
+        "image_condition_type": "token_replace",
     },
 }
 

scripts/convert_ltx_to_diffusers.py

@@ -74,6 +74,32 @@ VAE_091_RENAME_DICT = {
     "last_scale_shift_table": "scale_shift_table",
 }
 
+VAE_095_RENAME_DICT = {
+    # decoder
+    "up_blocks.0": "mid_block",
+    "up_blocks.1": "up_blocks.0.upsamplers.0",
+    "up_blocks.2": "up_blocks.0",
+    "up_blocks.3": "up_blocks.1.upsamplers.0",
+    "up_blocks.4": "up_blocks.1",
+    "up_blocks.5": "up_blocks.2.upsamplers.0",
+    "up_blocks.6": "up_blocks.2",
+    "up_blocks.7": "up_blocks.3.upsamplers.0",
+    "up_blocks.8": "up_blocks.3",
+    # encoder
+    "down_blocks.0": "down_blocks.0",
+    "down_blocks.1": "down_blocks.0.downsamplers.0",
+    "down_blocks.2": "down_blocks.1",
+    "down_blocks.3": "down_blocks.1.downsamplers.0",
+    "down_blocks.4": "down_blocks.2",
+    "down_blocks.5": "down_blocks.2.downsamplers.0",
+    "down_blocks.6": "down_blocks.3",
+    "down_blocks.7": "down_blocks.3.downsamplers.0",
+    "down_blocks.8": "mid_block",
+    # common
+    "last_time_embedder": "time_embedder",
+    "last_scale_shift_table": "scale_shift_table",
+}
+
 VAE_SPECIAL_KEYS_REMAP = {
     "per_channel_statistics.channel": remove_keys_,
     "per_channel_statistics.mean-of-means": remove_keys_,
@@ -81,10 +107,6 @@ VAE_SPECIAL_KEYS_REMAP = {
     "model.diffusion_model": remove_keys_,
 }
 
-VAE_091_SPECIAL_KEYS_REMAP = {
-    "timestep_scale_multiplier": remove_keys_,
-}
-
 
 def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
     state_dict = saved_dict
@@ -104,12 +126,16 @@ def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key:
 def convert_transformer(
     ckpt_path: str,
     dtype: torch.dtype,
+    version: str = "0.9.0",
 ):
     PREFIX_KEY = "model.diffusion_model."
 
     original_state_dict = get_state_dict(load_file(ckpt_path))
+    config = {}
+    if version == "0.9.5":
+        config["_use_causal_rope_fix"] = True
     with init_empty_weights():
-        transformer = LTXVideoTransformer3DModel()
+        transformer = LTXVideoTransformer3DModel(**config)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -161,12 +187,19 @@ def get_vae_config(version: str) -> Dict[str, Any]:
             "out_channels": 3,
             "latent_channels": 128,
             "block_out_channels": (128, 256, 512, 512),
+            "down_block_types": (
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+            ),
             "decoder_block_out_channels": (128, 256, 512, 512),
             "layers_per_block": (4, 3, 3, 3, 4),
             "decoder_layers_per_block": (4, 3, 3, 3, 4),
             "spatio_temporal_scaling": (True, True, True, False),
             "decoder_spatio_temporal_scaling": (True, True, True, False),
             "decoder_inject_noise": (False, False, False, False, False),
+            "downsample_type": ("conv", "conv", "conv", "conv"),
             "upsample_residual": (False, False, False, False),
             "upsample_factor": (1, 1, 1, 1),
             "patch_size": 4,
@@ -183,12 +216,19 @@ def get_vae_config(version: str) -> Dict[str, Any]:
             "out_channels": 3,
             "latent_channels": 128,
             "block_out_channels": (128, 256, 512, 512),
+            "down_block_types": (
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+                "LTXVideoDownBlock3D",
+            ),
             "decoder_block_out_channels": (256, 512, 1024),
             "layers_per_block": (4, 3, 3, 3, 4),
             "decoder_layers_per_block": (5, 6, 7, 8),
             "spatio_temporal_scaling": (True, True, True, False),
             "decoder_spatio_temporal_scaling": (True, True, True),
             "decoder_inject_noise": (True, True, True, False),
+            "downsample_type": ("conv", "conv", "conv", "conv"),
             "upsample_residual": (True, True, True),
             "upsample_factor": (2, 2, 2),
             "timestep_conditioning": True,
@@ -200,7 +240,38 @@ def get_vae_config(version: str) -> Dict[str, Any]:
             "decoder_causal": False,
         }
         VAE_KEYS_RENAME_DICT.update(VAE_091_RENAME_DICT)
-        VAE_SPECIAL_KEYS_REMAP.update(VAE_091_SPECIAL_KEYS_REMAP)
+    elif version == "0.9.5":
+        config = {
+            "in_channels": 3,
+            "out_channels": 3,
+            "latent_channels": 128,
+            "block_out_channels": (128, 256, 512, 1024, 2048),
+            "down_block_types": (
+                "LTXVideo095DownBlock3D",
+                "LTXVideo095DownBlock3D",
+                "LTXVideo095DownBlock3D",
+                "LTXVideo095DownBlock3D",
+            ),
+            "decoder_block_out_channels": (256, 512, 1024),
+            "layers_per_block": (4, 6, 6, 2, 2),
+            "decoder_layers_per_block": (5, 5, 5, 5),
+            "spatio_temporal_scaling": (True, True, True, True),
+            "decoder_spatio_temporal_scaling": (True, True, True),
+            "decoder_inject_noise": (False, False, False, False),
+            "downsample_type": ("spatial", "temporal", "spatiotemporal", "spatiotemporal"),
+            "upsample_residual": (True, True, True),
+            "upsample_factor": (2, 2, 2),
+            "timestep_conditioning": True,
+            "patch_size": 4,
+            "patch_size_t": 1,
+            "resnet_norm_eps": 1e-6,
+            "scaling_factor": 1.0,
+            "encoder_causal": True,
+            "decoder_causal": False,
+            "spatial_compression_ratio": 32,
+            "temporal_compression_ratio": 8,
+        }
+        VAE_KEYS_RENAME_DICT.update(VAE_095_RENAME_DICT)
     return config
 
 
@@ -223,7 +294,7 @@ def get_args():
     parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
     parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
     parser.add_argument(
-        "--version", type=str, default="0.9.0", choices=["0.9.0", "0.9.1"], help="Version of the LTX model"
+        "--version", type=str, default="0.9.0", choices=["0.9.0", "0.9.1", "0.9.5"], help="Version of the LTX model"
     )
     return parser.parse_args()
 
@@ -277,14 +348,17 @@ if __name__ == "__main__":
         for param in text_encoder.parameters():
             param.data = param.data.contiguous()
 
-        scheduler = FlowMatchEulerDiscreteScheduler(
-            use_dynamic_shifting=True,
-            base_shift=0.95,
-            max_shift=2.05,
-            base_image_seq_len=1024,
-            max_image_seq_len=4096,
-            shift_terminal=0.1,
-        )
+        if args.version == "0.9.5":
+            scheduler = FlowMatchEulerDiscreteScheduler(use_dynamic_shifting=False)
+        else:
+            scheduler = FlowMatchEulerDiscreteScheduler(
+                use_dynamic_shifting=True,
+                base_shift=0.95,
+                max_shift=2.05,
+                base_image_seq_len=1024,
+                max_image_seq_len=4096,
+                shift_terminal=0.1,
+            )
 
         pipe = LTXPipeline(
             scheduler=scheduler,

scripts/convert_lumina_to_diffusers.py

@@ -5,7 +5,7 @@ import torch
 from safetensors.torch import load_file
 from transformers import AutoModel, AutoTokenizer
 
-from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, LuminaNextDiT2DModel, LuminaText2ImgPipeline
+from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, LuminaNextDiT2DModel, LuminaPipeline
 
 
 def main(args):
@@ -115,7 +115,7 @@ def main(args):
         tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
         text_encoder = AutoModel.from_pretrained("google/gemma-2b")
 
-        pipeline = LuminaText2ImgPipeline(
+        pipeline = LuminaPipeline(
             tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, scheduler=scheduler
         )
         pipeline.save_pretrained(args.dump_path)

scripts/convert_sana_to_diffusers.py

@@ -16,7 +16,9 @@ from diffusers import (
     DPMSolverMultistepScheduler,
     FlowMatchEulerDiscreteScheduler,
     SanaPipeline,
+    SanaSprintPipeline,
     SanaTransformer2DModel,
+    SCMScheduler,
 )
 from diffusers.models.modeling_utils import load_model_dict_into_meta
 from diffusers.utils.import_utils import is_accelerate_available
@@ -25,6 +27,10 @@ from diffusers.utils.import_utils import is_accelerate_available
 CTX = init_empty_weights if is_accelerate_available else nullcontext
 
 ckpt_ids = [
+    "Efficient-Large-Model/Sana_Sprint_0.6B_1024px/checkpoints/Sana_Sprint_0.6B_1024px.pth"
+    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px/checkpoints/Sana_Sprint_1.6B_1024px.pth"
+    "Efficient-Large-Model/SANA1.5_4.8B_1024px/checkpoints/SANA1.5_4.8B_1024px.pth",
+    "Efficient-Large-Model/SANA1.5_1.6B_1024px/checkpoints/SANA1.5_1.6B_1024px.pth",
     "Efficient-Large-Model/Sana_1600M_4Kpx_BF16/checkpoints/Sana_1600M_4Kpx_BF16.pth",
     "Efficient-Large-Model/Sana_1600M_2Kpx_BF16/checkpoints/Sana_1600M_2Kpx_BF16.pth",
     "Efficient-Large-Model/Sana_1600M_1024px_MultiLing/checkpoints/Sana_1600M_1024px_MultiLing.pth",
@@ -72,15 +78,42 @@ def main(args):
     converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
     converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
 
-    # AdaLN-single LN
-    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
-        "t_embedder.mlp.0.weight"
-    )
-    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
-    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
-        "t_embedder.mlp.2.weight"
-    )
-    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+    # Handle different time embedding structure based on model type
+
+    if args.model_type in ["SanaSprint_1600M_P1_D20", "SanaSprint_600M_P1_D28"]:
+        # For Sana Sprint, the time embedding structure is different
+        converted_state_dict["time_embed.timestep_embedder.linear_1.weight"] = state_dict.pop(
+            "t_embedder.mlp.0.weight"
+        )
+        converted_state_dict["time_embed.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+        converted_state_dict["time_embed.timestep_embedder.linear_2.weight"] = state_dict.pop(
+            "t_embedder.mlp.2.weight"
+        )
+        converted_state_dict["time_embed.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+        # Guidance embedder for Sana Sprint
+        converted_state_dict["time_embed.guidance_embedder.linear_1.weight"] = state_dict.pop(
+            "cfg_embedder.mlp.0.weight"
+        )
+        converted_state_dict["time_embed.guidance_embedder.linear_1.bias"] = state_dict.pop("cfg_embedder.mlp.0.bias")
+        converted_state_dict["time_embed.guidance_embedder.linear_2.weight"] = state_dict.pop(
+            "cfg_embedder.mlp.2.weight"
+        )
+        converted_state_dict["time_embed.guidance_embedder.linear_2.bias"] = state_dict.pop("cfg_embedder.mlp.2.bias")
+    else:
+        # Original Sana time embedding structure
+        converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+            "t_embedder.mlp.0.weight"
+        )
+        converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop(
+            "t_embedder.mlp.0.bias"
+        )
+        converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+            "t_embedder.mlp.2.weight"
+        )
+        converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop(
+            "t_embedder.mlp.2.bias"
+        )
 
     # Shared norm.
     converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
@@ -96,14 +129,22 @@ def main(args):
         flow_shift = 3.0
 
     # model config
-    if args.model_type == "SanaMS_1600M_P1_D20":
+    if args.model_type in ["SanaMS_1600M_P1_D20", "SanaSprint_1600M_P1_D20", "SanaMS1.5_1600M_P1_D20"]:
         layer_num = 20
-    elif args.model_type == "SanaMS_600M_P1_D28":
+    elif args.model_type in ["SanaMS_600M_P1_D28", "SanaSprint_600M_P1_D28"]:
         layer_num = 28
+    elif args.model_type == "SanaMS_4800M_P1_D60":
+        layer_num = 60
     else:
         raise ValueError(f"{args.model_type} is not supported.")
     # Positional embedding interpolation scale.
     interpolation_scale = {512: None, 1024: None, 2048: 1.0, 4096: 2.0}
+    qk_norm = (
+        "rms_norm_across_heads"
+        if args.model_type
+        in ["SanaMS1.5_1600M_P1_D20", "SanaMS1.5_4800M_P1_D60", "SanaSprint_600M_P1_D28", "SanaSprint_1600M_P1_D20"]
+        else None
+    )
 
     for depth in range(layer_num):
         # Transformer blocks.
@@ -117,6 +158,14 @@ def main(args):
         converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
         converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
         converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        if qk_norm is not None:
+            # Add Q/K normalization for self-attention (attn1) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.k_norm.weight"
+            )
         # Projection.
         converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
             f"blocks.{depth}.attn.proj.weight"
@@ -154,6 +203,14 @@ def main(args):
         converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
         converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
         converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+        if qk_norm is not None:
+            # Add Q/K normalization for cross-attention (attn2) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.k_norm.weight"
+            )
 
         converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
             f"blocks.{depth}.cross_attn.proj.weight"
@@ -169,24 +226,37 @@ def main(args):
 
     # Transformer
     with CTX():
-        transformer = SanaTransformer2DModel(
-            in_channels=32,
-            out_channels=32,
-            num_attention_heads=model_kwargs[args.model_type]["num_attention_heads"],
-            attention_head_dim=model_kwargs[args.model_type]["attention_head_dim"],
-            num_layers=model_kwargs[args.model_type]["num_layers"],
-            num_cross_attention_heads=model_kwargs[args.model_type]["num_cross_attention_heads"],
-            cross_attention_head_dim=model_kwargs[args.model_type]["cross_attention_head_dim"],
-            cross_attention_dim=model_kwargs[args.model_type]["cross_attention_dim"],
-            caption_channels=2304,
-            mlp_ratio=2.5,
-            attention_bias=False,
-            sample_size=args.image_size // 32,
-            patch_size=1,
-            norm_elementwise_affine=False,
-            norm_eps=1e-6,
-            interpolation_scale=interpolation_scale[args.image_size],
-        )
+        transformer_kwargs = {
+            "in_channels": 32,
+            "out_channels": 32,
+            "num_attention_heads": model_kwargs[args.model_type]["num_attention_heads"],
+            "attention_head_dim": model_kwargs[args.model_type]["attention_head_dim"],
+            "num_layers": model_kwargs[args.model_type]["num_layers"],
+            "num_cross_attention_heads": model_kwargs[args.model_type]["num_cross_attention_heads"],
+            "cross_attention_head_dim": model_kwargs[args.model_type]["cross_attention_head_dim"],
+            "cross_attention_dim": model_kwargs[args.model_type]["cross_attention_dim"],
+            "caption_channels": 2304,
+            "mlp_ratio": 2.5,
+            "attention_bias": False,
+            "sample_size": args.image_size // 32,
+            "patch_size": 1,
+            "norm_elementwise_affine": False,
+            "norm_eps": 1e-6,
+            "interpolation_scale": interpolation_scale[args.image_size],
+        }
+
+        # Add qk_norm parameter for Sana Sprint
+        if args.model_type in [
+            "SanaMS1.5_1600M_P1_D20",
+            "SanaMS1.5_4800M_P1_D60",
+            "SanaSprint_600M_P1_D28",
+            "SanaSprint_1600M_P1_D20",
+        ]:
+            transformer_kwargs["qk_norm"] = "rms_norm_across_heads"
+        if args.model_type in ["SanaSprint_1600M_P1_D20", "SanaSprint_600M_P1_D28"]:
+            transformer_kwargs["guidance_embeds"] = True
+
+        transformer = SanaTransformer2DModel(**transformer_kwargs)
 
     if is_accelerate_available():
         load_model_dict_into_meta(transformer, converted_state_dict)
@@ -196,6 +266,8 @@ def main(args):
     try:
         state_dict.pop("y_embedder.y_embedding")
         state_dict.pop("pos_embed")
+        state_dict.pop("logvar_linear.weight")
+        state_dict.pop("logvar_linear.bias")
     except KeyError:
         print("y_embedder.y_embedding or pos_embed not found in the state_dict")
 
@@ -210,47 +282,74 @@ def main(args):
         print(
             colored(
                 f"Only saving transformer model of {args.model_type}. "
-                f"Set --save_full_pipeline to save the whole SanaPipeline",
+                f"Set --save_full_pipeline to save the whole Pipeline",
                 "green",
                 attrs=["bold"],
             )
         )
         transformer.save_pretrained(
-            os.path.join(args.dump_path, "transformer"), safe_serialization=True, max_shard_size="5GB", variant=variant
+            os.path.join(args.dump_path, "transformer"), safe_serialization=True, max_shard_size="5GB"
         )
     else:
-        print(colored(f"Saving the whole SanaPipeline containing {args.model_type}", "green", attrs=["bold"]))
+        print(colored(f"Saving the whole Pipeline containing {args.model_type}", "green", attrs=["bold"]))
         # VAE
-        ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", torch_dtype=torch.float32)
+        ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers", torch_dtype=torch.float32)
 
         # Text Encoder
-        text_encoder_model_path = "google/gemma-2-2b-it"
+        text_encoder_model_path = "Efficient-Large-Model/gemma-2-2b-it"
         tokenizer = AutoTokenizer.from_pretrained(text_encoder_model_path)
         tokenizer.padding_side = "right"
         text_encoder = AutoModelForCausalLM.from_pretrained(
             text_encoder_model_path, torch_dtype=torch.bfloat16
         ).get_decoder()
 
-        # Scheduler
-        if args.scheduler_type == "flow-dpm_solver":
-            scheduler = DPMSolverMultistepScheduler(
-                flow_shift=flow_shift,
-                use_flow_sigmas=True,
-                prediction_type="flow_prediction",
-            )
-        elif args.scheduler_type == "flow-euler":
-            scheduler = FlowMatchEulerDiscreteScheduler(shift=flow_shift)
-        else:
-            raise ValueError(f"Scheduler type {args.scheduler_type} is not supported")
+        # Choose the appropriate pipeline and scheduler based on model type
+        if args.model_type in ["SanaSprint_1600M_P1_D20", "SanaSprint_600M_P1_D28"]:
+            # Force SCM Scheduler for Sana Sprint regardless of scheduler_type
+            if args.scheduler_type != "scm":
+                print(
+                    colored(
+                        f"Warning: Overriding scheduler_type '{args.scheduler_type}' to 'scm' for SanaSprint model",
+                        "yellow",
+                        attrs=["bold"],
+                    )
+                )
 
-        pipe = SanaPipeline(
-            tokenizer=tokenizer,
-            text_encoder=text_encoder,
-            transformer=transformer,
-            vae=ae,
-            scheduler=scheduler,
-        )
-        pipe.save_pretrained(args.dump_path, safe_serialization=True, max_shard_size="5GB", variant=variant)
+            # SCM Scheduler for Sana Sprint
+            scheduler_config = {
+                "prediction_type": "trigflow",
+                "sigma_data": 0.5,
+            }
+            scheduler = SCMScheduler(**scheduler_config)
+            pipe = SanaSprintPipeline(
+                tokenizer=tokenizer,
+                text_encoder=text_encoder,
+                transformer=transformer,
+                vae=ae,
+                scheduler=scheduler,
+            )
+        else:
+            # Original Sana scheduler
+            if args.scheduler_type == "flow-dpm_solver":
+                scheduler = DPMSolverMultistepScheduler(
+                    flow_shift=flow_shift,
+                    use_flow_sigmas=True,
+                    prediction_type="flow_prediction",
+                )
+            elif args.scheduler_type == "flow-euler":
+                scheduler = FlowMatchEulerDiscreteScheduler(shift=flow_shift)
+            else:
+                raise ValueError(f"Scheduler type {args.scheduler_type} is not supported")
+
+            pipe = SanaPipeline(
+                tokenizer=tokenizer,
+                text_encoder=text_encoder,
+                transformer=transformer,
+                vae=ae,
+                scheduler=scheduler,
+            )
+
+        pipe.save_pretrained(args.dump_path, safe_serialization=True, max_shard_size="5GB")
 
 
 DTYPE_MAPPING = {
@@ -259,12 +358,6 @@ DTYPE_MAPPING = {
     "bf16": torch.bfloat16,
 }
 
-VARIANT_MAPPING = {
-    "fp32": None,
-    "fp16": "fp16",
-    "bf16": "bf16",
-}
-
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
@@ -281,10 +374,24 @@ if __name__ == "__main__":
         help="Image size of pretrained model, 512, 1024, 2048 or 4096.",
     )
     parser.add_argument(
-        "--model_type", default="SanaMS_1600M_P1_D20", type=str, choices=["SanaMS_1600M_P1_D20", "SanaMS_600M_P1_D28"]
+        "--model_type",
+        default="SanaMS_1600M_P1_D20",
+        type=str,
+        choices=[
+            "SanaMS_1600M_P1_D20",
+            "SanaMS_600M_P1_D28",
+            "SanaMS1.5_1600M_P1_D20",
+            "SanaMS1.5_4800M_P1_D60",
+            "SanaSprint_1600M_P1_D20",
+            "SanaSprint_600M_P1_D28",
+        ],
     )
     parser.add_argument(
-        "--scheduler_type", default="flow-dpm_solver", type=str, choices=["flow-dpm_solver", "flow-euler"]
+        "--scheduler_type",
+        default="flow-dpm_solver",
+        type=str,
+        choices=["flow-dpm_solver", "flow-euler", "scm"],
+        help="Scheduler type to use. Use 'scm' for Sana Sprint models.",
     )
     parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
     parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipelien elemets in one.")
@@ -309,10 +416,41 @@ if __name__ == "__main__":
             "cross_attention_dim": 1152,
             "num_layers": 28,
         },
+        "SanaMS1.5_1600M_P1_D20": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 20,
+        },
+        "SanaMS1.5_4800M_P1_D60": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 60,
+        },
+        "SanaSprint_600M_P1_D28": {
+            "num_attention_heads": 36,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 16,
+            "cross_attention_head_dim": 72,
+            "cross_attention_dim": 1152,
+            "num_layers": 28,
+        },
+        "SanaSprint_1600M_P1_D20": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 20,
+        },
     }
 
     device = "cuda" if torch.cuda.is_available() else "cpu"
     weight_dtype = DTYPE_MAPPING[args.dtype]
-    variant = VARIANT_MAPPING[args.dtype]
 
     main(args)

src/diffusers/__init__.py

@@ -131,8 +131,10 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["hooks"].extend(
         [
+            "FasterCacheConfig",
             "HookRegistry",
             "PyramidAttentionBroadcastConfig",
+            "apply_faster_cache",
             "apply_pyramid_attention_broadcast",
         ]
     )
@@ -271,6 +273,7 @@ else:
             "RePaintScheduler",
             "SASolverScheduler",
             "SchedulerMixin",
+            "SCMScheduler",
             "ScoreSdeVeScheduler",
             "TCDScheduler",
             "UnCLIPScheduler",
@@ -345,6 +348,7 @@ else:
             "CogVideoXPipeline",
             "CogVideoXVideoToVideoPipeline",
             "CogView3PlusPipeline",
+            "CogView4ControlPipeline",
             "CogView4Pipeline",
             "ConsisIDPipeline",
             "CycleDiffusionPipeline",
@@ -401,9 +405,12 @@ else:
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
+            "LTXConditionPipeline",
             "LTXImageToVideoPipeline",
             "LTXPipeline",
+            "Lumina2Pipeline",
             "Lumina2Text2ImgPipeline",
+            "LuminaPipeline",
             "LuminaText2ImgPipeline",
             "MarigoldDepthPipeline",
             "MarigoldIntrinsicsPipeline",
@@ -419,6 +426,7 @@ else:
             "ReduxImageEncoder",
             "SanaPAGPipeline",
             "SanaPipeline",
+            "SanaSprintPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -699,7 +707,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from .utils.dummy_pt_objects import *  # noqa F403
     else:
-        from .hooks import HookRegistry, PyramidAttentionBroadcastConfig, apply_pyramid_attention_broadcast
+        from .hooks import (
+            FasterCacheConfig,
+            HookRegistry,
+            PyramidAttentionBroadcastConfig,
+            apply_faster_cache,
+            apply_pyramid_attention_broadcast,
+        )
         from .models import (
             AllegroTransformer3DModel,
             AsymmetricAutoencoderKL,
@@ -832,6 +846,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             RePaintScheduler,
             SASolverScheduler,
             SchedulerMixin,
+            SCMScheduler,
             ScoreSdeVeScheduler,
             TCDScheduler,
             UnCLIPScheduler,
@@ -887,6 +902,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogVideoXPipeline,
             CogVideoXVideoToVideoPipeline,
             CogView3PlusPipeline,
+            CogView4ControlPipeline,
             CogView4Pipeline,
             ConsisIDPipeline,
             CycleDiffusionPipeline,
@@ -943,9 +959,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            LTXConditionPipeline,
             LTXImageToVideoPipeline,
             LTXPipeline,
+            Lumina2Pipeline,
             Lumina2Text2ImgPipeline,
+            LuminaPipeline,
             LuminaText2ImgPipeline,
             MarigoldDepthPipeline,
             MarigoldIntrinsicsPipeline,
@@ -961,6 +980,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ReduxImageEncoder,
             SanaPAGPipeline,
             SanaPipeline,
+            SanaSprintPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,

src/diffusers/hooks/__init__.py

@@ -2,6 +2,7 @@ from ..utils import is_torch_available
 
 
 if is_torch_available():
+    from .faster_cache import FasterCacheConfig, apply_faster_cache
     from .group_offloading import apply_group_offloading
     from .hooks import HookRegistry, ModelHook
     from .layerwise_casting import apply_layerwise_casting, apply_layerwise_casting_hook

src/diffusers/hooks/faster_cache.py

@@ -0,0 +1,653 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from dataclasses import dataclass
+from typing import Any, Callable, List, Optional, Tuple
+
+import torch
+
+from ..models.attention_processor import Attention, MochiAttention
+from ..models.modeling_outputs import Transformer2DModelOutput
+from ..utils import logging
+from .hooks import HookRegistry, ModelHook
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+_FASTER_CACHE_DENOISER_HOOK = "faster_cache_denoiser"
+_FASTER_CACHE_BLOCK_HOOK = "faster_cache_block"
+_ATTENTION_CLASSES = (Attention, MochiAttention)
+_SPATIAL_ATTENTION_BLOCK_IDENTIFIERS = (
+    "^blocks.*attn",
+    "^transformer_blocks.*attn",
+    "^single_transformer_blocks.*attn",
+)
+_TEMPORAL_ATTENTION_BLOCK_IDENTIFIERS = ("^temporal_transformer_blocks.*attn",)
+_TRANSFORMER_BLOCK_IDENTIFIERS = _SPATIAL_ATTENTION_BLOCK_IDENTIFIERS + _TEMPORAL_ATTENTION_BLOCK_IDENTIFIERS
+_UNCOND_COND_INPUT_KWARGS_IDENTIFIERS = (
+    "hidden_states",
+    "encoder_hidden_states",
+    "timestep",
+    "attention_mask",
+    "encoder_attention_mask",
+)
+
+
+@dataclass
+class FasterCacheConfig:
+    r"""
+    Configuration for [FasterCache](https://huggingface.co/papers/2410.19355).
+
+    Attributes:
+        spatial_attention_block_skip_range (`int`, defaults to `2`):
+            Calculate the attention states every `N` iterations. If this is set to `N`, the attention computation will
+            be skipped `N - 1` times (i.e., cached attention states will be re-used) before computing the new attention
+            states again.
+        temporal_attention_block_skip_range (`int`, *optional*, defaults to `None`):
+            Calculate the attention states every `N` iterations. If this is set to `N`, the attention computation will
+            be skipped `N - 1` times (i.e., cached attention states will be re-used) before computing the new attention
+            states again.
+        spatial_attention_timestep_skip_range (`Tuple[float, float]`, defaults to `(-1, 681)`):
+            The timestep range within which the spatial attention computation can be skipped without a significant loss
+            in quality. This is to be determined by the user based on the underlying model. The first value in the
+            tuple is the lower bound and the second value is the upper bound. Typically, diffusion timesteps for
+            denoising are in the reversed range of 0 to 1000 (i.e. denoising starts at timestep 1000 and ends at
+            timestep 0). For the default values, this would mean that the spatial attention computation skipping will
+            be applicable only after denoising timestep 681 is reached, and continue until the end of the denoising
+            process.
+        temporal_attention_timestep_skip_range (`Tuple[float, float]`, *optional*, defaults to `None`):
+            The timestep range within which the temporal attention computation can be skipped without a significant
+            loss in quality. This is to be determined by the user based on the underlying model. The first value in the
+            tuple is the lower bound and the second value is the upper bound. Typically, diffusion timesteps for
+            denoising are in the reversed range of 0 to 1000 (i.e. denoising starts at timestep 1000 and ends at
+            timestep 0).
+        low_frequency_weight_update_timestep_range (`Tuple[int, int]`, defaults to `(99, 901)`):
+            The timestep range within which the low frequency weight scaling update is applied. The first value in the
+            tuple is the lower bound and the second value is the upper bound of the timestep range. The callback
+            function for the update is called only within this range.
+        high_frequency_weight_update_timestep_range (`Tuple[int, int]`, defaults to `(-1, 301)`):
+            The timestep range within which the high frequency weight scaling update is applied. The first value in the
+            tuple is the lower bound and the second value is the upper bound of the timestep range. The callback
+            function for the update is called only within this range.
+        alpha_low_frequency (`float`, defaults to `1.1`):
+            The weight to scale the low frequency updates by. This is used to approximate the unconditional branch from
+            the conditional branch outputs.
+        alpha_high_frequency (`float`, defaults to `1.1`):
+            The weight to scale the high frequency updates by. This is used to approximate the unconditional branch
+            from the conditional branch outputs.
+        unconditional_batch_skip_range (`int`, defaults to `5`):
+            Process the unconditional branch every `N` iterations. If this is set to `N`, the unconditional branch
+            computation will be skipped `N - 1` times (i.e., cached unconditional branch states will be re-used) before
+            computing the new unconditional branch states again.
+        unconditional_batch_timestep_skip_range (`Tuple[float, float]`, defaults to `(-1, 641)`):
+            The timestep range within which the unconditional branch computation can be skipped without a significant
+            loss in quality. This is to be determined by the user based on the underlying model. The first value in the
+            tuple is the lower bound and the second value is the upper bound.
+        spatial_attention_block_identifiers (`Tuple[str, ...]`, defaults to `("blocks.*attn1", "transformer_blocks.*attn1", "single_transformer_blocks.*attn1")`):
+            The identifiers to match the spatial attention blocks in the model. If the name of the block contains any
+            of these identifiers, FasterCache will be applied to that block. This can either be the full layer names,
+            partial layer names, or regex patterns. Matching will always be done using a regex match.
+        temporal_attention_block_identifiers (`Tuple[str, ...]`, defaults to `("temporal_transformer_blocks.*attn1",)`):
+            The identifiers to match the temporal attention blocks in the model. If the name of the block contains any
+            of these identifiers, FasterCache will be applied to that block. This can either be the full layer names,
+            partial layer names, or regex patterns. Matching will always be done using a regex match.
+        attention_weight_callback (`Callable[[torch.nn.Module], float]`, defaults to `None`):
+            The callback function to determine the weight to scale the attention outputs by. This function should take
+            the attention module as input and return a float value. This is used to approximate the unconditional
+            branch from the conditional branch outputs. If not provided, the default weight is 0.5 for all timesteps.
+            Typically, as described in the paper, this weight should gradually increase from 0 to 1 as the inference
+            progresses. Users are encouraged to experiment and provide custom weight schedules that take into account
+            the number of inference steps and underlying model behaviour as denoising progresses.
+        low_frequency_weight_callback (`Callable[[torch.nn.Module], float]`, defaults to `None`):
+            The callback function to determine the weight to scale the low frequency updates by. If not provided, the
+            default weight is 1.1 for timesteps within the range specified (as described in the paper).
+        high_frequency_weight_callback (`Callable[[torch.nn.Module], float]`, defaults to `None`):
+            The callback function to determine the weight to scale the high frequency updates by. If not provided, the
+            default weight is 1.1 for timesteps within the range specified (as described in the paper).
+        tensor_format (`str`, defaults to `"BCFHW"`):
+            The format of the input tensors. This should be one of `"BCFHW"`, `"BFCHW"`, or `"BCHW"`. The format is
+            used to split individual latent frames in order for low and high frequency components to be computed.
+        is_guidance_distilled (`bool`, defaults to `False`):
+            Whether the model is guidance distilled or not. If the model is guidance distilled, FasterCache will not be
+            applied at the denoiser-level to skip the unconditional branch computation (as there is none).
+        _unconditional_conditional_input_kwargs_identifiers (`List[str]`, defaults to `("hidden_states", "encoder_hidden_states", "timestep", "attention_mask", "encoder_attention_mask")`):
+            The identifiers to match the input kwargs that contain the batchwise-concatenated unconditional and
+            conditional inputs. If the name of the input kwargs contains any of these identifiers, FasterCache will
+            split the inputs into unconditional and conditional branches. This must be a list of exact input kwargs
+            names that contain the batchwise-concatenated unconditional and conditional inputs.
+    """
+
+    # In the paper and codebase, they hardcode these values to 2. However, it can be made configurable
+    # after some testing. We default to 2 if these parameters are not provided.
+    spatial_attention_block_skip_range: int = 2
+    temporal_attention_block_skip_range: Optional[int] = None
+
+    spatial_attention_timestep_skip_range: Tuple[int, int] = (-1, 681)
+    temporal_attention_timestep_skip_range: Tuple[int, int] = (-1, 681)
+
+    # Indicator functions for low/high frequency as mentioned in Equation 11 of the paper
+    low_frequency_weight_update_timestep_range: Tuple[int, int] = (99, 901)
+    high_frequency_weight_update_timestep_range: Tuple[int, int] = (-1, 301)
+
+    # ⍺1 and ⍺2 as mentioned in Equation 11 of the paper
+    alpha_low_frequency: float = 1.1
+    alpha_high_frequency: float = 1.1
+
+    # n as described in CFG-Cache explanation in the paper - dependant on the model
+    unconditional_batch_skip_range: int = 5
+    unconditional_batch_timestep_skip_range: Tuple[int, int] = (-1, 641)
+
+    spatial_attention_block_identifiers: Tuple[str, ...] = _SPATIAL_ATTENTION_BLOCK_IDENTIFIERS
+    temporal_attention_block_identifiers: Tuple[str, ...] = _TEMPORAL_ATTENTION_BLOCK_IDENTIFIERS
+
+    attention_weight_callback: Callable[[torch.nn.Module], float] = None
+    low_frequency_weight_callback: Callable[[torch.nn.Module], float] = None
+    high_frequency_weight_callback: Callable[[torch.nn.Module], float] = None
+
+    tensor_format: str = "BCFHW"
+    is_guidance_distilled: bool = False
+
+    current_timestep_callback: Callable[[], int] = None
+
+    _unconditional_conditional_input_kwargs_identifiers: List[str] = _UNCOND_COND_INPUT_KWARGS_IDENTIFIERS
+
+    def __repr__(self) -> str:
+        return (
+            f"FasterCacheConfig(\n"
+            f"  spatial_attention_block_skip_range={self.spatial_attention_block_skip_range},\n"
+            f"  temporal_attention_block_skip_range={self.temporal_attention_block_skip_range},\n"
+            f"  spatial_attention_timestep_skip_range={self.spatial_attention_timestep_skip_range},\n"
+            f"  temporal_attention_timestep_skip_range={self.temporal_attention_timestep_skip_range},\n"
+            f"  low_frequency_weight_update_timestep_range={self.low_frequency_weight_update_timestep_range},\n"
+            f"  high_frequency_weight_update_timestep_range={self.high_frequency_weight_update_timestep_range},\n"
+            f"  alpha_low_frequency={self.alpha_low_frequency},\n"
+            f"  alpha_high_frequency={self.alpha_high_frequency},\n"
+            f"  unconditional_batch_skip_range={self.unconditional_batch_skip_range},\n"
+            f"  unconditional_batch_timestep_skip_range={self.unconditional_batch_timestep_skip_range},\n"
+            f"  spatial_attention_block_identifiers={self.spatial_attention_block_identifiers},\n"
+            f"  temporal_attention_block_identifiers={self.temporal_attention_block_identifiers},\n"
+            f"  tensor_format={self.tensor_format},\n"
+            f")"
+        )
+
+
+class FasterCacheDenoiserState:
+    r"""
+    State for [FasterCache](https://huggingface.co/papers/2410.19355) top-level denoiser module.
+    """
+
+    def __init__(self) -> None:
+        self.iteration: int = 0
+        self.low_frequency_delta: torch.Tensor = None
+        self.high_frequency_delta: torch.Tensor = None
+
+    def reset(self):
+        self.iteration = 0
+        self.low_frequency_delta = None
+        self.high_frequency_delta = None
+
+
+class FasterCacheBlockState:
+    r"""
+    State for [FasterCache](https://huggingface.co/papers/2410.19355). Every underlying block that FasterCache is
+    applied to will have an instance of this state.
+    """
+
+    def __init__(self) -> None:
+        self.iteration: int = 0
+        self.batch_size: int = None
+        self.cache: Tuple[torch.Tensor, torch.Tensor] = None
+
+    def reset(self):
+        self.iteration = 0
+        self.batch_size = None
+        self.cache = None
+
+
+class FasterCacheDenoiserHook(ModelHook):
+    _is_stateful = True
+
+    def __init__(
+        self,
+        unconditional_batch_skip_range: int,
+        unconditional_batch_timestep_skip_range: Tuple[int, int],
+        tensor_format: str,
+        is_guidance_distilled: bool,
+        uncond_cond_input_kwargs_identifiers: List[str],
+        current_timestep_callback: Callable[[], int],
+        low_frequency_weight_callback: Callable[[torch.nn.Module], torch.Tensor],
+        high_frequency_weight_callback: Callable[[torch.nn.Module], torch.Tensor],
+    ) -> None:
+        super().__init__()
+
+        self.unconditional_batch_skip_range = unconditional_batch_skip_range
+        self.unconditional_batch_timestep_skip_range = unconditional_batch_timestep_skip_range
+        # We can't easily detect what args are to be split in unconditional and conditional branches. We
+        # can only do it for kwargs, hence they are the only ones we split. The args are passed as-is.
+        # If a model is to be made compatible with FasterCache, the user must ensure that the inputs that
+        # contain batchwise-concatenated unconditional and conditional inputs are passed as kwargs.
+        self.uncond_cond_input_kwargs_identifiers = uncond_cond_input_kwargs_identifiers
+        self.tensor_format = tensor_format
+        self.is_guidance_distilled = is_guidance_distilled
+
+        self.current_timestep_callback = current_timestep_callback
+        self.low_frequency_weight_callback = low_frequency_weight_callback
+        self.high_frequency_weight_callback = high_frequency_weight_callback
+
+    def initialize_hook(self, module):
+        self.state = FasterCacheDenoiserState()
+        return module
+
+    @staticmethod
+    def _get_cond_input(input: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Note: this method assumes that the input tensor is batchwise-concatenated with unconditional inputs
+        # followed by conditional inputs.
+        _, cond = input.chunk(2, dim=0)
+        return cond
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs) -> Any:
+        # Split the unconditional and conditional inputs. We only want to infer the conditional branch if the
+        # requirements for skipping the unconditional branch are met as described in the paper.
+        # We skip the unconditional branch only if the following conditions are met:
+        #   1. We have completed at least one iteration of the denoiser
+        #   2. The current timestep is within the range specified by the user. This is the optimal timestep range
+        #      where approximating the unconditional branch from the computation of the conditional branch is possible
+        #      without a significant loss in quality.
+        #   3. The current iteration is not a multiple of the unconditional batch skip range. This is done so that
+        #      we compute the unconditional branch at least once every few iterations to ensure minimal quality loss.
+        is_within_timestep_range = (
+            self.unconditional_batch_timestep_skip_range[0]
+            < self.current_timestep_callback()
+            < self.unconditional_batch_timestep_skip_range[1]
+        )
+        should_skip_uncond = (
+            self.state.iteration > 0
+            and is_within_timestep_range
+            and self.state.iteration % self.unconditional_batch_skip_range != 0
+            and not self.is_guidance_distilled
+        )
+
+        if should_skip_uncond:
+            is_any_kwarg_uncond = any(k in self.uncond_cond_input_kwargs_identifiers for k in kwargs.keys())
+            if is_any_kwarg_uncond:
+                logger.debug("FasterCache - Skipping unconditional branch computation")
+                args = tuple([self._get_cond_input(arg) if torch.is_tensor(arg) else arg for arg in args])
+                kwargs = {
+                    k: v if k not in self.uncond_cond_input_kwargs_identifiers else self._get_cond_input(v)
+                    for k, v in kwargs.items()
+                }
+
+        output = self.fn_ref.original_forward(*args, **kwargs)
+
+        if self.is_guidance_distilled:
+            self.state.iteration += 1
+            return output
+
+        if torch.is_tensor(output):
+            hidden_states = output
+        elif isinstance(output, (tuple, Transformer2DModelOutput)):
+            hidden_states = output[0]
+
+        batch_size = hidden_states.size(0)
+
+        if should_skip_uncond:
+            self.state.low_frequency_delta = self.state.low_frequency_delta * self.low_frequency_weight_callback(
+                module
+            )
+            self.state.high_frequency_delta = self.state.high_frequency_delta * self.high_frequency_weight_callback(
+                module
+            )
+
+            if self.tensor_format == "BCFHW":
+                hidden_states = hidden_states.permute(0, 2, 1, 3, 4)
+            if self.tensor_format == "BCFHW" or self.tensor_format == "BFCHW":
+                hidden_states = hidden_states.flatten(0, 1)
+
+            low_freq_cond, high_freq_cond = _split_low_high_freq(hidden_states.float())
+
+            # Approximate/compute the unconditional branch outputs as described in Equation 9 and 10 of the paper
+            low_freq_uncond = self.state.low_frequency_delta + low_freq_cond
+            high_freq_uncond = self.state.high_frequency_delta + high_freq_cond
+            uncond_freq = low_freq_uncond + high_freq_uncond
+
+            uncond_states = torch.fft.ifftshift(uncond_freq)
+            uncond_states = torch.fft.ifft2(uncond_states).real
+
+            if self.tensor_format == "BCFHW" or self.tensor_format == "BFCHW":
+                uncond_states = uncond_states.unflatten(0, (batch_size, -1))
+                hidden_states = hidden_states.unflatten(0, (batch_size, -1))
+            if self.tensor_format == "BCFHW":
+                uncond_states = uncond_states.permute(0, 2, 1, 3, 4)
+                hidden_states = hidden_states.permute(0, 2, 1, 3, 4)
+
+            # Concatenate the approximated unconditional and predicted conditional branches
+            uncond_states = uncond_states.to(hidden_states.dtype)
+            hidden_states = torch.cat([uncond_states, hidden_states], dim=0)
+        else:
+            uncond_states, cond_states = hidden_states.chunk(2, dim=0)
+            if self.tensor_format == "BCFHW":
+                uncond_states = uncond_states.permute(0, 2, 1, 3, 4)
+                cond_states = cond_states.permute(0, 2, 1, 3, 4)
+            if self.tensor_format == "BCFHW" or self.tensor_format == "BFCHW":
+                uncond_states = uncond_states.flatten(0, 1)
+                cond_states = cond_states.flatten(0, 1)
+
+            low_freq_uncond, high_freq_uncond = _split_low_high_freq(uncond_states.float())
+            low_freq_cond, high_freq_cond = _split_low_high_freq(cond_states.float())
+            self.state.low_frequency_delta = low_freq_uncond - low_freq_cond
+            self.state.high_frequency_delta = high_freq_uncond - high_freq_cond
+
+        self.state.iteration += 1
+        if torch.is_tensor(output):
+            output = hidden_states
+        elif isinstance(output, tuple):
+            output = (hidden_states, *output[1:])
+        else:
+            output.sample = hidden_states
+
+        return output
+
+    def reset_state(self, module: torch.nn.Module) -> torch.nn.Module:
+        self.state.reset()
+        return module
+
+
+class FasterCacheBlockHook(ModelHook):
+    _is_stateful = True
+
+    def __init__(
+        self,
+        block_skip_range: int,
+        timestep_skip_range: Tuple[int, int],
+        is_guidance_distilled: bool,
+        weight_callback: Callable[[torch.nn.Module], float],
+        current_timestep_callback: Callable[[], int],
+    ) -> None:
+        super().__init__()
+
+        self.block_skip_range = block_skip_range
+        self.timestep_skip_range = timestep_skip_range
+        self.is_guidance_distilled = is_guidance_distilled
+
+        self.weight_callback = weight_callback
+        self.current_timestep_callback = current_timestep_callback
+
+    def initialize_hook(self, module):
+        self.state = FasterCacheBlockState()
+        return module
+
+    def _compute_approximated_attention_output(
+        self, t_2_output: torch.Tensor, t_output: torch.Tensor, weight: float, batch_size: int
+    ) -> torch.Tensor:
+        if t_2_output.size(0) != batch_size:
+            # The cache t_2_output contains both batchwise-concatenated unconditional-conditional branch outputs. Just
+            # take the conditional branch outputs.
+            assert t_2_output.size(0) == 2 * batch_size
+            t_2_output = t_2_output[batch_size:]
+        if t_output.size(0) != batch_size:
+            # The cache t_output contains both batchwise-concatenated unconditional-conditional branch outputs. Just
+            # take the conditional branch outputs.
+            assert t_output.size(0) == 2 * batch_size
+            t_output = t_output[batch_size:]
+        return t_output + (t_output - t_2_output) * weight
+
+    def new_forward(self, module: torch.nn.Module, *args, **kwargs) -> Any:
+        batch_size = [
+            *[arg.size(0) for arg in args if torch.is_tensor(arg)],
+            *[v.size(0) for v in kwargs.values() if torch.is_tensor(v)],
+        ][0]
+        if self.state.batch_size is None:
+            # Will be updated on first forward pass through the denoiser
+            self.state.batch_size = batch_size
+
+        # If we have to skip due to the skip conditions, then let's skip as expected.
+        # But, we can't skip if the denoiser wants to infer both unconditional and conditional branches. This
+        # is because the expected output shapes of attention layer will not match if we only return values from
+        # the cache (which only caches conditional branch outputs). So, if state.batch_size (which is the true
+        # unconditional-conditional batch size) is same as the current batch size, we don't perform the layer
+        # skip. Otherwise, we conditionally skip the layer based on what state.skip_callback returns.
+        is_within_timestep_range = (
+            self.timestep_skip_range[0] < self.current_timestep_callback() < self.timestep_skip_range[1]
+        )
+        if not is_within_timestep_range:
+            should_skip_attention = False
+        else:
+            should_compute_attention = self.state.iteration > 0 and self.state.iteration % self.block_skip_range == 0
+            should_skip_attention = not should_compute_attention
+        if should_skip_attention:
+            should_skip_attention = self.is_guidance_distilled or self.state.batch_size != batch_size
+
+        if should_skip_attention:
+            logger.debug("FasterCache - Skipping attention and using approximation")
+            if torch.is_tensor(self.state.cache[-1]):
+                t_2_output, t_output = self.state.cache
+                weight = self.weight_callback(module)
+                output = self._compute_approximated_attention_output(t_2_output, t_output, weight, batch_size)
+            else:
+                # The cache contains multiple tensors from past N iterations (N=2 for FasterCache). We need to handle all of them.
+                # Diffusers blocks can return multiple tensors - let's call them [A, B, C, ...] for simplicity.
+                # In our cache, we would have [[A_1, B_1, C_1, ...], [A_2, B_2, C_2, ...], ...] where each list is the output from
+                # a forward pass of the block. We need to compute the approximated output for each of these tensors.
+                # The zip(*state.cache) operation will give us [(A_1, A_2, ...), (B_1, B_2, ...), (C_1, C_2, ...), ...] which
+                # allows us to compute the approximated attention output for each tensor in the cache.
+                output = ()
+                for t_2_output, t_output in zip(*self.state.cache):
+                    result = self._compute_approximated_attention_output(
+                        t_2_output, t_output, self.weight_callback(module), batch_size
+                    )
+                    output += (result,)
+        else:
+            logger.debug("FasterCache - Computing attention")
+            output = self.fn_ref.original_forward(*args, **kwargs)
+
+        # Note that the following condition for getting hidden_states should suffice since Diffusers blocks either return
+        # a single hidden_states tensor, or a tuple of (hidden_states, encoder_hidden_states) tensors. We need to handle
+        # both cases.
+        if torch.is_tensor(output):
+            cache_output = output
+            if not self.is_guidance_distilled and cache_output.size(0) == self.state.batch_size:
+                # The output here can be both unconditional-conditional branch outputs or just conditional branch outputs.
+                # This is determined at the higher-level denoiser module. We only want to cache the conditional branch outputs.
+                cache_output = cache_output.chunk(2, dim=0)[1]
+        else:
+            # Cache all return values and perform the same operation as above
+            cache_output = ()
+            for out in output:
+                if not self.is_guidance_distilled and out.size(0) == self.state.batch_size:
+                    out = out.chunk(2, dim=0)[1]
+                cache_output += (out,)
+
+        if self.state.cache is None:
+            self.state.cache = [cache_output, cache_output]
+        else:
+            self.state.cache = [self.state.cache[-1], cache_output]
+
+        self.state.iteration += 1
+        return output
+
+    def reset_state(self, module: torch.nn.Module) -> torch.nn.Module:
+        self.state.reset()
+        return module
+
+
+def apply_faster_cache(module: torch.nn.Module, config: FasterCacheConfig) -> None:
+    r"""
+    Applies [FasterCache](https://huggingface.co/papers/2410.19355) to a given pipeline.
+
+    Args:
+        pipeline (`DiffusionPipeline`):
+            The diffusion pipeline to apply FasterCache to.
+        config (`Optional[FasterCacheConfig]`, `optional`, defaults to `None`):
+            The configuration to use for FasterCache.
+
+    Example:
+    ```python
+    >>> import torch
+    >>> from diffusers import CogVideoXPipeline, FasterCacheConfig, apply_faster_cache
+
+    >>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+    >>> pipe.to("cuda")
+
+    >>> config = FasterCacheConfig(
+    ...     spatial_attention_block_skip_range=2,
+    ...     spatial_attention_timestep_skip_range=(-1, 681),
+    ...     low_frequency_weight_update_timestep_range=(99, 641),
+    ...     high_frequency_weight_update_timestep_range=(-1, 301),
+    ...     spatial_attention_block_identifiers=["transformer_blocks"],
+    ...     attention_weight_callback=lambda _: 0.3,
+    ...     tensor_format="BFCHW",
+    ... )
+    >>> apply_faster_cache(pipe.transformer, config)
+    ```
+    """
+
+    logger.warning(
+        "FasterCache is a purely experimental feature and may not work as expected. Not all models support FasterCache. "
+        "The API is subject to change in future releases, with no guarantee of backward compatibility. Please report any issues at "
+        "https://github.com/huggingface/diffusers/issues."
+    )
+
+    if config.attention_weight_callback is None:
+        # If the user has not provided a weight callback, we default to 0.5 for all timesteps.
+        # In the paper, they recommend using a gradually increasing weight from 0 to 1 as the inference progresses, but
+        # this depends from model-to-model. It is required by the user to provide a weight callback if they want to
+        # use a different weight function. Defaulting to 0.5 works well in practice for most cases.
+        logger.warning(
+            "No `attention_weight_callback` provided when enabling FasterCache. Defaulting to using a weight of 0.5 for all timesteps."
+        )
+        config.attention_weight_callback = lambda _: 0.5
+
+    if config.low_frequency_weight_callback is None:
+        logger.debug(
+            "Low frequency weight callback not provided when enabling FasterCache. Defaulting to behaviour described in the paper."
+        )
+
+        def low_frequency_weight_callback(module: torch.nn.Module) -> float:
+            is_within_range = (
+                config.low_frequency_weight_update_timestep_range[0]
+                < config.current_timestep_callback()
+                < config.low_frequency_weight_update_timestep_range[1]
+            )
+            return config.alpha_low_frequency if is_within_range else 1.0
+
+        config.low_frequency_weight_callback = low_frequency_weight_callback
+
+    if config.high_frequency_weight_callback is None:
+        logger.debug(
+            "High frequency weight callback not provided when enabling FasterCache. Defaulting to behaviour described in the paper."
+        )
+
+        def high_frequency_weight_callback(module: torch.nn.Module) -> float:
+            is_within_range = (
+                config.high_frequency_weight_update_timestep_range[0]
+                < config.current_timestep_callback()
+                < config.high_frequency_weight_update_timestep_range[1]
+            )
+            return config.alpha_high_frequency if is_within_range else 1.0
+
+        config.high_frequency_weight_callback = high_frequency_weight_callback
+
+    supported_tensor_formats = ["BCFHW", "BFCHW", "BCHW"]  # TODO(aryan): Support BSC for LTX Video
+    if config.tensor_format not in supported_tensor_formats:
+        raise ValueError(f"`tensor_format` must be one of {supported_tensor_formats}, but got {config.tensor_format}.")
+
+    _apply_faster_cache_on_denoiser(module, config)
+
+    for name, submodule in module.named_modules():
+        if not isinstance(submodule, _ATTENTION_CLASSES):
+            continue
+        if any(re.search(identifier, name) is not None for identifier in _TRANSFORMER_BLOCK_IDENTIFIERS):
+            _apply_faster_cache_on_attention_class(name, submodule, config)
+
+
+def _apply_faster_cache_on_denoiser(module: torch.nn.Module, config: FasterCacheConfig) -> None:
+    hook = FasterCacheDenoiserHook(
+        config.unconditional_batch_skip_range,
+        config.unconditional_batch_timestep_skip_range,
+        config.tensor_format,
+        config.is_guidance_distilled,
+        config._unconditional_conditional_input_kwargs_identifiers,
+        config.current_timestep_callback,
+        config.low_frequency_weight_callback,
+        config.high_frequency_weight_callback,
+    )
+    registry = HookRegistry.check_if_exists_or_initialize(module)
+    registry.register_hook(hook, _FASTER_CACHE_DENOISER_HOOK)
+
+
+def _apply_faster_cache_on_attention_class(name: str, module: Attention, config: FasterCacheConfig) -> None:
+    is_spatial_self_attention = (
+        any(re.search(identifier, name) is not None for identifier in config.spatial_attention_block_identifiers)
+        and config.spatial_attention_block_skip_range is not None
+        and not getattr(module, "is_cross_attention", False)
+    )
+    is_temporal_self_attention = (
+        any(re.search(identifier, name) is not None for identifier in config.temporal_attention_block_identifiers)
+        and config.temporal_attention_block_skip_range is not None
+        and not module.is_cross_attention
+    )
+
+    block_skip_range, timestep_skip_range, block_type = None, None, None
+    if is_spatial_self_attention:
+        block_skip_range = config.spatial_attention_block_skip_range
+        timestep_skip_range = config.spatial_attention_timestep_skip_range
+        block_type = "spatial"
+    elif is_temporal_self_attention:
+        block_skip_range = config.temporal_attention_block_skip_range
+        timestep_skip_range = config.temporal_attention_timestep_skip_range
+        block_type = "temporal"
+
+    if block_skip_range is None or timestep_skip_range is None:
+        logger.debug(
+            f'Unable to apply FasterCache to the selected layer: "{name}" because it does '
+            f"not match any of the required criteria for spatial or temporal attention layers. Note, "
+            f"however, that this layer may still be valid for applying PAB. Please specify the correct "
+            f"block identifiers in the configuration or use the specialized `apply_faster_cache_on_module` "
+            f"function to apply FasterCache to this layer."
+        )
+        return
+
+    logger.debug(f"Enabling FasterCache ({block_type}) for layer: {name}")
+    hook = FasterCacheBlockHook(
+        block_skip_range,
+        timestep_skip_range,
+        config.is_guidance_distilled,
+        config.attention_weight_callback,
+        config.current_timestep_callback,
+    )
+    registry = HookRegistry.check_if_exists_or_initialize(module)
+    registry.register_hook(hook, _FASTER_CACHE_BLOCK_HOOK)
+
+
+# Reference: https://github.com/Vchitect/FasterCache/blob/fab32c15014636dc854948319c0a9a8d92c7acb4/scripts/latte/faster_cache_sample_latte.py#L127C1-L143C39
+@torch.no_grad()
+def _split_low_high_freq(x):
+    fft = torch.fft.fft2(x)
+    fft_shifted = torch.fft.fftshift(fft)
+    height, width = x.shape[-2:]
+    radius = min(height, width) // 5
+
+    y_grid, x_grid = torch.meshgrid(torch.arange(height), torch.arange(width))
+    center_x, center_y = width // 2, height // 2
+    mask = (x_grid - center_x) ** 2 + (y_grid - center_y) ** 2 <= radius**2
+
+    low_freq_mask = mask.unsqueeze(0).unsqueeze(0).to(x.device)
+    high_freq_mask = ~low_freq_mask
+
+    low_freq_fft = fft_shifted * low_freq_mask
+    high_freq_fft = fft_shifted * high_freq_mask
+
+    return low_freq_fft, high_freq_fft

src/diffusers/hooks/group_offloading.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from contextlib import nullcontext
+from contextlib import contextmanager, nullcontext
 from typing import Dict, List, Optional, Set, Tuple
 
 import torch
@@ -56,7 +56,7 @@ class ModuleGroup:
         buffers: Optional[List[torch.Tensor]] = None,
         non_blocking: bool = False,
         stream: Optional[torch.cuda.Stream] = None,
-        cpu_param_dict: Optional[Dict[torch.nn.Parameter, torch.Tensor]] = None,
+        low_cpu_mem_usage=False,
         onload_self: bool = True,
     ) -> None:
         self.modules = modules
@@ -64,15 +64,50 @@ class ModuleGroup:
         self.onload_device = onload_device
         self.offload_leader = offload_leader
         self.onload_leader = onload_leader
-        self.parameters = parameters
-        self.buffers = buffers
+        self.parameters = parameters or []
+        self.buffers = buffers or []
         self.non_blocking = non_blocking or stream is not None
         self.stream = stream
-        self.cpu_param_dict = cpu_param_dict
         self.onload_self = onload_self
+        self.low_cpu_mem_usage = low_cpu_mem_usage
 
-        if self.stream is not None and self.cpu_param_dict is None:
-            raise ValueError("cpu_param_dict must be provided when using stream for data transfer.")
+        self.cpu_param_dict = self._init_cpu_param_dict()
+
+    def _init_cpu_param_dict(self):
+        cpu_param_dict = {}
+        if self.stream is None:
+            return cpu_param_dict
+
+        for module in self.modules:
+            for param in module.parameters():
+                cpu_param_dict[param] = param.data.cpu() if self.low_cpu_mem_usage else param.data.cpu().pin_memory()
+            for buffer in module.buffers():
+                cpu_param_dict[buffer] = (
+                    buffer.data.cpu() if self.low_cpu_mem_usage else buffer.data.cpu().pin_memory()
+                )
+
+        for param in self.parameters:
+            cpu_param_dict[param] = param.data.cpu() if self.low_cpu_mem_usage else param.data.cpu().pin_memory()
+
+        for buffer in self.buffers:
+            cpu_param_dict[buffer] = buffer.data.cpu() if self.low_cpu_mem_usage else buffer.data.cpu().pin_memory()
+
+        return cpu_param_dict
+
+    @contextmanager
+    def _pinned_memory_tensors(self):
+        pinned_dict = {}
+        try:
+            for param, tensor in self.cpu_param_dict.items():
+                if not tensor.is_pinned():
+                    pinned_dict[param] = tensor.pin_memory()
+                else:
+                    pinned_dict[param] = tensor
+
+            yield pinned_dict
+
+        finally:
+            pinned_dict = None
 
     def onload_(self):
         r"""Onloads the group of modules to the onload_device."""
@@ -82,12 +117,30 @@ class ModuleGroup:
             self.stream.synchronize()
 
         with context:
-            for group_module in self.modules:
-                group_module.to(self.onload_device, non_blocking=self.non_blocking)
-            if self.parameters is not None:
+            if self.stream is not None:
+                with self._pinned_memory_tensors() as pinned_memory:
+                    for group_module in self.modules:
+                        for param in group_module.parameters():
+                            param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
+                        for buffer in group_module.buffers():
+                            buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
+
+                    for param in self.parameters:
+                        param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
+
+                    for buffer in self.buffers:
+                        buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
+
+            else:
+                for group_module in self.modules:
+                    for param in group_module.parameters():
+                        param.data = param.data.to(self.onload_device, non_blocking=self.non_blocking)
+                    for buffer in group_module.buffers():
+                        buffer.data = buffer.data.to(self.onload_device, non_blocking=self.non_blocking)
+
                 for param in self.parameters:
                     param.data = param.data.to(self.onload_device, non_blocking=self.non_blocking)
-            if self.buffers is not None:
+
                 for buffer in self.buffers:
                     buffer.data = buffer.data.to(self.onload_device, non_blocking=self.non_blocking)
 
@@ -98,15 +151,18 @@ class ModuleGroup:
             for group_module in self.modules:
                 for param in group_module.parameters():
                     param.data = self.cpu_param_dict[param]
+            for param in self.parameters:
+                param.data = self.cpu_param_dict[param]
+            for buffer in self.buffers:
+                buffer.data = self.cpu_param_dict[buffer]
+
         else:
             for group_module in self.modules:
                 group_module.to(self.offload_device, non_blocking=self.non_blocking)
-            if self.parameters is not None:
-                for param in self.parameters:
-                    param.data = param.data.to(self.offload_device, non_blocking=self.non_blocking)
-            if self.buffers is not None:
-                for buffer in self.buffers:
-                    buffer.data = buffer.data.to(self.offload_device, non_blocking=self.non_blocking)
+            for param in self.parameters:
+                param.data = param.data.to(self.offload_device, non_blocking=self.non_blocking)
+            for buffer in self.buffers:
+                buffer.data = buffer.data.to(self.offload_device, non_blocking=self.non_blocking)
 
 
 class GroupOffloadingHook(ModelHook):
@@ -172,6 +228,13 @@ class LazyPrefetchGroupOffloadingHook(ModelHook):
         self._layer_execution_tracker_module_names = set()
 
     def initialize_hook(self, module):
+        def make_execution_order_update_callback(current_name, current_submodule):
+            def callback():
+                logger.debug(f"Adding {current_name} to the execution order")
+                self.execution_order.append((current_name, current_submodule))
+
+            return callback
+
         # To every submodule that contains a group offloading hook (at this point, no prefetching is enabled for any
         # of the groups), we add a layer execution tracker hook that will be used to determine the order in which the
         # layers are executed during the forward pass.
@@ -183,14 +246,8 @@ class LazyPrefetchGroupOffloadingHook(ModelHook):
             group_offloading_hook = registry.get_hook(_GROUP_OFFLOADING)
 
             if group_offloading_hook is not None:
-
-                def make_execution_order_update_callback(current_name, current_submodule):
-                    def callback():
-                        logger.debug(f"Adding {current_name} to the execution order")
-                        self.execution_order.append((current_name, current_submodule))
-
-                    return callback
-
+                # For the first forward pass, we have to load in a blocking manner
+                group_offloading_hook.group.non_blocking = False
                 layer_tracker_hook = LayerExecutionTrackerHook(make_execution_order_update_callback(name, submodule))
                 registry.register_hook(layer_tracker_hook, _LAYER_EXECUTION_TRACKER)
                 self._layer_execution_tracker_module_names.add(name)
@@ -220,6 +277,7 @@ class LazyPrefetchGroupOffloadingHook(ModelHook):
         # Remove the layer execution tracker hooks from the submodules
         base_module_registry = module._diffusers_hook
         registries = [submodule._diffusers_hook for _, submodule in self.execution_order]
+        group_offloading_hooks = [registry.get_hook(_GROUP_OFFLOADING) for registry in registries]
 
         for i in range(num_executed):
             registries[i].remove_hook(_LAYER_EXECUTION_TRACKER, recurse=False)
@@ -227,8 +285,13 @@ class LazyPrefetchGroupOffloadingHook(ModelHook):
         # Remove the current lazy prefetch group offloading hook so that it doesn't interfere with the next forward pass
         base_module_registry.remove_hook(_LAZY_PREFETCH_GROUP_OFFLOADING, recurse=False)
 
-        # Apply lazy prefetching by setting required attributes
-        group_offloading_hooks = [registry.get_hook(_GROUP_OFFLOADING) for registry in registries]
+        # LazyPrefetchGroupOffloadingHook is only used with streams, so we know that non_blocking should be True.
+        # We disable non_blocking for the first forward pass, but need to enable it for the subsequent passes to
+        # see the benefits of prefetching.
+        for hook in group_offloading_hooks:
+            hook.group.non_blocking = True
+
+        # Set required attributes for prefetching
         if num_executed > 0:
             base_module_group_offloading_hook = base_module_registry.get_hook(_GROUP_OFFLOADING)
             base_module_group_offloading_hook.next_group = group_offloading_hooks[0].group
@@ -268,6 +331,7 @@ def apply_group_offloading(
     num_blocks_per_group: Optional[int] = None,
     non_blocking: bool = False,
     use_stream: bool = False,
+    low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
     Applies group offloading to the internal layers of a torch.nn.Module. To understand what group offloading is, and
@@ -314,6 +378,10 @@ def apply_group_offloading(
         use_stream (`bool`, defaults to `False`):
             If True, offloading and onloading is done asynchronously using a CUDA stream. This can be useful for
             overlapping computation and data transfer.
+        low_cpu_mem_usage (`bool`, defaults to `False`):
+            If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. This
+            option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
+            the CPU memory is a bottleneck but may counteract the benefits of using streams.
 
     Example:
         ```python
@@ -349,10 +417,12 @@ def apply_group_offloading(
             raise ValueError("num_blocks_per_group must be provided when using offload_type='block_level'.")
 
         _apply_group_offloading_block_level(
-            module, num_blocks_per_group, offload_device, onload_device, non_blocking, stream
+            module, num_blocks_per_group, offload_device, onload_device, non_blocking, stream, low_cpu_mem_usage
         )
     elif offload_type == "leaf_level":
-        _apply_group_offloading_leaf_level(module, offload_device, onload_device, non_blocking, stream)
+        _apply_group_offloading_leaf_level(
+            module, offload_device, onload_device, non_blocking, stream, low_cpu_mem_usage
+        )
     else:
         raise ValueError(f"Unsupported offload_type: {offload_type}")
 
@@ -364,6 +434,7 @@ def _apply_group_offloading_block_level(
     onload_device: torch.device,
     non_blocking: bool,
     stream: Optional[torch.cuda.Stream] = None,
+    low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
     This function applies offloading to groups of torch.nn.ModuleList or torch.nn.Sequential blocks. In comparison to
@@ -384,13 +455,6 @@ def _apply_group_offloading_block_level(
             for overlapping computation and data transfer.
     """
 
-    # Create a pinned CPU parameter dict for async data transfer if streams are to be used
-    cpu_param_dict = None
-    if stream is not None:
-        for param in module.parameters():
-            param.data = param.data.cpu().pin_memory()
-        cpu_param_dict = {param: param.data for param in module.parameters()}
-
     # Create module groups for ModuleList and Sequential blocks
     modules_with_group_offloading = set()
     unmatched_modules = []
@@ -411,7 +475,7 @@ def _apply_group_offloading_block_level(
                 onload_leader=current_modules[0],
                 non_blocking=non_blocking,
                 stream=stream,
-                cpu_param_dict=cpu_param_dict,
+                low_cpu_mem_usage=low_cpu_mem_usage,
                 onload_self=stream is None,
             )
             matched_module_groups.append(group)
@@ -448,7 +512,6 @@ def _apply_group_offloading_block_level(
         buffers=buffers,
         non_blocking=False,
         stream=None,
-        cpu_param_dict=None,
         onload_self=True,
     )
     next_group = matched_module_groups[0] if len(matched_module_groups) > 0 else None
@@ -461,6 +524,7 @@ def _apply_group_offloading_leaf_level(
     onload_device: torch.device,
     non_blocking: bool,
     stream: Optional[torch.cuda.Stream] = None,
+    low_cpu_mem_usage: bool = False,
 ) -> None:
     r"""
     This function applies offloading to groups of leaf modules in a torch.nn.Module. This method has minimal memory
@@ -483,13 +547,6 @@ def _apply_group_offloading_leaf_level(
             for overlapping computation and data transfer.
     """
 
-    # Create a pinned CPU parameter dict for async data transfer if streams are to be used
-    cpu_param_dict = None
-    if stream is not None:
-        for param in module.parameters():
-            param.data = param.data.cpu().pin_memory()
-        cpu_param_dict = {param: param.data for param in module.parameters()}
-
     # Create module groups for leaf modules and apply group offloading hooks
     modules_with_group_offloading = set()
     for name, submodule in module.named_modules():
@@ -503,7 +560,7 @@ def _apply_group_offloading_leaf_level(
             onload_leader=submodule,
             non_blocking=non_blocking,
             stream=stream,
-            cpu_param_dict=cpu_param_dict,
+            low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )
         _apply_group_offloading_hook(submodule, group, None)
@@ -548,7 +605,7 @@ def _apply_group_offloading_leaf_level(
             buffers=buffers,
             non_blocking=non_blocking,
             stream=stream,
-            cpu_param_dict=cpu_param_dict,
+            low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )
         _apply_group_offloading_hook(parent_module, group, None)
@@ -567,7 +624,7 @@ def _apply_group_offloading_leaf_level(
             buffers=None,
             non_blocking=False,
             stream=None,
-            cpu_param_dict=None,
+            low_cpu_mem_usage=low_cpu_mem_usage,
             onload_self=True,
         )
         _apply_lazy_group_offloading_hook(module, unmatched_group, None)

src/diffusers/hooks/pyramid_attention_broadcast.py

@@ -26,8 +26,8 @@ from .hooks import HookRegistry, ModelHook
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
+_PYRAMID_ATTENTION_BROADCAST_HOOK = "pyramid_attention_broadcast"
 _ATTENTION_CLASSES = (Attention, MochiAttention)
-
 _SPATIAL_ATTENTION_BLOCK_IDENTIFIERS = ("blocks", "transformer_blocks", "single_transformer_blocks")
 _TEMPORAL_ATTENTION_BLOCK_IDENTIFIERS = ("temporal_transformer_blocks",)
 _CROSS_ATTENTION_BLOCK_IDENTIFIERS = ("blocks", "transformer_blocks")
@@ -87,7 +87,7 @@ class PyramidAttentionBroadcastConfig:
 
     def __repr__(self) -> str:
         return (
-            f"PyramidAttentionBroadcastConfig("
+            f"PyramidAttentionBroadcastConfig(\n"
             f"  spatial_attention_block_skip_range={self.spatial_attention_block_skip_range},\n"
             f"  temporal_attention_block_skip_range={self.temporal_attention_block_skip_range},\n"
             f"  cross_attention_block_skip_range={self.cross_attention_block_skip_range},\n"
@@ -175,10 +175,7 @@ class PyramidAttentionBroadcastHook(ModelHook):
         return module
 
 
-def apply_pyramid_attention_broadcast(
-    module: torch.nn.Module,
-    config: PyramidAttentionBroadcastConfig,
-):
+def apply_pyramid_attention_broadcast(module: torch.nn.Module, config: PyramidAttentionBroadcastConfig):
     r"""
     Apply [Pyramid Attention Broadcast](https://huggingface.co/papers/2408.12588) to a given pipeline.
 
@@ -311,4 +308,4 @@ def _apply_pyramid_attention_broadcast_hook(
     """
     registry = HookRegistry.check_if_exists_or_initialize(module)
     hook = PyramidAttentionBroadcastHook(timestep_skip_range, block_skip_range, current_timestep_callback)
-    registry.register_hook(hook, "pyramid_attention_broadcast")
+    registry.register_hook(hook, _PYRAMID_ATTENTION_BROADCAST_HOOK)

src/diffusers/loaders/ip_adapter.py

@@ -804,9 +804,7 @@ class SD3IPAdapterMixin:
                     }
 
                     self.register_modules(
-                        feature_extractor=SiglipImageProcessor.from_pretrained(image_encoder_subfolder, **kwargs).to(
-                            self.device, dtype=self.dtype
-                        ),
+                        feature_extractor=SiglipImageProcessor.from_pretrained(image_encoder_subfolder, **kwargs),
                         image_encoder=SiglipVisionModel.from_pretrained(
                             image_encoder_subfolder, torch_dtype=self.dtype, **kwargs
                         ).to(self.device),

src/diffusers/loaders/lora_base.py

@@ -423,8 +423,12 @@ def _load_lora_into_text_encoder(
         # Unsafe code />
 
     if prefix is not None and not state_dict:
-        logger.info(
-            f"No LoRA keys associated to {text_encoder.__class__.__name__} found with the {prefix=}. This is safe to ignore if LoRA state dict didn't originally have any {text_encoder.__class__.__name__} related params. Open an issue if you think it's unexpected: https://github.com/huggingface/diffusers/issues/new"
+        logger.warning(
+            f"No LoRA keys associated to {text_encoder.__class__.__name__} found with the {prefix=}. "
+            "This is safe to ignore if LoRA state dict didn't originally have any "
+            f"{text_encoder.__class__.__name__} related params. You can also try specifying `prefix=None` "
+            "to resolve the warning. Otherwise, open an issue if you think it's unexpected: "
+            "https://github.com/huggingface/diffusers/issues/new"
         )
 
 

src/diffusers/loaders/lora_conversion_utils.py

@@ -1348,3 +1348,56 @@ def _convert_non_diffusers_lumina2_lora_to_diffusers(state_dict):
         converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
 
     return converted_state_dict
+
+
+def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
+    converted_state_dict = {}
+    original_state_dict = {k[len("diffusion_model.") :]: v for k, v in state_dict.items()}
+
+    num_blocks = len({k.split("blocks.")[1].split(".")[0] for k in original_state_dict})
+    is_i2v_lora = any("k_img" in k for k in original_state_dict) and any("v_img" in k for k in original_state_dict)
+
+    for i in range(num_blocks):
+        # Self-attention
+        for o, c in zip(["q", "k", "v", "o"], ["to_q", "to_k", "to_v", "to_out.0"]):
+            converted_state_dict[f"blocks.{i}.attn1.{c}.lora_A.weight"] = original_state_dict.pop(
+                f"blocks.{i}.self_attn.{o}.lora_A.weight"
+            )
+            converted_state_dict[f"blocks.{i}.attn1.{c}.lora_B.weight"] = original_state_dict.pop(
+                f"blocks.{i}.self_attn.{o}.lora_B.weight"
+            )
+
+        # Cross-attention
+        for o, c in zip(["q", "k", "v", "o"], ["to_q", "to_k", "to_v", "to_out.0"]):
+            converted_state_dict[f"blocks.{i}.attn2.{c}.lora_A.weight"] = original_state_dict.pop(
+                f"blocks.{i}.cross_attn.{o}.lora_A.weight"
+            )
+            converted_state_dict[f"blocks.{i}.attn2.{c}.lora_B.weight"] = original_state_dict.pop(
+                f"blocks.{i}.cross_attn.{o}.lora_B.weight"
+            )
+
+        if is_i2v_lora:
+            for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
+                converted_state_dict[f"blocks.{i}.attn2.{c}.lora_A.weight"] = original_state_dict.pop(
+                    f"blocks.{i}.cross_attn.{o}.lora_A.weight"
+                )
+                converted_state_dict[f"blocks.{i}.attn2.{c}.lora_B.weight"] = original_state_dict.pop(
+                    f"blocks.{i}.cross_attn.{o}.lora_B.weight"
+                )
+
+        # FFN
+        for o, c in zip(["ffn.0", "ffn.2"], ["net.0.proj", "net.2"]):
+            converted_state_dict[f"blocks.{i}.ffn.{c}.lora_A.weight"] = original_state_dict.pop(
+                f"blocks.{i}.{o}.lora_A.weight"
+            )
+            converted_state_dict[f"blocks.{i}.ffn.{c}.lora_B.weight"] = original_state_dict.pop(
+                f"blocks.{i}.{o}.lora_B.weight"
+            )
+
+    if len(original_state_dict) > 0:
+        raise ValueError(f"`state_dict` should be empty at this point but has {original_state_dict.keys()=}")
+
+    for key in list(converted_state_dict.keys()):
+        converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
+
+    return converted_state_dict

src/diffusers/loaders/lora_pipeline.py

@@ -42,6 +42,7 @@ from .lora_conversion_utils import (
     _convert_kohya_flux_lora_to_diffusers,
     _convert_non_diffusers_lora_to_diffusers,
     _convert_non_diffusers_lumina2_lora_to_diffusers,
+    _convert_non_diffusers_wan_lora_to_diffusers,
     _convert_xlabs_flux_lora_to_diffusers,
     _maybe_map_sgm_blocks_to_diffusers,
 )
@@ -451,7 +452,11 @@ class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     def unfuse_lora(self, components: List[str] = ["unet", "text_encoder"], **kwargs):
@@ -472,7 +477,7 @@ class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
                 Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                 LoRA parameters then it won't have any effect.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
@@ -891,7 +896,11 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     def unfuse_lora(self, components: List[str] = ["unet", "text_encoder", "text_encoder_2"], **kwargs):
@@ -912,7 +921,7 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
                 Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                 LoRA parameters then it won't have any effect.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class SD3LoraLoaderMixin(LoraBaseMixin):
@@ -1290,7 +1299,11 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.StableDiffusionXLLoraLoaderMixin.unfuse_lora with unet->transformer
@@ -1312,7 +1325,7 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
                 Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
                 LoRA parameters then it won't have any effect.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class FluxLoraLoaderMixin(LoraBaseMixin):
@@ -1828,7 +1841,11 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
             )
 
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder"], **kwargs):
@@ -1849,7 +1866,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
         if hasattr(transformer, "_transformer_norm_layers") and transformer._transformer_norm_layers:
             transformer.load_state_dict(transformer._transformer_norm_layers, strict=False)
 
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
     # We override this here account for `_transformer_norm_layers` and `_overwritten_params`.
     def unload_lora_weights(self, reset_to_overwritten_params=False):
@@ -2548,7 +2565,11 @@ class CogVideoXLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
@@ -2566,7 +2587,7 @@ class CogVideoXLoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class Mochi1LoraLoaderMixin(LoraBaseMixin):
@@ -2852,7 +2873,11 @@ class Mochi1LoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -2871,7 +2896,7 @@ class Mochi1LoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class LTXVideoLoraLoaderMixin(LoraBaseMixin):
@@ -3157,7 +3182,11 @@ class LTXVideoLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -3176,7 +3205,7 @@ class LTXVideoLoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class SanaLoraLoaderMixin(LoraBaseMixin):
@@ -3462,7 +3491,11 @@ class SanaLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -3481,7 +3514,7 @@ class SanaLoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class HunyuanVideoLoraLoaderMixin(LoraBaseMixin):
@@ -3770,7 +3803,11 @@ class HunyuanVideoLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -3789,7 +3826,7 @@ class HunyuanVideoLoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class Lumina2LoraLoaderMixin(LoraBaseMixin):
@@ -4079,7 +4116,11 @@ class Lumina2LoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.unfuse_lora
@@ -4098,7 +4139,7 @@ class Lumina2LoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class WanLoraLoaderMixin(LoraBaseMixin):
@@ -4111,7 +4152,6 @@ class WanLoraLoaderMixin(LoraBaseMixin):
 
     @classmethod
     @validate_hf_hub_args
-    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.lora_state_dict
     def lora_state_dict(
         cls,
         pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
@@ -4198,6 +4238,8 @@ class WanLoraLoaderMixin(LoraBaseMixin):
             user_agent=user_agent,
             allow_pickle=allow_pickle,
         )
+        if any(k.startswith("diffusion_model.") for k in state_dict):
+            state_dict = _convert_non_diffusers_wan_lora_to_diffusers(state_dict)
 
         is_dora_scale_present = any("dora_scale" in k for k in state_dict)
         if is_dora_scale_present:
@@ -4207,7 +4249,33 @@ class WanLoraLoaderMixin(LoraBaseMixin):
 
         return state_dict
 
-    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
+    @classmethod
+    def _maybe_expand_t2v_lora_for_i2v(
+        cls,
+        transformer: torch.nn.Module,
+        state_dict,
+    ):
+        if transformer.config.image_dim is None:
+            return state_dict
+
+        if any(k.startswith("transformer.blocks.") for k in state_dict):
+            num_blocks = len({k.split("blocks.")[1].split(".")[0] for k in state_dict})
+            is_i2v_lora = any("add_k_proj" in k for k in state_dict) and any("add_v_proj" in k for k in state_dict)
+
+            if is_i2v_lora:
+                return state_dict
+
+            for i in range(num_blocks):
+                for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
+                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_A.weight"] = torch.zeros_like(
+                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"]
+                    )
+                    state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.weight"] = torch.zeros_like(
+                        state_dict[f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"]
+                    )
+
+        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
     ):
@@ -4245,7 +4313,11 @@ class WanLoraLoaderMixin(LoraBaseMixin):
 
         # 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)
-
+        # convert T2V LoRA to I2V LoRA (when loaded to Wan I2V) by adding zeros for the additional (missing) _img layers
+        state_dict = self._maybe_expand_t2v_lora_for_i2v(
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            state_dict=state_dict,
+        )
         is_correct_format = all("lora" in key for key in state_dict.keys())
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
@@ -4384,7 +4456,11 @@ class WanLoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -4403,7 +4479,7 @@ class WanLoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class CogView4LoraLoaderMixin(LoraBaseMixin):
@@ -4689,7 +4765,11 @@ class CogView4LoraLoaderMixin(LoraBaseMixin):
         ```
         """
         super().fuse_lora(
-            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
         )
 
     # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
@@ -4708,7 +4788,7 @@ class CogView4LoraLoaderMixin(LoraBaseMixin):
             components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
             unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
         """
-        super().unfuse_lora(components=components)
+        super().unfuse_lora(components=components, **kwargs)
 
 
 class LoraLoaderMixin(StableDiffusionLoraLoaderMixin):

src/diffusers/loaders/peft.py

@@ -354,8 +354,12 @@ class PeftAdapterMixin:
             # Unsafe code />
 
         if prefix is not None and not state_dict:
-            logger.info(
-                f"No LoRA keys associated to {self.__class__.__name__} found with the {prefix=}. This is safe to ignore if LoRA state dict didn't originally have any {self.__class__.__name__} related params. Open an issue if you think it's unexpected: https://github.com/huggingface/diffusers/issues/new"
+            logger.warning(
+                f"No LoRA keys associated to {self.__class__.__name__} found with the {prefix=}. "
+                "This is safe to ignore if LoRA state dict didn't originally have any "
+                f"{self.__class__.__name__} related params. You can also try specifying `prefix=None` "
+                "to resolve the warning. Otherwise, open an issue if you think it's unexpected: "
+                "https://github.com/huggingface/diffusers/issues/new"
             )
 
     def save_lora_adapter(

src/diffusers/loaders/single_file.py

@@ -360,12 +360,12 @@ class FromSingleFileMixin:
         cache_dir = kwargs.pop("cache_dir", None)
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
+        torch_dtype = kwargs.pop("torch_dtype", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
 
         is_legacy_loading = False
 
-        if not isinstance(torch_dtype, torch.dtype):
+        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
             torch_dtype = torch.float32
             logger.warning(
                 f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."

src/diffusers/loaders/single_file_model.py

@@ -255,12 +255,12 @@ class FromOriginalModelMixin:
         subfolder = kwargs.pop("subfolder", None)
         revision = kwargs.pop("revision", None)
         config_revision = kwargs.pop("config_revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
+        torch_dtype = kwargs.pop("torch_dtype", None)
         quantization_config = kwargs.pop("quantization_config", None)
         device = kwargs.pop("device", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
 
-        if not isinstance(torch_dtype, torch.dtype):
+        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
             torch_dtype = torch.float32
             logger.warning(
                 f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."
@@ -282,6 +282,7 @@ class FromOriginalModelMixin:
         if quantization_config is not None:
             hf_quantizer = DiffusersAutoQuantizer.from_config(quantization_config)
             hf_quantizer.validate_environment()
+            torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype)
 
         else:
             hf_quantizer = None

src/diffusers/loaders/textual_inversion.py

@@ -449,9 +449,9 @@ class TextualInversionLoaderMixin:
 
         # 7.5 Offload the model again
         if is_model_cpu_offload:
-            self.enable_model_cpu_offload()
+            self.enable_model_cpu_offload(device=device)
         elif is_sequential_cpu_offload:
-            self.enable_sequential_cpu_offload()
+            self.enable_sequential_cpu_offload(device=device)
 
         # / Unsafe Code >
 

src/diffusers/models/attention_processor.py

@@ -741,10 +741,14 @@ class Attention(nn.Module):
 
         if out_dim == 3:
             if attention_mask.shape[0] < batch_size * head_size:
-                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+                attention_mask = attention_mask.repeat_interleave(
+                    head_size, dim=0, output_size=attention_mask.shape[0] * head_size
+                )
         elif out_dim == 4:
             attention_mask = attention_mask.unsqueeze(1)
-            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+            attention_mask = attention_mask.repeat_interleave(
+                head_size, dim=1, output_size=attention_mask.shape[1] * head_size
+            )
 
         return attention_mask
 
@@ -2335,7 +2339,9 @@ class FluxAttnProcessor2_0:
             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 = 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)
         hidden_states = hidden_states.to(query.dtype)
@@ -3704,8 +3710,10 @@ class StableAudioAttnProcessor2_0:
         if kv_heads != attn.heads:
             # if GQA or MQA, repeat the key/value heads to reach the number of query heads.
             heads_per_kv_head = attn.heads // kv_heads
-            key = torch.repeat_interleave(key, heads_per_kv_head, dim=1)
-            value = torch.repeat_interleave(value, heads_per_kv_head, dim=1)
+            key = torch.repeat_interleave(key, heads_per_kv_head, dim=1, output_size=key.shape[1] * heads_per_kv_head)
+            value = torch.repeat_interleave(
+                value, heads_per_kv_head, dim=1, output_size=value.shape[1] * heads_per_kv_head
+            )
 
         if attn.norm_q is not None:
             query = attn.norm_q(query)
@@ -6012,6 +6020,11 @@ class SanaLinearAttnProcessor2_0:
         key = attn.to_k(encoder_hidden_states)
         value = attn.to_v(encoder_hidden_states)
 
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
         query = query.transpose(1, 2).unflatten(1, (attn.heads, -1))
         key = key.transpose(1, 2).unflatten(1, (attn.heads, -1)).transpose(2, 3)
         value = value.transpose(1, 2).unflatten(1, (attn.heads, -1))

src/diffusers/models/autoencoders/autoencoder_dc.py

@@ -190,7 +190,7 @@ class DCUpBlock2d(nn.Module):
             x = F.pixel_shuffle(x, self.factor)
 
         if self.shortcut:
-            y = hidden_states.repeat_interleave(self.repeats, dim=1)
+            y = hidden_states.repeat_interleave(self.repeats, dim=1, output_size=hidden_states.shape[1] * self.repeats)
             y = F.pixel_shuffle(y, self.factor)
             hidden_states = x + y
         else:
@@ -361,7 +361,9 @@ class Decoder(nn.Module):
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         if self.in_shortcut:
-            x = hidden_states.repeat_interleave(self.in_shortcut_repeats, dim=1)
+            x = hidden_states.repeat_interleave(
+                self.in_shortcut_repeats, dim=1, output_size=hidden_states.shape[1] * self.in_shortcut_repeats
+            )
             hidden_states = self.conv_in(hidden_states) + x
         else:
             hidden_states = self.conv_in(hidden_states)

src/diffusers/models/autoencoders/autoencoder_kl_allegro.py

@@ -103,7 +103,7 @@ class AllegroTemporalConvLayer(nn.Module):
         if self.down_sample:
             identity = hidden_states[:, :, ::2]
         elif self.up_sample:
-            identity = hidden_states.repeat_interleave(2, dim=2)
+            identity = hidden_states.repeat_interleave(2, dim=2, output_size=hidden_states.shape[2] * 2)
         else:
             identity = hidden_states
 

src/diffusers/models/autoencoders/autoencoder_kl_ltx.py

@@ -196,6 +196,55 @@ class LTXVideoResnetBlock3d(nn.Module):
         return hidden_states
 
 
+class LTXVideoDownsampler3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        is_causal: bool = True,
+        padding_mode: str = "zeros",
+    ) -> None:
+        super().__init__()
+
+        self.stride = stride if isinstance(stride, tuple) else (stride, stride, stride)
+        self.group_size = (in_channels * stride[0] * stride[1] * stride[2]) // out_channels
+
+        out_channels = out_channels // (self.stride[0] * self.stride[1] * self.stride[2])
+
+        self.conv = LTXVideoCausalConv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=3,
+            stride=1,
+            is_causal=is_causal,
+            padding_mode=padding_mode,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = torch.cat([hidden_states[:, :, : self.stride[0] - 1], hidden_states], dim=2)
+
+        residual = (
+            hidden_states.unflatten(4, (-1, self.stride[2]))
+            .unflatten(3, (-1, self.stride[1]))
+            .unflatten(2, (-1, self.stride[0]))
+        )
+        residual = residual.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(1, 4)
+        residual = residual.unflatten(1, (-1, self.group_size))
+        residual = residual.mean(dim=2)
+
+        hidden_states = self.conv(hidden_states)
+        hidden_states = (
+            hidden_states.unflatten(4, (-1, self.stride[2]))
+            .unflatten(3, (-1, self.stride[1]))
+            .unflatten(2, (-1, self.stride[0]))
+        )
+        hidden_states = hidden_states.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(1, 4)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
 class LTXVideoUpsampler3d(nn.Module):
     def __init__(
         self,
@@ -204,6 +253,7 @@ class LTXVideoUpsampler3d(nn.Module):
         is_causal: bool = True,
         residual: bool = False,
         upscale_factor: int = 1,
+        padding_mode: str = "zeros",
     ) -> None:
         super().__init__()
 
@@ -219,6 +269,7 @@ class LTXVideoUpsampler3d(nn.Module):
             kernel_size=3,
             stride=1,
             is_causal=is_causal,
+            padding_mode=padding_mode,
         )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -352,6 +403,118 @@ class LTXVideoDownBlock3D(nn.Module):
         return hidden_states
 
 
+class LTXVideo095DownBlock3D(nn.Module):
+    r"""
+    Down block used in the LTXVideo model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        spatio_temporal_scale (`bool`, defaults to `True`):
+            Whether or not to use a downsampling layer. If not used, output dimension would be same as input dimension.
+            Whether or not to downsample across temporal dimension.
+        is_causal (`bool`, defaults to `True`):
+            Whether this layer behaves causally (future frames depend only on past frames) or not.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        spatio_temporal_scale: bool = True,
+        is_causal: bool = True,
+        downsample_type: str = "conv",
+    ):
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+
+        resnets = []
+        for _ in range(num_layers):
+            resnets.append(
+                LTXVideoResnetBlock3d(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    dropout=dropout,
+                    eps=resnet_eps,
+                    non_linearity=resnet_act_fn,
+                    is_causal=is_causal,
+                )
+            )
+        self.resnets = nn.ModuleList(resnets)
+
+        self.downsamplers = None
+        if spatio_temporal_scale:
+            self.downsamplers = nn.ModuleList()
+
+            if downsample_type == "conv":
+                self.downsamplers.append(
+                    LTXVideoCausalConv3d(
+                        in_channels=in_channels,
+                        out_channels=in_channels,
+                        kernel_size=3,
+                        stride=(2, 2, 2),
+                        is_causal=is_causal,
+                    )
+                )
+            elif downsample_type == "spatial":
+                self.downsamplers.append(
+                    LTXVideoDownsampler3d(
+                        in_channels=in_channels, out_channels=out_channels, stride=(1, 2, 2), is_causal=is_causal
+                    )
+                )
+            elif downsample_type == "temporal":
+                self.downsamplers.append(
+                    LTXVideoDownsampler3d(
+                        in_channels=in_channels, out_channels=out_channels, stride=(2, 1, 1), is_causal=is_causal
+                    )
+                )
+            elif downsample_type == "spatiotemporal":
+                self.downsamplers.append(
+                    LTXVideoDownsampler3d(
+                        in_channels=in_channels, out_channels=out_channels, stride=(2, 2, 2), is_causal=is_causal
+                    )
+                )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        generator: Optional[torch.Generator] = None,
+    ) -> torch.Tensor:
+        r"""Forward method of the `LTXDownBlock3D` class."""
+
+        for i, resnet in enumerate(self.resnets):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb, generator)
+            else:
+                hidden_states = resnet(hidden_states, temb, generator)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+        return hidden_states
+
+
 # Adapted from diffusers.models.autoencoders.autoencoder_kl_cogvideox.CogVideoMidBlock3d
 class LTXVideoMidBlock3d(nn.Module):
     r"""
@@ -593,8 +756,15 @@ class LTXVideoEncoder3d(nn.Module):
         in_channels: int = 3,
         out_channels: int = 128,
         block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
+        down_block_types: Tuple[str, ...] = (
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+        ),
         spatio_temporal_scaling: Tuple[bool, ...] = (True, True, True, False),
         layers_per_block: Tuple[int, ...] = (4, 3, 3, 3, 4),
+        downsample_type: Tuple[str, ...] = ("conv", "conv", "conv", "conv"),
         patch_size: int = 4,
         patch_size_t: int = 1,
         resnet_norm_eps: float = 1e-6,
@@ -617,20 +787,37 @@ class LTXVideoEncoder3d(nn.Module):
         )
 
         # down blocks
-        num_block_out_channels = len(block_out_channels)
+        is_ltx_095 = down_block_types[-1] == "LTXVideo095DownBlock3D"
+        num_block_out_channels = len(block_out_channels) - (1 if is_ltx_095 else 0)
         self.down_blocks = nn.ModuleList([])
         for i in range(num_block_out_channels):
             input_channel = output_channel
-            output_channel = block_out_channels[i + 1] if i + 1 < num_block_out_channels else block_out_channels[i]
+            if not is_ltx_095:
+                output_channel = block_out_channels[i + 1] if i + 1 < num_block_out_channels else block_out_channels[i]
+            else:
+                output_channel = block_out_channels[i + 1]
 
-            down_block = LTXVideoDownBlock3D(
-                in_channels=input_channel,
-                out_channels=output_channel,
-                num_layers=layers_per_block[i],
-                resnet_eps=resnet_norm_eps,
-                spatio_temporal_scale=spatio_temporal_scaling[i],
-                is_causal=is_causal,
-            )
+            if down_block_types[i] == "LTXVideoDownBlock3D":
+                down_block = LTXVideoDownBlock3D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    num_layers=layers_per_block[i],
+                    resnet_eps=resnet_norm_eps,
+                    spatio_temporal_scale=spatio_temporal_scaling[i],
+                    is_causal=is_causal,
+                )
+            elif down_block_types[i] == "LTXVideo095DownBlock3D":
+                down_block = LTXVideo095DownBlock3D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    num_layers=layers_per_block[i],
+                    resnet_eps=resnet_norm_eps,
+                    spatio_temporal_scale=spatio_temporal_scaling[i],
+                    is_causal=is_causal,
+                    downsample_type=downsample_type[i],
+                )
+            else:
+                raise ValueError(f"Unknown down block type: {down_block_types[i]}")
 
             self.down_blocks.append(down_block)
 
@@ -794,7 +981,9 @@ class LTXVideoDecoder3d(nn.Module):
         # timestep embedding
         self.time_embedder = None
         self.scale_shift_table = None
+        self.timestep_scale_multiplier = None
         if timestep_conditioning:
+            self.timestep_scale_multiplier = nn.Parameter(torch.tensor(1000.0, dtype=torch.float32))
             self.time_embedder = PixArtAlphaCombinedTimestepSizeEmbeddings(output_channel * 2, 0)
             self.scale_shift_table = nn.Parameter(torch.randn(2, output_channel) / output_channel**0.5)
 
@@ -803,6 +992,9 @@ class LTXVideoDecoder3d(nn.Module):
     def forward(self, hidden_states: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
         hidden_states = self.conv_in(hidden_states)
 
+        if self.timestep_scale_multiplier is not None:
+            temb = temb * self.timestep_scale_multiplier
+
         if torch.is_grad_enabled() and self.gradient_checkpointing:
             hidden_states = self._gradient_checkpointing_func(self.mid_block, hidden_states, temb)
 
@@ -891,12 +1083,19 @@ class AutoencoderKLLTXVideo(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         out_channels: int = 3,
         latent_channels: int = 128,
         block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
+        down_block_types: Tuple[str, ...] = (
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+            "LTXVideoDownBlock3D",
+        ),
         decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
         layers_per_block: Tuple[int, ...] = (4, 3, 3, 3, 4),
         decoder_layers_per_block: Tuple[int, ...] = (4, 3, 3, 3, 4),
         spatio_temporal_scaling: Tuple[bool, ...] = (True, True, True, False),
         decoder_spatio_temporal_scaling: Tuple[bool, ...] = (True, True, True, False),
         decoder_inject_noise: Tuple[bool, ...] = (False, False, False, False, False),
+        downsample_type: Tuple[str, ...] = ("conv", "conv", "conv", "conv"),
         upsample_residual: Tuple[bool, ...] = (False, False, False, False),
         upsample_factor: Tuple[int, ...] = (1, 1, 1, 1),
         timestep_conditioning: bool = False,
@@ -906,6 +1105,8 @@ class AutoencoderKLLTXVideo(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         scaling_factor: float = 1.0,
         encoder_causal: bool = True,
         decoder_causal: bool = False,
+        spatial_compression_ratio: int = None,
+        temporal_compression_ratio: int = None,
     ) -> None:
         super().__init__()
 
@@ -913,8 +1114,10 @@ class AutoencoderKLLTXVideo(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             in_channels=in_channels,
             out_channels=latent_channels,
             block_out_channels=block_out_channels,
+            down_block_types=down_block_types,
             spatio_temporal_scaling=spatio_temporal_scaling,
             layers_per_block=layers_per_block,
+            downsample_type=downsample_type,
             patch_size=patch_size,
             patch_size_t=patch_size_t,
             resnet_norm_eps=resnet_norm_eps,
@@ -941,8 +1144,16 @@ class AutoencoderKLLTXVideo(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         self.register_buffer("latents_mean", latents_mean, persistent=True)
         self.register_buffer("latents_std", latents_std, persistent=True)
 
-        self.spatial_compression_ratio = patch_size * 2 ** sum(spatio_temporal_scaling)
-        self.temporal_compression_ratio = patch_size_t * 2 ** sum(spatio_temporal_scaling)
+        self.spatial_compression_ratio = (
+            patch_size * 2 ** sum(spatio_temporal_scaling)
+            if spatial_compression_ratio is None
+            else spatial_compression_ratio
+        )
+        self.temporal_compression_ratio = (
+            patch_size_t * 2 ** sum(spatio_temporal_scaling)
+            if temporal_compression_ratio is None
+            else temporal_compression_ratio
+        )
 
         # When decoding a batch of video latents at a time, one can save memory by slicing across the batch dimension
         # to perform decoding of a single video latent at a time.

src/diffusers/models/autoencoders/autoencoder_kl_mochi.py

@@ -426,7 +426,9 @@ class FourierFeatures(nn.Module):
         w = w.repeat(num_channels)[None, :, None, None, None]  # [1, num_channels * num_freqs, 1, 1, 1]
 
         # Interleaved repeat of input channels to match w
-        h = inputs.repeat_interleave(num_freqs, dim=1)  # [B, C * num_freqs, T, H, W]
+        h = inputs.repeat_interleave(
+            num_freqs, dim=1, output_size=inputs.shape[1] * num_freqs
+        )  # [B, C * num_freqs, T, H, W]
         # Scale channels by frequency.
         h = w * h
 

src/diffusers/models/cache_utils.py

@@ -24,6 +24,7 @@ class CacheMixin:
 
     Supported caching techniques:
         - [Pyramid Attention Broadcast](https://huggingface.co/papers/2408.12588)
+        - [FasterCache](https://huggingface.co/papers/2410.19355)
     """
 
     _cache_config = None
@@ -59,17 +60,31 @@ class CacheMixin:
         ```
         """
 
-        from ..hooks import PyramidAttentionBroadcastConfig, apply_pyramid_attention_broadcast
+        from ..hooks import (
+            FasterCacheConfig,
+            PyramidAttentionBroadcastConfig,
+            apply_faster_cache,
+            apply_pyramid_attention_broadcast,
+        )
+
+        if self.is_cache_enabled:
+            raise ValueError(
+                f"Caching has already been enabled with {type(self._cache_config)}. To apply a new caching technique, please disable the existing one first."
+            )
 
         if isinstance(config, PyramidAttentionBroadcastConfig):
             apply_pyramid_attention_broadcast(self, config)
+        elif isinstance(config, FasterCacheConfig):
+            apply_faster_cache(self, config)
         else:
             raise ValueError(f"Cache config {type(config)} is not supported.")
 
         self._cache_config = config
 
     def disable_cache(self) -> None:
-        from ..hooks import HookRegistry, PyramidAttentionBroadcastConfig
+        from ..hooks import FasterCacheConfig, HookRegistry, PyramidAttentionBroadcastConfig
+        from ..hooks.faster_cache import _FASTER_CACHE_BLOCK_HOOK, _FASTER_CACHE_DENOISER_HOOK
+        from ..hooks.pyramid_attention_broadcast import _PYRAMID_ATTENTION_BROADCAST_HOOK
 
         if self._cache_config is None:
             logger.warning("Caching techniques have not been enabled, so there's nothing to disable.")
@@ -77,7 +92,11 @@ class CacheMixin:
 
         if isinstance(self._cache_config, PyramidAttentionBroadcastConfig):
             registry = HookRegistry.check_if_exists_or_initialize(self)
-            registry.remove_hook("pyramid_attention_broadcast", recurse=True)
+            registry.remove_hook(_PYRAMID_ATTENTION_BROADCAST_HOOK, recurse=True)
+        elif isinstance(self._cache_config, FasterCacheConfig):
+            registry = HookRegistry.check_if_exists_or_initialize(self)
+            registry.remove_hook(_FASTER_CACHE_DENOISER_HOOK, recurse=True)
+            registry.remove_hook(_FASTER_CACHE_BLOCK_HOOK, recurse=True)
         else:
             raise ValueError(f"Cache config {type(self._cache_config)} is not supported.")
 

src/diffusers/models/controlnets/controlnet_sparsectrl.py

@@ -687,7 +687,7 @@ class SparseControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         t_emb = t_emb.to(dtype=sample.dtype)
 
         emb = self.time_embedding(t_emb, timestep_cond)
-        emb = emb.repeat_interleave(sample_num_frames, dim=0)
+        emb = emb.repeat_interleave(sample_num_frames, dim=0, output_size=emb.shape[0] * sample_num_frames)
 
         # 2. pre-process
         batch_size, channels, num_frames, height, width = sample.shape

src/diffusers/models/embeddings.py

@@ -139,7 +139,9 @@ def get_3d_sincos_pos_embed(
 
     # 3. Concat
     pos_embed_spatial = pos_embed_spatial[None, :, :]
-    pos_embed_spatial = pos_embed_spatial.repeat_interleave(temporal_size, dim=0)  # [T, H*W, D // 4 * 3]
+    pos_embed_spatial = pos_embed_spatial.repeat_interleave(
+        temporal_size, dim=0, output_size=pos_embed_spatial.shape[0] * temporal_size
+    )  # [T, H*W, D // 4 * 3]
 
     pos_embed_temporal = pos_embed_temporal[:, None, :]
     pos_embed_temporal = pos_embed_temporal.repeat_interleave(
@@ -334,7 +336,7 @@ def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
             " `from_numpy` is no longer required."
             "  Pass `output_type='pt' to use the new version now."
         )
-        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        deprecate("output_type=='np'", "0.34.0", deprecation_message, standard_warn=False)
         return get_1d_sincos_pos_embed_from_grid_np(embed_dim=embed_dim, pos=pos)
     if embed_dim % 2 != 0:
         raise ValueError("embed_dim must be divisible by 2")
@@ -1152,10 +1154,13 @@ def get_1d_rotary_pos_embed(
         / linear_factor
     )  # [D/2]
     freqs = torch.outer(pos, freqs)  # type: ignore   # [S, D/2]
+    is_npu = freqs.device.type == "npu"
+    if is_npu:
+        freqs = freqs.float()
     if use_real and repeat_interleave_real:
         # 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]
+        freqs_cos = freqs.cos().repeat_interleave(2, dim=1, output_size=freqs.shape[1] * 2).float()  # [S, D]
+        freqs_sin = freqs.sin().repeat_interleave(2, dim=1, output_size=freqs.shape[1] * 2).float()  # [S, D]
         return freqs_cos, freqs_sin
     elif use_real:
         # stable audio, allegro

src/diffusers/models/modeling_utils.py

@@ -37,7 +37,6 @@ from torch import Tensor, nn
 from typing_extensions import Self
 
 from .. import __version__
-from ..hooks import apply_group_offloading, apply_layerwise_casting
 from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer
 from ..quantizers.quantization_config import QuantizationMethod
 from ..utils import (
@@ -504,6 +503,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             non_blocking (`bool`, *optional*, defaults to `False`):
                 If `True`, the weight casting operations are non-blocking.
         """
+        from ..hooks import apply_layerwise_casting
 
         user_provided_patterns = True
         if skip_modules_pattern is None:
@@ -546,6 +546,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         num_blocks_per_group: Optional[int] = None,
         non_blocking: bool = False,
         use_stream: bool = False,
+        low_cpu_mem_usage=False,
     ) -> None:
         r"""
         Activates group offloading for the current model.
@@ -569,6 +570,8 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
             ... )
             ```
         """
+        from ..hooks import apply_group_offloading
+
         if getattr(self, "enable_tiling", None) is not None and getattr(self, "use_tiling", False) and use_stream:
             msg = (
                 "Applying group offloading on autoencoders, with CUDA streams, may not work as expected if the first "
@@ -584,7 +587,14 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 f"open an issue at https://github.com/huggingface/diffusers/issues."
             )
         apply_group_offloading(
-            self, onload_device, offload_device, offload_type, num_blocks_per_group, non_blocking, use_stream
+            self,
+            onload_device,
+            offload_device,
+            offload_type,
+            num_blocks_per_group,
+            non_blocking,
+            use_stream,
+            low_cpu_mem_usage=low_cpu_mem_usage,
         )
 
     def save_pretrained(
@@ -870,7 +880,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         local_files_only = kwargs.pop("local_files_only", None)
         token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
+        torch_dtype = kwargs.pop("torch_dtype", None)
         subfolder = kwargs.pop("subfolder", None)
         device_map = kwargs.pop("device_map", None)
         max_memory = kwargs.pop("max_memory", None)
@@ -883,7 +893,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
 
-        if not isinstance(torch_dtype, torch.dtype):
+        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
             torch_dtype = torch.float32
             logger.warning(
                 f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."

src/diffusers/models/normalization.py

@@ -550,16 +550,6 @@ class RMSNorm(nn.Module):
             hidden_states = torch_npu.npu_rms_norm(hidden_states, self.weight, epsilon=self.eps)[0]
             if self.bias is not None:
                 hidden_states = hidden_states + self.bias
-        elif is_torch_version(">=", "2.4"):
-            if self.weight is not None:
-                # convert into half-precision if necessary
-                if self.weight.dtype in [torch.float16, torch.bfloat16]:
-                    hidden_states = hidden_states.to(self.weight.dtype)
-            hidden_states = nn.functional.rms_norm(
-                hidden_states, normalized_shape=(hidden_states.shape[-1],), weight=self.weight, eps=self.eps
-            )
-            if self.bias is not None:
-                hidden_states = hidden_states + self.bias
         else:
             input_dtype = hidden_states.dtype
             variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)

src/diffusers/models/resnet.py

@@ -366,7 +366,7 @@ class ResnetBlock2D(nn.Module):
         hidden_states = self.conv2(hidden_states)
 
         if self.conv_shortcut is not None:
-            input_tensor = self.conv_shortcut(input_tensor)
+            input_tensor = self.conv_shortcut(input_tensor.contiguous())
 
         output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
 

src/diffusers/models/transformers/latte_transformer_3d.py

@@ -227,13 +227,17 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
         # Prepare text embeddings for spatial block
         # batch_size num_tokens hidden_size -> (batch_size * num_frame) num_tokens hidden_size
         encoder_hidden_states = self.caption_projection(encoder_hidden_states)  # 3 120 1152
-        encoder_hidden_states_spatial = encoder_hidden_states.repeat_interleave(num_frame, dim=0).view(
-            -1, encoder_hidden_states.shape[-2], encoder_hidden_states.shape[-1]
-        )
+        encoder_hidden_states_spatial = encoder_hidden_states.repeat_interleave(
+            num_frame, dim=0, output_size=encoder_hidden_states.shape[0] * num_frame
+        ).view(-1, encoder_hidden_states.shape[-2], encoder_hidden_states.shape[-1])
 
         # Prepare timesteps for spatial and temporal block
-        timestep_spatial = timestep.repeat_interleave(num_frame, dim=0).view(-1, timestep.shape[-1])
-        timestep_temp = timestep.repeat_interleave(num_patches, dim=0).view(-1, timestep.shape[-1])
+        timestep_spatial = timestep.repeat_interleave(
+            num_frame, dim=0, output_size=timestep.shape[0] * num_frame
+        ).view(-1, timestep.shape[-1])
+        timestep_temp = timestep.repeat_interleave(
+            num_patches, dim=0, output_size=timestep.shape[0] * num_patches
+        ).view(-1, timestep.shape[-1])
 
         # Spatial and temporal transformer blocks
         for i, (spatial_block, temp_block) in enumerate(
@@ -299,7 +303,9 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
                 ).permute(0, 2, 1, 3)
                 hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])
 
-        embedded_timestep = embedded_timestep.repeat_interleave(num_frame, dim=0).view(-1, embedded_timestep.shape[-1])
+        embedded_timestep = embedded_timestep.repeat_interleave(
+            num_frame, dim=0, output_size=embedded_timestep.shape[0] * num_frame
+        ).view(-1, embedded_timestep.shape[-1])
         shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
         hidden_states = self.norm_out(hidden_states)
         # Modulation

src/diffusers/models/transformers/prior_transformer.py

@@ -353,7 +353,11 @@ class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Pef
             attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
             attention_mask = F.pad(attention_mask, (0, self.additional_embeddings), value=0.0)
             attention_mask = (attention_mask[:, None, :] + self.causal_attention_mask).to(hidden_states.dtype)
-            attention_mask = attention_mask.repeat_interleave(self.config.num_attention_heads, dim=0)
+            attention_mask = attention_mask.repeat_interleave(
+                self.config.num_attention_heads,
+                dim=0,
+                output_size=attention_mask.shape[0] * self.config.num_attention_heads,
+            )
 
         if self.norm_in is not None:
             hidden_states = self.norm_in(hidden_states)

src/diffusers/models/transformers/sana_transformer.py

@@ -15,6 +15,7 @@
 from typing import Any, Dict, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
@@ -23,10 +24,9 @@ from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_
 from ..attention_processor import (
     Attention,
     AttentionProcessor,
-    AttnProcessor2_0,
     SanaLinearAttnProcessor2_0,
 )
-from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNormSingle, RMSNorm
@@ -96,6 +96,95 @@ class SanaModulatedNorm(nn.Module):
         return hidden_states
 
 
+class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+
+    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
+
+        guidance_proj = self.guidance_condition_proj(guidance)
+        guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
+        conditioning = timesteps_emb + guidance_emb
+
+        return self.linear(self.silu(conditioning)), conditioning
+
+
+class SanaAttnProcessor2_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("SanaAttnProcessor2_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.Tensor] = None,
+    ) -> torch.Tensor:
+        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)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        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)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        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)
+
+        # 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, 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)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class SanaTransformerBlock(nn.Module):
     r"""
     Transformer block introduced in [Sana](https://huggingface.co/papers/2410.10629).
@@ -115,6 +204,7 @@ class SanaTransformerBlock(nn.Module):
         norm_eps: float = 1e-6,
         attention_out_bias: bool = True,
         mlp_ratio: float = 2.5,
+        qk_norm: Optional[str] = None,
     ) -> None:
         super().__init__()
 
@@ -124,6 +214,8 @@ class SanaTransformerBlock(nn.Module):
             query_dim=dim,
             heads=num_attention_heads,
             dim_head=attention_head_dim,
+            kv_heads=num_attention_heads if qk_norm is not None else None,
+            qk_norm=qk_norm,
             dropout=dropout,
             bias=attention_bias,
             cross_attention_dim=None,
@@ -135,13 +227,15 @@ class SanaTransformerBlock(nn.Module):
             self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
             self.attn2 = Attention(
                 query_dim=dim,
+                qk_norm=qk_norm,
+                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
                 cross_attention_dim=cross_attention_dim,
                 heads=num_cross_attention_heads,
                 dim_head=cross_attention_head_dim,
                 dropout=dropout,
                 bias=True,
                 out_bias=attention_out_bias,
-                processor=AttnProcessor2_0(),
+                processor=SanaAttnProcessor2_0(),
             )
 
         # 3. Feed-forward
@@ -232,6 +326,10 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
             Whether to use elementwise affinity in the normalization layer.
         norm_eps (`float`, defaults to `1e-6`):
             The epsilon value for the normalization layer.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for the query and key.
+        timestep_scale (`float`, defaults to `1.0`):
+            The scale to use for the timesteps.
     """
 
     _supports_gradient_checkpointing = True
@@ -258,6 +356,10 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         norm_elementwise_affine: bool = False,
         norm_eps: float = 1e-6,
         interpolation_scale: Optional[int] = None,
+        guidance_embeds: bool = False,
+        guidance_embeds_scale: float = 0.1,
+        qk_norm: Optional[str] = None,
+        timestep_scale: float = 1.0,
     ) -> None:
         super().__init__()
 
@@ -276,7 +378,10 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
 
         # 2. Additional condition embeddings
-        self.time_embed = AdaLayerNormSingle(inner_dim)
+        if guidance_embeds:
+            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
+        else:
+            self.time_embed = AdaLayerNormSingle(inner_dim)
 
         self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
         self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
@@ -296,6 +401,7 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     norm_elementwise_affine=norm_elementwise_affine,
                     norm_eps=norm_eps,
                     mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
                 )
                 for _ in range(num_layers)
             ]
@@ -372,7 +478,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
-        timestep: torch.LongTensor,
+        timestep: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
         encoder_attention_mask: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -423,9 +530,14 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         hidden_states = self.patch_embed(hidden_states)
 
-        timestep, embedded_timestep = self.time_embed(
-            timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
-        )
+        if guidance is not None:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, guidance=guidance, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
 
         encoder_hidden_states = self.caption_projection(encoder_hidden_states)
         encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])

src/diffusers/models/transformers/transformer_cogview4.py

@@ -23,6 +23,7 @@ from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ..attention import FeedForward
 from ..attention_processor import Attention
+from ..cache_utils import CacheMixin
 from ..embeddings import CogView3CombinedTimestepSizeEmbeddings
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -126,7 +127,8 @@ class CogView4AttnProcessor:
         attention_mask: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        text_seq_length = encoder_hidden_states.size(1)
+        batch_size, text_seq_length, embed_dim = encoder_hidden_states.shape
+        batch_size, image_seq_length, embed_dim = hidden_states.shape
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
         # 1. QKV projections
@@ -156,6 +158,15 @@ class CogView4AttnProcessor:
             )
 
         # 4. Attention
+        if attention_mask is not None:
+            text_attention_mask = attention_mask.float().to(query.device)
+            actual_text_seq_length = text_attention_mask.size(1)
+            new_attention_mask = torch.zeros((batch_size, text_seq_length + image_seq_length), device=query.device)
+            new_attention_mask[:, :actual_text_seq_length] = text_attention_mask
+            new_attention_mask = new_attention_mask.unsqueeze(2)
+            attention_mask_matrix = new_attention_mask @ new_attention_mask.transpose(1, 2)
+            attention_mask = (attention_mask_matrix > 0).unsqueeze(1).to(query.dtype)
+
         hidden_states = F.scaled_dot_product_attention(
             query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
         )
@@ -203,6 +214,8 @@ class CogView4TransformerBlock(nn.Module):
         encoder_hidden_states: torch.Tensor,
         temb: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
     ) -> torch.Tensor:
         # 1. Timestep conditioning
         (
@@ -223,6 +236,8 @@ class CogView4TransformerBlock(nn.Module):
             hidden_states=norm_hidden_states,
             encoder_hidden_states=norm_encoder_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            attention_mask=attention_mask,
+            **kwargs,
         )
         hidden_states = hidden_states + attn_hidden_states * gate_msa.unsqueeze(1)
         encoder_hidden_states = encoder_hidden_states + attn_encoder_hidden_states * c_gate_msa.unsqueeze(1)
@@ -289,7 +304,7 @@ class CogView4RotaryPosEmbed(nn.Module):
         return (freqs.cos(), freqs.sin())
 
 
-class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
     r"""
     Args:
         patch_size (`int`, defaults to `2`):
@@ -386,6 +401,8 @@ class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         crop_coords: torch.Tensor,
         attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
     ) -> Union[torch.Tensor, Transformer2DModelOutput]:
         if attention_kwargs is not None:
             attention_kwargs = attention_kwargs.copy()
@@ -421,11 +438,11 @@ class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         for block in self.transformer_blocks:
             if torch.is_grad_enabled() and self.gradient_checkpointing:
                 hidden_states, encoder_hidden_states = self._gradient_checkpointing_func(
-                    block, hidden_states, encoder_hidden_states, temb, image_rotary_emb
+                    block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, attention_mask, **kwargs
                 )
             else:
                 hidden_states, encoder_hidden_states = block(
-                    hidden_states, encoder_hidden_states, temb, image_rotary_emb
+                    hidden_states, encoder_hidden_states, temb, image_rotary_emb, attention_mask, **kwargs
                 )
 
         # 4. Output norm & projection

src/diffusers/models/transformers/transformer_hunyuan_video.py

@@ -27,13 +27,15 @@ from ..attention import FeedForward
 from ..attention_processor import Attention, AttentionProcessor
 from ..cache_utils import CacheMixin
 from ..embeddings import (
-    CombinedTimestepGuidanceTextProjEmbeddings,
     CombinedTimestepTextProjEmbeddings,
+    PixArtAlphaTextProjection,
+    TimestepEmbedding,
+    Timesteps,
     get_1d_rotary_pos_embed,
 )
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
-from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
+from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle, FP32LayerNorm
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -173,6 +175,141 @@ class HunyuanVideoAdaNorm(nn.Module):
         return gate_msa, gate_mlp
 
 
+class HunyuanVideoTokenReplaceAdaLayerNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, norm_type: str = "layer_norm", bias: bool = True):
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=bias)
+
+        if norm_type == "layer_norm":
+            self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+        elif norm_type == "fp32_layer_norm":
+            self.norm = FP32LayerNorm(embedding_dim, elementwise_affine=False, bias=False)
+        else:
+            raise ValueError(
+                f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        emb: torch.Tensor,
+        token_replace_emb: torch.Tensor,
+        first_frame_num_tokens: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        token_replace_emb = self.linear(self.silu(token_replace_emb))
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
+        tr_shift_msa, tr_scale_msa, tr_gate_msa, tr_shift_mlp, tr_scale_mlp, tr_gate_mlp = token_replace_emb.chunk(
+            6, dim=1
+        )
+
+        norm_hidden_states = self.norm(hidden_states)
+        hidden_states_zero = (
+            norm_hidden_states[:, :first_frame_num_tokens] * (1 + tr_scale_msa[:, None]) + tr_shift_msa[:, None]
+        )
+        hidden_states_orig = (
+            norm_hidden_states[:, first_frame_num_tokens:] * (1 + scale_msa[:, None]) + shift_msa[:, None]
+        )
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+
+        return (
+            hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            tr_gate_msa,
+            tr_shift_mlp,
+            tr_scale_mlp,
+            tr_gate_mlp,
+        )
+
+
+class HunyuanVideoTokenReplaceAdaLayerNormZeroSingle(nn.Module):
+    def __init__(self, embedding_dim: int, norm_type: str = "layer_norm", bias: bool = True):
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 3 * embedding_dim, bias=bias)
+
+        if norm_type == "layer_norm":
+            self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+        else:
+            raise ValueError(
+                f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        emb: torch.Tensor,
+        token_replace_emb: torch.Tensor,
+        first_frame_num_tokens: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        token_replace_emb = self.linear(self.silu(token_replace_emb))
+
+        shift_msa, scale_msa, gate_msa = emb.chunk(3, dim=1)
+        tr_shift_msa, tr_scale_msa, tr_gate_msa = token_replace_emb.chunk(3, dim=1)
+
+        norm_hidden_states = self.norm(hidden_states)
+        hidden_states_zero = (
+            norm_hidden_states[:, :first_frame_num_tokens] * (1 + tr_scale_msa[:, None]) + tr_shift_msa[:, None]
+        )
+        hidden_states_orig = (
+            norm_hidden_states[:, first_frame_num_tokens:] * (1 + scale_msa[:, None]) + shift_msa[:, None]
+        )
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+
+        return hidden_states, gate_msa, tr_gate_msa
+
+
+class HunyuanVideoConditionEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        pooled_projection_dim: int,
+        guidance_embeds: bool,
+        image_condition_type: Optional[str] = None,
+    ):
+        super().__init__()
+
+        self.image_condition_type = image_condition_type
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+        self.text_embedder = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim, act_fn="silu")
+
+        self.guidance_embedder = None
+        if guidance_embeds:
+            self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(
+        self, timestep: torch.Tensor, pooled_projection: torch.Tensor, guidance: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=pooled_projection.dtype))  # (N, D)
+        pooled_projections = self.text_embedder(pooled_projection)
+        conditioning = timesteps_emb + pooled_projections
+
+        token_replace_emb = None
+        if self.image_condition_type == "token_replace":
+            token_replace_timestep = torch.zeros_like(timestep)
+            token_replace_proj = self.time_proj(token_replace_timestep)
+            token_replace_emb = self.timestep_embedder(token_replace_proj.to(dtype=pooled_projection.dtype))
+            token_replace_emb = token_replace_emb + pooled_projections
+
+        if self.guidance_embedder is not None:
+            guidance_proj = self.time_proj(guidance)
+            guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=pooled_projection.dtype))
+            conditioning = conditioning + guidance_emb
+
+        return conditioning, token_replace_emb
+
+
 class HunyuanVideoIndividualTokenRefinerBlock(nn.Module):
     def __init__(
         self,
@@ -390,6 +527,8 @@ class HunyuanVideoSingleTransformerBlock(nn.Module):
         temb: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        *args,
+        **kwargs,
     ) -> torch.Tensor:
         text_seq_length = encoder_hidden_states.shape[1]
         hidden_states = torch.cat([hidden_states, encoder_hidden_states], dim=1)
@@ -468,6 +607,8 @@ class HunyuanVideoTransformerBlock(nn.Module):
         temb: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        *args,
+        **kwargs,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         # 1. Input normalization
         norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
@@ -503,6 +644,181 @@ class HunyuanVideoTransformerBlock(nn.Module):
         return hidden_states, encoder_hidden_states
 
 
+class HunyuanVideoTokenReplaceSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: str = "rms_norm",
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+        mlp_dim = int(hidden_size * mlp_ratio)
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(),
+            qk_norm=qk_norm,
+            eps=1e-6,
+            pre_only=True,
+        )
+
+        self.norm = HunyuanVideoTokenReplaceAdaLayerNormZeroSingle(hidden_size, norm_type="layer_norm")
+        self.proj_mlp = nn.Linear(hidden_size, mlp_dim)
+        self.act_mlp = nn.GELU(approximate="tanh")
+        self.proj_out = nn.Linear(hidden_size + mlp_dim, hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        token_replace_emb: torch.Tensor = None,
+        num_tokens: int = None,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, encoder_hidden_states], dim=1)
+
+        residual = hidden_states
+
+        # 1. Input normalization
+        norm_hidden_states, gate, tr_gate = self.norm(hidden_states, temb, token_replace_emb, num_tokens)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+
+        norm_hidden_states, norm_encoder_hidden_states = (
+            norm_hidden_states[:, :-text_seq_length, :],
+            norm_hidden_states[:, -text_seq_length:, :],
+        )
+
+        # 2. Attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+        attn_output = torch.cat([attn_output, context_attn_output], dim=1)
+
+        # 3. Modulation and residual connection
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+
+        proj_output = self.proj_out(hidden_states)
+        hidden_states_zero = proj_output[:, :num_tokens] * tr_gate.unsqueeze(1)
+        hidden_states_orig = proj_output[:, num_tokens:] * gate.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        hidden_states = hidden_states + residual
+
+        hidden_states, encoder_hidden_states = (
+            hidden_states[:, :-text_seq_length, :],
+            hidden_states[:, -text_seq_length:, :],
+        )
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoTokenReplaceTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float,
+        qk_norm: str = "rms_norm",
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+
+        self.norm1 = HunyuanVideoTokenReplaceAdaLayerNormZero(hidden_size, norm_type="layer_norm")
+        self.norm1_context = AdaLayerNormZero(hidden_size, norm_type="layer_norm")
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            added_kv_proj_dim=hidden_size,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            context_pre_only=False,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(),
+            qk_norm=qk_norm,
+            eps=1e-6,
+        )
+
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+        self.norm2_context = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff_context = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        token_replace_emb: torch.Tensor = None,
+        num_tokens: int = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # 1. Input normalization
+        (
+            norm_hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            tr_gate_msa,
+            tr_shift_mlp,
+            tr_scale_mlp,
+            tr_gate_mlp,
+        ) = self.norm1(hidden_states, temb, token_replace_emb, num_tokens)
+        norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+            encoder_hidden_states, emb=temb
+        )
+
+        # 2. Joint attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=freqs_cis,
+        )
+
+        # 3. Modulation and residual connection
+        hidden_states_zero = hidden_states[:, :num_tokens] + attn_output[:, :num_tokens] * tr_gate_msa.unsqueeze(1)
+        hidden_states_orig = hidden_states[:, num_tokens:] + attn_output[:, num_tokens:] * gate_msa.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+
+        hidden_states_zero = norm_hidden_states[:, :num_tokens] * (1 + tr_scale_mlp[:, None]) + tr_shift_mlp[:, None]
+        hidden_states_orig = norm_hidden_states[:, num_tokens:] * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        norm_hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+        # 4. Feed-forward
+        ff_output = self.ff(norm_hidden_states)
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+
+        hidden_states_zero = hidden_states[:, :num_tokens] + ff_output[:, :num_tokens] * tr_gate_mlp.unsqueeze(1)
+        hidden_states_orig = hidden_states[:, num_tokens:] + ff_output[:, num_tokens:] * gate_mlp.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return hidden_states, encoder_hidden_states
+
+
 class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
     r"""
     A Transformer model for video-like data used in [HunyuanVideo](https://huggingface.co/tencent/HunyuanVideo).
@@ -540,6 +856,10 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
             The value of theta to use in the RoPE layer.
         rope_axes_dim (`Tuple[int]`, defaults to `(16, 56, 56)`):
             The dimensions of the axes to use in the RoPE layer.
+        image_condition_type (`str`, *optional*, defaults to `None`):
+            The type of image conditioning to use. If `None`, no image conditioning is used. If `latent_concat`, the
+            image is concatenated to the latent stream. If `token_replace`, the image is used to replace first-frame
+            tokens in the latent stream and apply conditioning.
     """
 
     _supports_gradient_checkpointing = True
@@ -570,9 +890,16 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
         pooled_projection_dim: int = 768,
         rope_theta: float = 256.0,
         rope_axes_dim: Tuple[int] = (16, 56, 56),
+        image_condition_type: Optional[str] = None,
     ) -> None:
         super().__init__()
 
+        supported_image_condition_types = ["latent_concat", "token_replace"]
+        if image_condition_type is not None and image_condition_type not in supported_image_condition_types:
+            raise ValueError(
+                f"Invalid `image_condition_type` ({image_condition_type}). Supported ones are: {supported_image_condition_types}"
+            )
+
         inner_dim = num_attention_heads * attention_head_dim
         out_channels = out_channels or in_channels
 
@@ -582,33 +909,52 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
             text_embed_dim, num_attention_heads, attention_head_dim, num_layers=num_refiner_layers
         )
 
-        if guidance_embeds:
-            self.time_text_embed = CombinedTimestepGuidanceTextProjEmbeddings(inner_dim, pooled_projection_dim)
-        else:
-            self.time_text_embed = CombinedTimestepTextProjEmbeddings(inner_dim, pooled_projection_dim)
+        self.time_text_embed = HunyuanVideoConditionEmbedding(
+            inner_dim, pooled_projection_dim, guidance_embeds, image_condition_type
+        )
 
         # 2. RoPE
         self.rope = HunyuanVideoRotaryPosEmbed(patch_size, patch_size_t, rope_axes_dim, rope_theta)
 
         # 3. Dual stream transformer blocks
-        self.transformer_blocks = nn.ModuleList(
-            [
-                HunyuanVideoTransformerBlock(
-                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
-                )
-                for _ in range(num_layers)
-            ]
-        )
+        if image_condition_type == "token_replace":
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTokenReplaceTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                    )
+                    for _ in range(num_layers)
+                ]
+            )
+        else:
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                    )
+                    for _ in range(num_layers)
+                ]
+            )
 
         # 4. Single stream transformer blocks
-        self.single_transformer_blocks = nn.ModuleList(
-            [
-                HunyuanVideoSingleTransformerBlock(
-                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
-                )
-                for _ in range(num_single_layers)
-            ]
-        )
+        if image_condition_type == "token_replace":
+            self.single_transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTokenReplaceSingleTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                    )
+                    for _ in range(num_single_layers)
+                ]
+            )
+        else:
+            self.single_transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoSingleTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                    )
+                    for _ in range(num_single_layers)
+                ]
+            )
 
         # 5. Output projection
         self.norm_out = AdaLayerNormContinuous(inner_dim, inner_dim, elementwise_affine=False, eps=1e-6)
@@ -707,15 +1053,13 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
         post_patch_num_frames = num_frames // p_t
         post_patch_height = height // p
         post_patch_width = width // p
+        first_frame_num_tokens = 1 * post_patch_height * post_patch_width
 
         # 1. RoPE
         image_rotary_emb = self.rope(hidden_states)
 
         # 2. Conditional embeddings
-        if self.config.guidance_embeds:
-            temb = self.time_text_embed(timestep, guidance, pooled_projections)
-        else:
-            temb = self.time_text_embed(timestep, pooled_projections)
+        temb, token_replace_emb = self.time_text_embed(timestep, pooled_projections, guidance)
 
         hidden_states = self.x_embedder(hidden_states)
         encoder_hidden_states = self.context_embedder(encoder_hidden_states, timestep, encoder_attention_mask)
@@ -746,6 +1090,8 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
                     temb,
                     attention_mask,
                     image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
                 )
 
             for block in self.single_transformer_blocks:
@@ -756,17 +1102,31 @@ class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin,
                     temb,
                     attention_mask,
                     image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
                 )
 
         else:
             for block in self.transformer_blocks:
                 hidden_states, encoder_hidden_states = block(
-                    hidden_states, encoder_hidden_states, temb, attention_mask, image_rotary_emb
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
                 )
 
             for block in self.single_transformer_blocks:
                 hidden_states, encoder_hidden_states = block(
-                    hidden_states, encoder_hidden_states, temb, attention_mask, image_rotary_emb
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
                 )
 
         # 5. Output projection

src/diffusers/models/transformers/transformer_ltx.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 
 import math
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -113,20 +113,19 @@ class LTXVideoRotaryPosEmbed(nn.Module):
         self.patch_size_t = patch_size_t
         self.theta = theta
 
-    def forward(
+    def _prepare_video_coords(
         self,
-        hidden_states: torch.Tensor,
+        batch_size: int,
         num_frames: int,
         height: int,
         width: int,
-        rope_interpolation_scale: Optional[Tuple[torch.Tensor, float, float]] = None,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        batch_size = hidden_states.size(0)
-
+        rope_interpolation_scale: Tuple[torch.Tensor, float, float],
+        device: torch.device,
+    ) -> torch.Tensor:
         # Always compute rope in fp32
-        grid_h = torch.arange(height, dtype=torch.float32, device=hidden_states.device)
-        grid_w = torch.arange(width, dtype=torch.float32, device=hidden_states.device)
-        grid_f = torch.arange(num_frames, dtype=torch.float32, device=hidden_states.device)
+        grid_h = torch.arange(height, dtype=torch.float32, device=device)
+        grid_w = torch.arange(width, dtype=torch.float32, device=device)
+        grid_f = torch.arange(num_frames, dtype=torch.float32, device=device)
         grid = torch.meshgrid(grid_f, grid_h, grid_w, indexing="ij")
         grid = torch.stack(grid, dim=0)
         grid = grid.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
@@ -138,6 +137,38 @@ class LTXVideoRotaryPosEmbed(nn.Module):
 
         grid = grid.flatten(2, 4).transpose(1, 2)
 
+        return grid
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        num_frames: Optional[int] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        rope_interpolation_scale: Optional[Tuple[torch.Tensor, float, float]] = None,
+        video_coords: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = hidden_states.size(0)
+
+        if video_coords is None:
+            grid = self._prepare_video_coords(
+                batch_size,
+                num_frames,
+                height,
+                width,
+                rope_interpolation_scale=rope_interpolation_scale,
+                device=hidden_states.device,
+            )
+        else:
+            grid = torch.stack(
+                [
+                    video_coords[:, 0] / self.base_num_frames,
+                    video_coords[:, 1] / self.base_height,
+                    video_coords[:, 2] / self.base_width,
+                ],
+                dim=-1,
+            )
+
         start = 1.0
         end = self.theta
         freqs = self.theta ** torch.linspace(
@@ -367,10 +398,11 @@ class LTXVideoTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin
         encoder_hidden_states: torch.Tensor,
         timestep: torch.LongTensor,
         encoder_attention_mask: torch.Tensor,
-        num_frames: int,
-        height: int,
-        width: int,
-        rope_interpolation_scale: Optional[Tuple[float, float, float]] = None,
+        num_frames: Optional[int] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        rope_interpolation_scale: Optional[Union[Tuple[float, float, float], torch.Tensor]] = None,
+        video_coords: Optional[torch.Tensor] = None,
         attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
     ) -> torch.Tensor:
@@ -389,7 +421,7 @@ class LTXVideoTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin
                     "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
                 )
 
-        image_rotary_emb = self.rope(hidden_states, num_frames, height, width, rope_interpolation_scale)
+        image_rotary_emb = self.rope(hidden_states, num_frames, height, width, rope_interpolation_scale, video_coords)
 
         # convert encoder_attention_mask to a bias the same way we do for attention_mask
         if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:

src/diffusers/models/transformers/transformer_wan.py

@@ -441,6 +441,14 @@ class WanTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrigi
 
         # 5. Output norm, projection & unpatchify
         shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
         hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
         hidden_states = self.proj_out(hidden_states)
 

src/diffusers/models/unets/unet_3d_condition.py

@@ -638,8 +638,10 @@ class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
         t_emb = t_emb.to(dtype=self.dtype)
 
         emb = self.time_embedding(t_emb, timestep_cond)
-        emb = emb.repeat_interleave(repeats=num_frames, dim=0)
-        encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
+        emb = emb.repeat_interleave(num_frames, dim=0, output_size=emb.shape[0] * num_frames)
+        encoder_hidden_states = encoder_hidden_states.repeat_interleave(
+            num_frames, dim=0, output_size=encoder_hidden_states.shape[0] * num_frames
+        )
 
         # 2. pre-process
         sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])

src/diffusers/models/unets/unet_i2vgen_xl.py

@@ -592,7 +592,7 @@ class I2VGenXLUNet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
 
         # 3. time + FPS embeddings.
         emb = t_emb + fps_emb
-        emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+        emb = emb.repeat_interleave(num_frames, dim=0, output_size=emb.shape[0] * num_frames)
 
         # 4. context embeddings.
         # The context embeddings consist of both text embeddings from the input prompt
@@ -620,7 +620,7 @@ class I2VGenXLUNet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
         image_emb = self.context_embedding(image_embeddings)
         image_emb = image_emb.view(-1, self.config.in_channels, self.config.cross_attention_dim)
         context_emb = torch.cat([context_emb, image_emb], dim=1)
-        context_emb = context_emb.repeat_interleave(repeats=num_frames, dim=0)
+        context_emb = context_emb.repeat_interleave(num_frames, dim=0, output_size=context_emb.shape[0] * num_frames)
 
         image_latents = image_latents.permute(0, 2, 1, 3, 4).reshape(
             image_latents.shape[0] * image_latents.shape[2],

src/diffusers/models/unets/unet_motion_model.py

@@ -2059,7 +2059,7 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
             aug_emb = self.add_embedding(add_embeds)
 
         emb = emb if aug_emb is None else emb + aug_emb
-        emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+        emb = emb.repeat_interleave(num_frames, dim=0, output_size=emb.shape[0] * num_frames)
 
         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:
@@ -2068,7 +2068,10 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
                 )
             image_embeds = added_cond_kwargs.get("image_embeds")
             image_embeds = self.encoder_hid_proj(image_embeds)
-            image_embeds = [image_embed.repeat_interleave(repeats=num_frames, dim=0) for image_embed in image_embeds]
+            image_embeds = [
+                image_embed.repeat_interleave(num_frames, dim=0, output_size=image_embed.shape[0] * num_frames)
+                for image_embed in image_embeds
+            ]
             encoder_hidden_states = (encoder_hidden_states, image_embeds)
 
         # 2. pre-process

src/diffusers/models/unets/unet_spatio_temporal_condition.py

@@ -431,9 +431,11 @@ class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionL
         sample = sample.flatten(0, 1)
         # Repeat the embeddings num_video_frames times
         # emb: [batch, channels] -> [batch * frames, channels]
-        emb = emb.repeat_interleave(num_frames, dim=0)
+        emb = emb.repeat_interleave(num_frames, dim=0, output_size=emb.shape[0] * num_frames)
         # encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
-        encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)
+        encoder_hidden_states = encoder_hidden_states.repeat_interleave(
+            num_frames, dim=0, output_size=encoder_hidden_states.shape[0] * num_frames
+        )
 
         # 2. pre-process
         sample = self.conv_in(sample)

src/diffusers/pipelines/__init__.py

@@ -154,7 +154,7 @@ else:
         "CogVideoXFunControlPipeline",
     ]
     _import_structure["cogview3"] = ["CogView3PlusPipeline"]
-    _import_structure["cogview4"] = ["CogView4Pipeline"]
+    _import_structure["cogview4"] = ["CogView4Pipeline", "CogView4ControlPipeline"]
     _import_structure["consisid"] = ["ConsisIDPipeline"]
     _import_structure["controlnet"].extend(
         [
@@ -264,9 +264,9 @@ else:
         ]
     )
     _import_structure["latte"] = ["LattePipeline"]
-    _import_structure["ltx"] = ["LTXPipeline", "LTXImageToVideoPipeline"]
-    _import_structure["lumina"] = ["LuminaText2ImgPipeline"]
-    _import_structure["lumina2"] = ["Lumina2Text2ImgPipeline"]
+    _import_structure["ltx"] = ["LTXPipeline", "LTXImageToVideoPipeline", "LTXConditionPipeline"]
+    _import_structure["lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"]
+    _import_structure["lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"]
     _import_structure["marigold"].extend(
         [
             "MarigoldDepthPipeline",
@@ -280,7 +280,7 @@ else:
     _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
     _import_structure["pia"] = ["PIAPipeline"]
     _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
-    _import_structure["sana"] = ["SanaPipeline"]
+    _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline"]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
     _import_structure["stable_audio"] = [
@@ -511,7 +511,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogVideoXVideoToVideoPipeline,
         )
         from .cogview3 import CogView3PlusPipeline
-        from .cogview4 import CogView4Pipeline
+        from .cogview4 import CogView4ControlPipeline, CogView4Pipeline
         from .consisid import ConsisIDPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
@@ -618,9 +618,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
         )
-        from .ltx import LTXImageToVideoPipeline, LTXPipeline
-        from .lumina import LuminaText2ImgPipeline
-        from .lumina2 import Lumina2Text2ImgPipeline
+        from .ltx import LTXConditionPipeline, LTXImageToVideoPipeline, LTXPipeline
+        from .lumina import LuminaPipeline, LuminaText2ImgPipeline
+        from .lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline
         from .marigold import (
             MarigoldDepthPipeline,
             MarigoldIntrinsicsPipeline,
@@ -651,7 +651,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
-        from .sana import SanaPipeline
+        from .sana import SanaPipeline, SanaSprintPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

src/diffusers/pipelines/auto_pipeline.py

@@ -22,7 +22,7 @@ from ..models.controlnets import ControlNetUnionModel
 from ..utils import is_sentencepiece_available
 from .aura_flow import AuraFlowPipeline
 from .cogview3 import CogView3PlusPipeline
-from .cogview4 import CogView4Pipeline
+from .cogview4 import CogView4ControlPipeline, CogView4Pipeline
 from .controlnet import (
     StableDiffusionControlNetImg2ImgPipeline,
     StableDiffusionControlNetInpaintPipeline,
@@ -69,8 +69,8 @@ from .kandinsky2_2 import (
 )
 from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
 from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
-from .lumina import LuminaText2ImgPipeline
-from .lumina2 import Lumina2Text2ImgPipeline
+from .lumina import LuminaPipeline
+from .lumina2 import Lumina2Pipeline
 from .pag import (
     HunyuanDiTPAGPipeline,
     PixArtSigmaPAGPipeline,
@@ -141,10 +141,11 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("flux", FluxPipeline),
         ("flux-control", FluxControlPipeline),
         ("flux-controlnet", FluxControlNetPipeline),
-        ("lumina", LuminaText2ImgPipeline),
-        ("lumina2", Lumina2Text2ImgPipeline),
+        ("lumina", LuminaPipeline),
+        ("lumina2", Lumina2Pipeline),
         ("cogview3", CogView3PlusPipeline),
         ("cogview4", CogView4Pipeline),
+        ("cogview4-control", CogView4ControlPipeline),
     ]
 )
 

src/diffusers/pipelines/cogview4/__init__.py

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_cogview4"] = ["CogView4Pipeline"]
+    _import_structure["pipeline_cogview4_control"] = ["CogView4ControlPipeline"]
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
         if not (is_transformers_available() and is_torch_available()):
@@ -31,6 +32,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
     else:
         from .pipeline_cogview4 import CogView4Pipeline
+        from .pipeline_cogview4_control import CogView4ControlPipeline
 else:
     import sys
 

src/diffusers/pipelines/cogview4/pipeline_cogview4.py

@@ -389,14 +389,18 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
     def num_timesteps(self):
         return self._num_timesteps
 
-    @property
-    def interrupt(self):
-        return self._interrupt
-
     @property
     def attention_kwargs(self):
         return self._attention_kwargs
 
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -533,6 +537,7 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
         )
         self._guidance_scale = guidance_scale
         self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
         self._interrupt = False
 
         # Default call parameters
@@ -610,6 +615,7 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
                 if self.interrupt:
                     continue
 
+                self._current_timestep = t
                 latent_model_input = latents.to(transformer_dtype)
 
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
@@ -661,6 +667,8 @@ class CogView4Pipeline(DiffusionPipeline, CogView4LoraLoaderMixin):
                 if XLA_AVAILABLE:
                     xm.mark_step()
 
+        self._current_timestep = None
+
         if not output_type == "latent":
             latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
             image = self.vae.decode(latents, return_dict=False, generator=generator)[0]

src/diffusers/pipelines/cogview4/pipeline_cogview4_control.py

@@ -0,0 +1,734 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, GlmModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...models import AutoencoderKL, CogView4Transformer2DModel
+from ...pipelines.pipeline_utils import DiffusionPipeline
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from .pipeline_output import CogView4PipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import CogView4ControlPipeline
+
+        >>> pipe = CogView4ControlPipeline.from_pretrained("THUDM/CogView4-6B-Control", torch_dtype=torch.bfloat16)
+        >>> control_image = load_image(
+        ...     "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/bird_canny.png"
+        ... )
+        >>> prompt = "A bird in space"
+        >>> image = pipe(prompt, control_image=control_image, height=1024, width=1024, guidance_scale=3.5).images[0]
+        >>> image.save("cogview4-control.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.cogview4.pipeline_cogview4.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    base_shift: float = 0.25,
+    max_shift: float = 0.75,
+) -> float:
+    m = (image_seq_len / base_seq_len) ** 0.5
+    mu = m * max_shift + base_shift
+    return mu
+
+
+# Copied from diffusers.pipelines.cogview4.pipeline_cogview4.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+    accepts_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+
+    if timesteps is not None and sigmas is not None:
+        if not accepts_timesteps and not accepts_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep or sigma schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif timesteps is not None and sigmas is None:
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif timesteps is None and sigmas is not None:
+        if not accepts_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class CogView4ControlPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using CogView4.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`GLMModel`]):
+            Frozen text-encoder. CogView4 uses [glm-4-9b-hf](https://huggingface.co/THUDM/glm-4-9b-hf).
+        tokenizer (`PreTrainedTokenizer`):
+            Tokenizer of class
+            [PreTrainedTokenizer](https://huggingface.co/docs/transformers/main/en/main_classes/tokenizer#transformers.PreTrainedTokenizer).
+        transformer ([`CogView4Transformer2DModel`]):
+            A text conditioned `CogView4Transformer2DModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: GlmModel,
+        vae: AutoencoderKL,
+        transformer: CogView4Transformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Copied from diffusers.pipelines.cogview4.pipeline_cogview4.CogView4Pipeline._get_glm_embeds
+    def _get_glm_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        max_sequence_length: int = 1024,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="longest",  # not use max length
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+        current_length = text_input_ids.shape[1]
+        pad_length = (16 - (current_length % 16)) % 16
+        if pad_length > 0:
+            pad_ids = torch.full(
+                (text_input_ids.shape[0], pad_length),
+                fill_value=self.tokenizer.pad_token_id,
+                dtype=text_input_ids.dtype,
+                device=text_input_ids.device,
+            )
+            text_input_ids = torch.cat([pad_ids, text_input_ids], dim=1)
+        prompt_embeds = self.text_encoder(
+            text_input_ids.to(self.text_encoder.device), output_hidden_states=True
+        ).hidden_states[-2]
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.cogview4.pipeline_cogview4.CogView4Pipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                Number of images that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+            max_sequence_length (`int`, defaults to `1024`):
+                Maximum sequence length in encoded prompt. Can be set to other values but may lead to poorer results.
+        """
+        device = device or self._execution_device
+
+        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_embeds = self._get_glm_embeds(prompt, max_sequence_length, device, dtype)
+
+        seq_len = prompt_embeds.size(1)
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_glm_embeds(negative_prompt, max_sequence_length, device, dtype)
+
+            seq_len = negative_prompt_embeds.size(1)
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        if latents is not None:
+            return latents.to(device)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        if isinstance(image, torch.Tensor):
+            pass
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0, output_size=image.shape[0] * repeat_by)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        control_image: PipelineImageInput = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 5.0,
+        num_images_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        output_type: str = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 1024,
+    ) -> Union[CogView4PipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. If not provided, it is set to 1024.
+            width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. If not provided it is set to 1024.
+            num_inference_steps (`int`, *optional*, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            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*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.pipeline_CogView4.CogView4PipelineOutput`] instead of a plain
+                tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `224`):
+                Maximum sequence length in encoded prompt. Can be set to other values but may lead to poorer results.
+        Examples:
+
+        Returns:
+            [`~pipelines.cogview4.pipeline_CogView4.CogView4PipelineOutput`] or `tuple`:
+            [`~pipelines.cogview4.pipeline_CogView4.CogView4PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = (height, width)
+
+        # Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        # Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            self.do_classifier_free_guidance,
+            num_images_per_prompt=num_images_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        # Prepare latents
+        latent_channels = self.transformer.config.in_channels // 2
+
+        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=self.vae.dtype,
+        )
+        height, width = control_image.shape[-2:]
+
+        vae_shift_factor = 0
+
+        control_image = self.vae.encode(control_image).latent_dist.sample()
+        control_image = (control_image - vae_shift_factor) * self.vae.config.scaling_factor
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # Prepare additional timestep conditions
+        original_size = torch.tensor([original_size], dtype=prompt_embeds.dtype, device=device)
+        target_size = torch.tensor([target_size], dtype=prompt_embeds.dtype, device=device)
+        crops_coords_top_left = torch.tensor([crops_coords_top_left], dtype=prompt_embeds.dtype, device=device)
+
+        original_size = original_size.repeat(batch_size * num_images_per_prompt, 1)
+        target_size = target_size.repeat(batch_size * num_images_per_prompt, 1)
+        crops_coords_top_left = crops_coords_top_left.repeat(batch_size * num_images_per_prompt, 1)
+
+        # Prepare timesteps
+        image_seq_len = ((height // self.vae_scale_factor) * (width // self.vae_scale_factor)) // (
+            self.transformer.config.patch_size**2
+        )
+
+        timesteps = (
+            np.linspace(self.scheduler.config.num_train_timesteps, 1.0, num_inference_steps)
+            if timesteps is None
+            else np.array(timesteps)
+        )
+        timesteps = timesteps.astype(np.int64).astype(np.float32)
+        sigmas = timesteps / self.scheduler.config.num_train_timesteps if sigmas is None else sigmas
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("base_shift", 0.25),
+            self.scheduler.config.get("max_shift", 0.75),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas, mu=mu
+        )
+        self._num_timesteps = len(timesteps)
+        # Denoising loop
+        transformer_dtype = self.transformer.dtype
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = torch.cat([latents, control_image], dim=1).to(transformer_dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred_cond = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    original_size=original_size,
+                    target_size=target_size,
+                    crop_coords=crops_coords_top_left,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        timestep=timestep,
+                        original_size=original_size,
+                        target_size=target_size,
+                        crop_coords=crops_coords_top_left,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+                else:
+                    noise_pred = noise_pred_cond
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, self.scheduler.sigmas[i], callback_kwargs)
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
+            image = self.vae.decode(latents, return_dict=False, generator=generator)[0]
+        else:
+            image = latents
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return CogView4PipelineOutput(images=image)

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint.py

@@ -184,7 +184,14 @@ class StableDiffusionControlNetInpaintPipeline(
     model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
     _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
-    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "control_image"]
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "control_image",
+        "mask",
+        "masked_image_latents",
+    ]
 
     def __init__(
         self,

src/diffusers/pipelines/flux/pipeline_flux_control_img2img.py

@@ -533,7 +533,6 @@ class FluxControlImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSin
 
         return latents
 
-    # Copied from diffusers.pipelines.flux.pipeline_flux_img2img.FluxImg2ImgPipeline.prepare_latents
     def prepare_latents(
         self,
         image,

src/diffusers/pipelines/flux/pipeline_flux_controlnet.py

@@ -63,6 +63,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import FluxControlNetPipeline
         >>> from diffusers import FluxControlNetModel
 
+        >>> base_model = "black-forest-labs/FLUX.1-dev"
         >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny"
         >>> controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)
         >>> pipe = FluxControlNetPipeline.from_pretrained(

src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py

@@ -533,7 +533,6 @@ class FluxControlNetImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, From
 
         return latents
 
-    # Copied from diffusers.pipelines.flux.pipeline_flux_img2img.FluxImg2ImgPipeline.prepare_latents
     def prepare_latents(
         self,
         image,

src/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py

@@ -561,7 +561,6 @@ class FluxControlNetInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, From
 
         return latents
 
-    # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.prepare_latents
     def prepare_latents(
         self,
         image,
@@ -614,7 +613,6 @@ class FluxControlNetInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, From
         latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
         return latents, noise, image_latents, latent_image_ids
 
-    # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.prepare_mask_latents
     def prepare_mask_latents(
         self,
         mask,

src/diffusers/pipelines/flux/pipeline_flux_img2img.py

@@ -225,7 +225,10 @@ class FluxImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFile
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
         # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels
+        )
         self.tokenizer_max_length = (
             self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
         )
@@ -634,7 +637,10 @@ class FluxImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFile
             return latents.to(device=device, dtype=dtype), latent_image_ids
 
         image = image.to(device=device, dtype=dtype)
-        image_latents = self._encode_vae_image(image=image, generator=generator)
+        if image.shape[1] != self.latent_channels:
+            image_latents = self._encode_vae_image(image=image, generator=generator)
+        else:
+            image_latents = image
         if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
             # expand init_latents for batch_size
             additional_image_per_prompt = batch_size // image_latents.shape[0]

src/diffusers/pipelines/flux/pipeline_flux_inpaint.py

@@ -222,11 +222,13 @@ class FluxInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FluxIPAdapterM
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
         # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
         # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
-        latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
+        self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels
+        )
         self.mask_processor = VaeImageProcessor(
             vae_scale_factor=self.vae_scale_factor * 2,
-            vae_latent_channels=latent_channels,
+            vae_latent_channels=self.latent_channels,
             do_normalize=False,
             do_binarize=True,
             do_convert_grayscale=True,
@@ -653,7 +655,10 @@ class FluxInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FluxIPAdapterM
         latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
 
         image = image.to(device=device, dtype=dtype)
-        image_latents = self._encode_vae_image(image=image, generator=generator)
+        if image.shape[1] != self.latent_channels:
+            image_latents = self._encode_vae_image(image=image, generator=generator)
+        else:
+            image_latents = image
 
         if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
             # expand init_latents for batch_size
@@ -710,7 +715,9 @@ class FluxInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FluxIPAdapterM
         else:
             masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator)
 
-        masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+            masked_image_latents = (
+                masked_image_latents - self.vae.config.shift_factor
+            ) * self.vae.config.scaling_factor
 
         # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
         if mask.shape[0] < batch_size:

src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py

@@ -54,6 +54,7 @@ EXAMPLE_DOC_STRING = """
         >>> from diffusers import HunyuanVideoImageToVideoPipeline, HunyuanVideoTransformer3DModel
         >>> from diffusers.utils import load_image, export_to_video
 
+        >>> # Available checkpoints: hunyuanvideo-community/HunyuanVideo-I2V, hunyuanvideo-community/HunyuanVideo-I2V-33ch
         >>> model_id = "hunyuanvideo-community/HunyuanVideo-I2V"
         >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained(
         ...     model_id, subfolder="transformer", torch_dtype=torch.bfloat16
@@ -69,7 +70,12 @@ EXAMPLE_DOC_STRING = """
         ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png"
         ... )
 
-        >>> output = pipe(image=image, prompt=prompt).frames[0]
+        >>> # If using hunyuanvideo-community/HunyuanVideo-I2V
+        >>> output = pipe(image=image, prompt=prompt, guidance_scale=6.0).frames[0]
+
+        >>> # If using hunyuanvideo-community/HunyuanVideo-I2V-33ch
+        >>> output = pipe(image=image, prompt=prompt, guidance_scale=1.0, true_cfg_scale=1.0).frames[0]
+
         >>> export_to_video(output, "output.mp4", fps=15)
         ```
 """
@@ -399,7 +405,8 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
         max_sequence_length: int = 256,
-    ):
+        image_embed_interleave: int = 2,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         if prompt_embeds is None:
             prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
                 image,
@@ -409,6 +416,7 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
                 device=device,
                 dtype=dtype,
                 max_sequence_length=max_sequence_length,
+                image_embed_interleave=image_embed_interleave,
             )
 
         if pooled_prompt_embeds is None:
@@ -433,6 +441,8 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         prompt_embeds=None,
         callback_on_step_end_tensor_inputs=None,
         prompt_template=None,
+        true_cfg_scale=1.0,
+        guidance_scale=1.0,
     ):
         if height % 16 != 0 or width % 16 != 0:
             raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
@@ -471,6 +481,13 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
                     f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}"
                 )
 
+        if true_cfg_scale > 1.0 and guidance_scale > 1.0:
+            logger.warning(
+                "Both `true_cfg_scale` and `guidance_scale` are greater than 1.0. This will result in both "
+                "classifier-free guidance and embedded-guidance to be applied. This is not recommended "
+                "as it may lead to higher memory usage, slower inference and potentially worse results."
+            )
+
     def prepare_latents(
         self,
         image: torch.Tensor,
@@ -483,6 +500,7 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         device: Optional[torch.device] = None,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.Tensor] = None,
+        image_condition_type: str = "latent_concat",
     ) -> torch.Tensor:
         if isinstance(generator, list) and len(generator) != batch_size:
             raise ValueError(
@@ -497,10 +515,11 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         image = image.unsqueeze(2)  # [B, C, 1, H, W]
         if isinstance(generator, list):
             image_latents = [
-                retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i], "argmax")
+                for i in range(batch_size)
             ]
         else:
-            image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator) for img in image]
+            image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator, "argmax") for img in image]
 
         image_latents = torch.cat(image_latents, dim=0).to(dtype) * self.vae_scaling_factor
         image_latents = image_latents.repeat(1, 1, num_latent_frames, 1, 1)
@@ -513,6 +532,9 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         t = torch.tensor([0.999]).to(device=device)
         latents = latents * t + image_latents * (1 - t)
 
+        if image_condition_type == "token_replace":
+            image_latents = image_latents[:, :, :1]
+
         return latents, image_latents
 
     def enable_vae_slicing(self):
@@ -598,6 +620,7 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
         max_sequence_length: int = 256,
+        image_embed_interleave: Optional[int] = None,
     ):
         r"""
         The call function to the pipeline for generation.
@@ -704,12 +727,22 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
             prompt_embeds,
             callback_on_step_end_tensor_inputs,
             prompt_template,
+            true_cfg_scale,
+            guidance_scale,
         )
 
+        image_condition_type = self.transformer.config.image_condition_type
         has_neg_prompt = negative_prompt is not None or (
             negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None
         )
         do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        image_embed_interleave = (
+            image_embed_interleave
+            if image_embed_interleave is not None
+            else (
+                2 if image_condition_type == "latent_concat" else 4 if image_condition_type == "token_replace" else 1
+            )
+        )
 
         self._guidance_scale = guidance_scale
         self._attention_kwargs = attention_kwargs
@@ -729,7 +762,12 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         # 3. Prepare latent variables
         vae_dtype = self.vae.dtype
         image_tensor = self.video_processor.preprocess(image, height, width).to(device, vae_dtype)
-        num_channels_latents = (self.transformer.config.in_channels - 1) // 2
+
+        if image_condition_type == "latent_concat":
+            num_channels_latents = (self.transformer.config.in_channels - 1) // 2
+        elif image_condition_type == "token_replace":
+            num_channels_latents = self.transformer.config.in_channels
+
         latents, image_latents = self.prepare_latents(
             image_tensor,
             batch_size * num_videos_per_prompt,
@@ -741,10 +779,12 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
             device,
             generator,
             latents,
+            image_condition_type,
         )
-        image_latents[:, :, 1:] = 0
-        mask = image_latents.new_ones(image_latents.shape[0], 1, *image_latents.shape[2:])
-        mask[:, :, 1:] = 0
+        if image_condition_type == "latent_concat":
+            image_latents[:, :, 1:] = 0
+            mask = image_latents.new_ones(image_latents.shape[0], 1, *image_latents.shape[2:])
+            mask[:, :, 1:] = 0
 
         # 4. Encode input prompt
         transformer_dtype = self.transformer.dtype
@@ -759,6 +799,7 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
             prompt_attention_mask=prompt_attention_mask,
             device=device,
             max_sequence_length=max_sequence_length,
+            image_embed_interleave=image_embed_interleave,
         )
         prompt_embeds = prompt_embeds.to(transformer_dtype)
         prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)
@@ -782,10 +823,17 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
             negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype)
             negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype)
 
-        # 4. Prepare timesteps
+        # 5. Prepare timesteps
         sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas)
 
+        # 6. Prepare guidance condition
+        guidance = None
+        if self.transformer.config.guidance_embeds:
+            guidance = (
+                torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0
+            )
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         self._num_timesteps = len(timesteps)
@@ -796,16 +844,21 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
                     continue
 
                 self._current_timestep = t
-                latent_model_input = torch.cat([latents, image_latents, mask], dim=1).to(transformer_dtype)
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                 timestep = t.expand(latents.shape[0]).to(latents.dtype)
 
+                if image_condition_type == "latent_concat":
+                    latent_model_input = torch.cat([latents, image_latents, mask], dim=1).to(transformer_dtype)
+                elif image_condition_type == "token_replace":
+                    latent_model_input = torch.cat([image_latents, latents[:, :, 1:]], dim=2).to(transformer_dtype)
+
                 noise_pred = self.transformer(
                     hidden_states=latent_model_input,
                     timestep=timestep,
                     encoder_hidden_states=prompt_embeds,
                     encoder_attention_mask=prompt_attention_mask,
                     pooled_projections=pooled_prompt_embeds,
+                    guidance=guidance,
                     attention_kwargs=attention_kwargs,
                     return_dict=False,
                 )[0]
@@ -817,13 +870,20 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
                         encoder_hidden_states=negative_prompt_embeds,
                         encoder_attention_mask=negative_prompt_attention_mask,
                         pooled_projections=negative_pooled_prompt_embeds,
+                        guidance=guidance,
                         attention_kwargs=attention_kwargs,
                         return_dict=False,
                     )[0]
                     noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
 
                 # compute the previous noisy sample x_t -> x_t-1
-                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+                if image_condition_type == "latent_concat":
+                    latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+                elif image_condition_type == "token_replace":
+                    latents = latents = self.scheduler.step(
+                        noise_pred[:, :, 1:], t, latents[:, :, 1:], return_dict=False
+                    )[0]
+                    latents = torch.cat([image_latents, latents], dim=2)
 
                 if callback_on_step_end is not None:
                     callback_kwargs = {}
@@ -844,12 +904,16 @@ class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoader
         self._current_timestep = None
 
         if not output_type == "latent":
-            latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
+            latents = latents.to(self.vae.dtype) / self.vae_scaling_factor
             video = self.vae.decode(latents, return_dict=False)[0]
-            video = video[:, :, 4:, :, :]
+            if image_condition_type == "latent_concat":
+                video = video[:, :, 4:, :, :]
             video = self.video_processor.postprocess_video(video, output_type=output_type)
         else:
-            video = latents[:, :, 1:, :, :]
+            if image_condition_type == "latent_concat":
+                video = latents[:, :, 1:, :, :]
+            else:
+                video = latents
 
         # Offload all models
         self.maybe_free_model_hooks()

src/diffusers/pipelines/kolors/text_encoder.py

@@ -104,13 +104,6 @@ class RMSNorm(torch.nn.Module):
         return (self.weight * hidden_states).to(input_dtype)
 
 
-def _config_to_kwargs(args):
-    common_kwargs = {
-        "dtype": args.torch_dtype,
-    }
-    return common_kwargs
-
-
 class CoreAttention(torch.nn.Module):
     def __init__(self, config: ChatGLMConfig, layer_number):
         super(CoreAttention, self).__init__()
@@ -314,7 +307,6 @@ class SelfAttention(torch.nn.Module):
             self.qkv_hidden_size,
             bias=config.add_bias_linear or config.add_qkv_bias,
             device=device,
-            **_config_to_kwargs(config),
         )
 
         self.core_attention = CoreAttention(config, self.layer_number)
@@ -325,7 +317,6 @@ class SelfAttention(torch.nn.Module):
             config.hidden_size,
             bias=config.add_bias_linear,
             device=device,
-            **_config_to_kwargs(config),
         )
 
     def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
@@ -449,7 +440,6 @@ class MLP(torch.nn.Module):
             config.ffn_hidden_size * 2,
             bias=self.add_bias,
             device=device,
-            **_config_to_kwargs(config),
         )
 
         def swiglu(x):
@@ -459,9 +449,7 @@ class MLP(torch.nn.Module):
         self.activation_func = swiglu
 
         # Project back to h.
-        self.dense_4h_to_h = nn.Linear(
-            config.ffn_hidden_size, config.hidden_size, bias=self.add_bias, device=device, **_config_to_kwargs(config)
-        )
+        self.dense_4h_to_h = nn.Linear(config.ffn_hidden_size, config.hidden_size, bias=self.add_bias, device=device)
 
     def forward(self, hidden_states):
         # [s, b, 4hp]
@@ -488,18 +476,14 @@ class GLMBlock(torch.nn.Module):
 
         LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
         # Layernorm on the input data.
-        self.input_layernorm = LayerNormFunc(
-            config.hidden_size, eps=config.layernorm_epsilon, device=device, dtype=config.torch_dtype
-        )
+        self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device)
 
         # Self attention.
         self.self_attention = SelfAttention(config, layer_number, device=device)
         self.hidden_dropout = config.hidden_dropout
 
         # Layernorm on the attention output
-        self.post_attention_layernorm = LayerNormFunc(
-            config.hidden_size, eps=config.layernorm_epsilon, device=device, dtype=config.torch_dtype
-        )
+        self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device)
 
         # MLP
         self.mlp = MLP(config, device=device)
@@ -569,9 +553,7 @@ class GLMTransformer(torch.nn.Module):
         if self.post_layer_norm:
             LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
             # Final layer norm before output.
-            self.final_layernorm = LayerNormFunc(
-                config.hidden_size, eps=config.layernorm_epsilon, device=device, dtype=config.torch_dtype
-            )
+            self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device)
 
         self.gradient_checkpointing = False
 
@@ -679,9 +661,7 @@ class Embedding(torch.nn.Module):
 
         self.hidden_size = config.hidden_size
         # Word embeddings (parallel).
-        self.word_embeddings = nn.Embedding(
-            config.padded_vocab_size, self.hidden_size, dtype=config.torch_dtype, device=device
-        )
+        self.word_embeddings = nn.Embedding(config.padded_vocab_size, self.hidden_size, device=device)
         self.fp32_residual_connection = config.fp32_residual_connection
 
     def forward(self, input_ids):
@@ -784,16 +764,13 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
             config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
         )
 
-        self.rotary_pos_emb = RotaryEmbedding(
-            rotary_dim // 2, original_impl=config.original_rope, device=device, dtype=config.torch_dtype
-        )
+        self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device)
         self.encoder = init_method(GLMTransformer, config, **init_kwargs)
         self.output_layer = init_method(
             nn.Linear,
             config.hidden_size,
             config.padded_vocab_size,
             bias=False,
-            dtype=config.torch_dtype,
             **init_kwargs,
         )
         self.pre_seq_len = config.pre_seq_len

src/diffusers/pipelines/latte/pipeline_latte.py

@@ -817,7 +817,7 @@ class LattePipeline(DiffusionPipeline):
 
                 # predict noise model_output
                 noise_pred = self.transformer(
-                    latent_model_input,
+                    hidden_states=latent_model_input,
                     encoder_hidden_states=prompt_embeds,
                     timestep=current_timestep,
                     enable_temporal_attentions=enable_temporal_attentions,

src/diffusers/pipelines/ltx/__init__.py

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_ltx"] = ["LTXPipeline"]
+    _import_structure["pipeline_ltx_condition"] = ["LTXConditionPipeline"]
     _import_structure["pipeline_ltx_image2video"] = ["LTXImageToVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -34,6 +35,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_ltx import LTXPipeline
+        from .pipeline_ltx_condition import LTXConditionPipeline
         from .pipeline_ltx_image2video import LTXImageToVideoPipeline
 
 else:

src/diffusers/pipelines/ltx/pipeline_ltx.py

@@ -694,9 +694,8 @@ class LTXPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixi
         self._num_timesteps = len(timesteps)
 
         # 6. Prepare micro-conditions
-        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
         rope_interpolation_scale = (
-            1 / latent_frame_rate,
+            self.vae_temporal_compression_ratio / frame_rate,
             self.vae_spatial_compression_ratio,
             self.vae_spatial_compression_ratio,
         )

src/diffusers/pipelines/ltx/pipeline_ltx_image2video.py

@@ -764,9 +764,8 @@ class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLo
         self._num_timesteps = len(timesteps)
 
         # 6. Prepare micro-conditions
-        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
         rope_interpolation_scale = (
-            1 / latent_frame_rate,
+            self.vae_temporal_compression_ratio / frame_rate,
             self.vae_spatial_compression_ratio,
             self.vae_spatial_compression_ratio,
         )

src/diffusers/pipelines/lumina/__init__.py

@@ -22,7 +22,7 @@ except OptionalDependencyNotAvailable:
 
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
-    _import_structure["pipeline_lumina"] = ["LuminaText2ImgPipeline"]
+    _import_structure["pipeline_lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -32,7 +32,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
-        from .pipeline_lumina import LuminaText2ImgPipeline
+        from .pipeline_lumina import LuminaPipeline, LuminaText2ImgPipeline
 
 else:
     import sys

src/diffusers/pipelines/lumina/pipeline_lumina.py

@@ -30,6 +30,7 @@ from ...models.transformers.lumina_nextdit2d import LuminaNextDiT2DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
 from ...utils import (
     BACKENDS_MAPPING,
+    deprecate,
     is_bs4_available,
     is_ftfy_available,
     is_torch_xla_available,
@@ -60,11 +61,9 @@ EXAMPLE_DOC_STRING = """
     Examples:
         ```py
         >>> import torch
-        >>> from diffusers import LuminaText2ImgPipeline
+        >>> from diffusers import LuminaPipeline
 
-        >>> pipe = LuminaText2ImgPipeline.from_pretrained(
-        ...     "Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16
-        ... )
+        >>> pipe = LuminaPipeline.from_pretrained("Alpha-VLLM/Lumina-Next-SFT-diffusers", torch_dtype=torch.bfloat16)
         >>> # Enable memory optimizations.
         >>> pipe.enable_model_cpu_offload()
 
@@ -134,7 +133,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class LuminaText2ImgPipeline(DiffusionPipeline):
+class LuminaPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-image generation using Lumina-T2I.
 
@@ -932,3 +931,23 @@ class LuminaText2ImgPipeline(DiffusionPipeline):
             return (image,)
 
         return ImagePipelineOutput(images=image)
+
+
+class LuminaText2ImgPipeline(LuminaPipeline):
+    def __init__(
+        self,
+        transformer: LuminaNextDiT2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: GemmaPreTrainedModel,
+        tokenizer: Union[GemmaTokenizer, GemmaTokenizerFast],
+    ):
+        deprecation_message = "`LuminaText2ImgPipeline` has been renamed to `LuminaPipeline` and will be removed in a future version. Please use `LuminaPipeline` instead."
+        deprecate("diffusers.pipelines.lumina.pipeline_lumina.LuminaText2ImgPipeline", "0.34", deprecation_message)
+        super().__init__(
+            transformer=transformer,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+        )

src/diffusers/pipelines/lumina2/__init__.py

@@ -22,7 +22,7 @@ except OptionalDependencyNotAvailable:
 
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
-    _import_structure["pipeline_lumina2"] = ["Lumina2Text2ImgPipeline"]
+    _import_structure["pipeline_lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -32,7 +32,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
-        from .pipeline_lumina2 import Lumina2Text2ImgPipeline
+        from .pipeline_lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline
 
 else:
     import sys