update

Merge branch 'main' into custom-code-updates
2025-07-28 05:33:00 +02:00 · 2025-07-28 08:32:01 +05:30 · 2025-07-23 10:23:35 -10:00 · 2025-07-23 06:19:40 -10:00 · 2025-07-23 19:42:46 +05:30 · 2025-07-23 17:49:38 +05:30
114 changed files with 12316 additions and 1680 deletions
@@ -13,6 +13,7 @@ on:
      - "src/diffusers/loaders/peft.py"
      - "tests/pipelines/test_pipelines_common.py"
      - "tests/models/test_modeling_common.py"
+      - "examples/**/*.py"
  workflow_dispatch:

 concurrency:
@@ -1,36 +1,39 @@
- sections:
+- title: Get started
+  sections:
  - local: index
-    title: 🧨 Diffusers
+    title: Diffusers
+  - local: installation
+    title: Installation
  - local: quicktour
    title: Quicktour
  - local: stable_diffusion
    title: Effective and efficient diffusion
-  - local: installation
-    title: Installation
-  title: Get started
- sections:
-  - local: tutorials/tutorial_overview
-    title: Overview
-  - local: using-diffusers/write_own_pipeline
-    title: Understanding pipelines, models and schedulers
-  - local: tutorials/autopipeline
-    title: AutoPipeline
-  - local: tutorials/basic_training
-    title: Train a diffusion model
-  title: Tutorials
- sections:
+
+- title: DiffusionPipeline
+  isExpanded: false
+  sections:
  - local: using-diffusers/loading
    title: Load pipelines
+  - local: tutorials/autopipeline
+    title: AutoPipeline
  - local: using-diffusers/custom_pipeline_overview
    title: Load community pipelines and components
+  - local: using-diffusers/callback
+    title: Pipeline callbacks
+  - local: using-diffusers/reusing_seeds
+    title: Reproducible pipelines
  - local: using-diffusers/schedulers
    title: Load schedulers and models
+  - local: using-diffusers/scheduler_features
+    title: Scheduler features
  - local: using-diffusers/other-formats
    title: Model files and layouts
  - local: using-diffusers/push_to_hub
    title: Push files to the Hub
-  title: Load pipelines and adapters
- sections:
+
+- title: Adapters
+  isExpanded: false
+  sections:
  - local: tutorials/using_peft_for_inference
    title: LoRA
  - local: using-diffusers/ip_adapter
@@ -43,25 +46,12 @@
    title: DreamBooth
  - local: using-diffusers/textual_inversion_inference
    title: Textual inversion
-  title: Adapters
+
+- title: Inference
  isExpanded: false
- sections:
-  - local: using-diffusers/unconditional_image_generation
-    title: Unconditional image generation
-  - local: using-diffusers/conditional_image_generation
-    title: Text-to-image
-  - local: using-diffusers/img2img
-    title: Image-to-image
-  - local: using-diffusers/inpaint
-    title: Inpainting
-  - local: using-diffusers/text-img2vid
-    title: Video generation
-  - local: using-diffusers/depth2img
-    title: Depth-to-image
-  title: Generative tasks
- sections:
-  - local: using-diffusers/overview_techniques
-    title: Overview
+  sections:
+  - local: using-diffusers/weighted_prompts
+    title: Prompt techniques
  - local: using-diffusers/create_a_server
    title: Create a server
  - local: using-diffusers/batched_inference
@@ -76,14 +66,38 @@
    title: Reproducible pipelines
  - local: using-diffusers/image_quality
    title: Controlling image quality
-  - local: using-diffusers/weighted_prompts
-    title: Prompt techniques
-  title: Inference techniques
- sections:
-  - local: advanced_inference/outpaint
-    title: Outpainting
-  title: Advanced inference
- sections:
+
+- title: Inference optimization
+  isExpanded: false
+  sections:
+  - local: optimization/fp16
+    title: Accelerate inference
+  - local: optimization/cache
+    title: Caching
+  - local: optimization/memory
+    title: Reduce memory usage
+  - local: optimization/speed-memory-optims
+    title: Compile and offloading quantized models
+  - title: Community optimizations
+    sections:
+    - local: optimization/pruna
+      title: Pruna
+    - local: optimization/xformers
+      title: xFormers
+    - local: optimization/tome
+      title: Token merging
+    - local: optimization/deepcache
+      title: DeepCache
+    - local: optimization/tgate
+      title: TGATE
+    - local: optimization/xdit
+      title: xDiT
+    - local: optimization/para_attn
+      title: ParaAttention
+
+- title: Hybrid Inference
+  isExpanded: false
+  sections:
  - local: hybrid_inference/overview
    title: Overview
  - local: hybrid_inference/vae_decode
@@ -92,8 +106,10 @@
    title: VAE Encode
  - local: hybrid_inference/api_reference
    title: API Reference
-  title: Hybrid Inference
- sections:
+
+- title: Modular Diffusers
+  isExpanded: false
+  sections:
  - local: modular_diffusers/overview
    title: Overview
  - local: modular_diffusers/modular_pipeline
@@ -112,8 +128,88 @@
    title: Auto Pipeline Blocks
  - local: modular_diffusers/end_to_end_guide
    title: End-to-End Example
-  title: Modular Diffusers
- sections:
+
+- title: Training
+  isExpanded: false
+  sections:
+  - local: training/overview
+    title: Overview
+  - local: training/create_dataset
+    title: Create a dataset for training
+  - local: training/adapt_a_model
+    title: Adapt a model to a new task
+  - local: tutorials/basic_training
+    title: Train a diffusion model
+  - title: Models
+    sections:
+    - local: training/unconditional_training
+      title: Unconditional image generation
+    - local: training/text2image
+      title: Text-to-image
+    - local: training/sdxl
+      title: Stable Diffusion XL
+    - local: training/kandinsky
+      title: Kandinsky 2.2
+    - local: training/wuerstchen
+      title: Wuerstchen
+    - local: training/controlnet
+      title: ControlNet
+    - local: training/t2i_adapters
+      title: T2I-Adapters
+    - local: training/instructpix2pix
+      title: InstructPix2Pix
+    - local: training/cogvideox
+      title: CogVideoX
+  - title: Methods
+    sections:
+    - local: training/text_inversion
+      title: Textual Inversion
+    - local: training/dreambooth
+      title: DreamBooth
+    - local: training/lora
+      title: LoRA
+    - local: training/custom_diffusion
+      title: Custom Diffusion
+    - local: training/lcm_distill
+      title: Latent Consistency Distillation
+    - local: training/ddpo
+      title: Reinforcement learning training with DDPO
+
+- title: Quantization
+  isExpanded: false
+  sections:
+  - local: quantization/overview
+    title: Getting Started
+  - local: quantization/bitsandbytes
+    title: bitsandbytes
+  - local: quantization/gguf
+    title: gguf
+  - local: quantization/torchao
+    title: torchao
+  - local: quantization/quanto
+    title: quanto
+
+- title: Model accelerators and hardware
+  isExpanded: false
+  sections:
+  - local: using-diffusers/stable_diffusion_jax_how_to
+    title: JAX/Flax
+  - local: optimization/onnx
+    title: ONNX
+  - local: optimization/open_vino
+    title: OpenVINO
+  - local: optimization/coreml
+    title: Core ML
+  - local: optimization/mps
+    title: Metal Performance Shaders (MPS)
+  - local: optimization/habana
+    title: Intel Gaudi
+  - local: optimization/neuron
+    title: AWS Neuron
+
+- title: Specific pipeline examples
+  isExpanded: false
+  sections:
  - local: using-diffusers/consisid
    title: ConsisID
  - local: using-diffusers/sdxl
@@ -138,106 +234,30 @@
    title: Stable Video Diffusion
  - local: using-diffusers/marigold_usage
    title: Marigold Computer Vision
-  title: Specific pipeline examples
- sections:
-  - local: training/overview
-    title: Overview
-  - local: training/create_dataset
-    title: Create a dataset for training
-  - local: training/adapt_a_model
-    title: Adapt a model to a new task
-  - isExpanded: false
+
+- title: Resources
+  isExpanded: false
+  sections:
+  - title: Task recipes
    sections:
-    - local: training/unconditional_training
+    - local: using-diffusers/unconditional_image_generation
      title: Unconditional image generation
-    - local: training/text2image
+    - local: using-diffusers/conditional_image_generation
      title: Text-to-image
-    - local: training/sdxl
-      title: Stable Diffusion XL
-    - local: training/kandinsky
-      title: Kandinsky 2.2
-    - local: training/wuerstchen
-      title: Wuerstchen
-    - local: training/controlnet
-      title: ControlNet
-    - local: training/t2i_adapters
-      title: T2I-Adapters
-    - local: training/instructpix2pix
-      title: InstructPix2Pix
-    - local: training/cogvideox
-      title: CogVideoX
-    title: Models
-  - isExpanded: false
-    sections:
-    - local: training/text_inversion
-      title: Textual Inversion
-    - local: training/dreambooth
-      title: DreamBooth
-    - local: training/lora
-      title: LoRA
-    - local: training/custom_diffusion
-      title: Custom Diffusion
-    - local: training/lcm_distill
-      title: Latent Consistency Distillation
-    - local: training/ddpo
-      title: Reinforcement learning training with DDPO
-    title: Methods
-  title: Training
- sections:
-  - local: quantization/overview
-    title: Getting Started
-  - local: quantization/bitsandbytes
-    title: bitsandbytes
-  - local: quantization/gguf
-    title: gguf
-  - local: quantization/torchao
-    title: torchao
-  - local: quantization/quanto
-    title: quanto
-  title: Quantization Methods
- sections:
-  - local: optimization/fp16
-    title: Accelerate inference
-  - local: optimization/cache
-    title: Caching
-  - local: optimization/memory
-    title: Reduce memory usage
-  - local: optimization/speed-memory-optims
-    title: Compile and offloading quantized models
-  - local: optimization/pruna
-    title: Pruna
-  - local: optimization/xformers
-    title: xFormers
-  - local: optimization/tome
-    title: Token merging
-  - local: optimization/deepcache
-    title: DeepCache
-  - local: optimization/tgate
-    title: TGATE
-  - local: optimization/xdit
-    title: xDiT
-  - local: optimization/para_attn
-    title: ParaAttention
-  - sections:
-    - local: using-diffusers/stable_diffusion_jax_how_to
-      title: JAX/Flax
-    - local: optimization/onnx
-      title: ONNX
-    - local: optimization/open_vino
-      title: OpenVINO
-    - local: optimization/coreml
-      title: Core ML
-    title: Optimized model formats
-  - sections:
-    - local: optimization/mps
-      title: Metal Performance Shaders (MPS)
-    - local: optimization/habana
-      title: Intel Gaudi
-    - local: optimization/neuron
-      title: AWS Neuron
-    title: Optimized hardware
-  title: Accelerate inference and reduce memory
- sections:
+    - local: using-diffusers/img2img
+      title: Image-to-image
+    - local: using-diffusers/inpaint
+      title: Inpainting
+    - local: advanced_inference/outpaint
+      title: Outpainting
+    - local: using-diffusers/text-img2vid
+      title: Video generation
+    - local: using-diffusers/depth2img
+      title: Depth-to-image
+  - local: using-diffusers/write_own_pipeline
+    title: Understanding pipelines, models and schedulers
+  - local: community_projects
+    title: Projects built with Diffusers
  - local: conceptual/philosophy
    title: Philosophy
  - local: using-diffusers/controlling_generation
@@ -248,13 +268,11 @@
    title: Diffusers' Ethical Guidelines
  - local: conceptual/evaluation
    title: Evaluating Diffusion Models
-  title: Conceptual Guides
- sections:
-  - local: community_projects
-    title: Projects built with Diffusers
-  title: Community Projects
- sections:
-  - isExpanded: false
+
+- title: API
+  isExpanded: false
+  sections:
+  - title: Main Classes
    sections:
    - local: api/configuration
      title: Configuration
@@ -264,8 +282,7 @@
      title: Outputs
    - local: api/quantization
      title: Quantization
-    title: Main Classes
-  - isExpanded: false
+  - title: Loaders
    sections:
    - local: api/loaders/ip_adapter
      title: IP-Adapter
@@ -281,14 +298,14 @@
      title: SD3Transformer2D
    - local: api/loaders/peft
      title: PEFT
-    title: Loaders
-  - isExpanded: false
+  - title: Models
    sections:
    - local: api/models/overview
      title: Overview
    - local: api/models/auto_model
      title: AutoModel
-    - sections:
+    - title: ControlNets
+      sections:
      - local: api/models/controlnet
        title: ControlNetModel
      - local: api/models/controlnet_union
@@ -303,8 +320,8 @@
        title: SD3ControlNetModel
      - local: api/models/controlnet_sparsectrl
        title: SparseControlNetModel
-      title: ControlNets
-    - sections:
+    - title: Transformers
+      sections:
      - local: api/models/allegro_transformer3d
        title: AllegroTransformer3DModel
      - local: api/models/aura_flow_transformer2d
@@ -353,6 +370,8 @@
        title: SanaTransformer2DModel
      - local: api/models/sd3_transformer2d
        title: SD3Transformer2DModel
+      - local: api/models/skyreels_v2_transformer_3d
+        title: SkyReelsV2Transformer3DModel
      - local: api/models/stable_audio_transformer
        title: StableAudioDiTModel
      - local: api/models/transformer2d
@@ -361,8 +380,8 @@
        title: TransformerTemporalModel
      - local: api/models/wan_transformer_3d
        title: WanTransformer3DModel
-      title: Transformers
-    - sections:
+    - title: UNets
+      sections:
      - local: api/models/stable_cascade_unet
        title: StableCascadeUNet
      - local: api/models/unet
@@ -377,8 +396,8 @@
        title: UNetMotionModel
      - local: api/models/uvit2d
        title: UViT2DModel
-      title: UNets
-    - sections:
+    - title: VAEs
+      sections:
      - local: api/models/asymmetricautoencoderkl
        title: AsymmetricAutoencoderKL
      - local: api/models/autoencoder_dc
@@ -409,9 +428,7 @@
        title: Tiny AutoEncoder
      - local: api/models/vq
        title: VQModel
-      title: VAEs
-    title: Models
-  - isExpanded: false
+  - title: Pipelines
    sections:
    - local: api/pipelines/overview
      title: Overview
@@ -547,11 +564,14 @@
      title: Semantic Guidance
    - local: api/pipelines/shap_e
      title: Shap-E
+    - local: api/pipelines/skyreels_v2
+      title: SkyReels-V2
    - local: api/pipelines/stable_audio
      title: Stable Audio
    - local: api/pipelines/stable_cascade
      title: Stable Cascade
-    - sections:
+    - title: Stable Diffusion
+      sections:
      - local: api/pipelines/stable_diffusion/overview
        title: Overview
      - local: api/pipelines/stable_diffusion/depth2img
@@ -588,7 +608,6 @@
        title: T2I-Adapter
      - local: api/pipelines/stable_diffusion/text2img
        title: Text-to-image
-      title: Stable Diffusion
    - local: api/pipelines/stable_unclip
      title: Stable unCLIP
    - local: api/pipelines/text_to_video
@@ -607,8 +626,7 @@
      title: Wan
    - local: api/pipelines/wuerstchen
      title: Wuerstchen
-    title: Pipelines
-  - isExpanded: false
+  - title: Schedulers
    sections:
    - local: api/schedulers/overview
      title: Overview
@@ -678,8 +696,7 @@
      title: UniPCMultistepScheduler
    - local: api/schedulers/vq_diffusion
      title: VQDiffusionScheduler
-    title: Schedulers
-  - isExpanded: false
+  - title: Internal classes
    sections:
    - local: api/internal_classes_overview
      title: Overview
@@ -697,5 +714,3 @@
      title: VAE Image Processor
    - local: api/video_processor
      title: Video Processor
-    title: Internal classes
-  title: API
@@ -26,6 +26,7 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`HunyuanVideoLoraLoaderMixin`] provides similar functions for [HunyuanVideo](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hunyuan_video).
 - [`Lumina2LoraLoaderMixin`] provides similar functions for [Lumina2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/lumina2).
 - [`WanLoraLoaderMixin`] provides similar functions for [Wan](https://huggingface.co/docs/diffusers/main/en/api/pipelines/wan).
+- [`SkyReelsV2LoraLoaderMixin`] provides similar functions for [SkyReels-V2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/skyreels_v2).
 - [`CogView4LoraLoaderMixin`] provides similar functions for [CogView4](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogview4).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`HiDreamImageLoraLoaderMixin`] provides similar functions for [HiDream Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/hidream)
@@ -92,6 +93,10 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse

 [[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin

+## SkyReelsV2LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.SkyReelsV2LoraLoaderMixin
+
 ## AmusedLoraLoaderMixin

 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin
@@ -100,6 +105,6 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse

 [[autodoc]] loaders.lora_pipeline.HiDreamImageLoraLoaderMixin

-## WanLoraLoaderMixin
+## LoraBaseMixin

-[[autodoc]] loaders.lora_pipeline.WanLoraLoaderMixin
+[[autodoc]] loaders.lora_base.LoraBaseMixin
@@ -0,0 +1,30 @@
+<!-- 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. -->
+
+# SkyReelsV2Transformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2) by the Skywork AI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import SkyReelsV2Transformer3DModel
+
+transformer = SkyReelsV2Transformer3DModel.from_pretrained("Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## SkyReelsV2Transformer3DModel
+
+[[autodoc]] SkyReelsV2Transformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
@@ -0,0 +1,367 @@
+<!-- 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. -->
+
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
+</div>
+
+# SkyReels-V2: Infinite-length Film Generative model
+
+[SkyReels-V2](https://huggingface.co/papers/2504.13074) by the SkyReels Team.
+
+*Recent advances in video generation have been driven by diffusion models and autoregressive frameworks, yet critical challenges persist in harmonizing prompt adherence, visual quality, motion dynamics, and duration: compromises in motion dynamics to enhance temporal visual quality, constrained video duration (5-10 seconds) to prioritize resolution, and inadequate shot-aware generation stemming from general-purpose MLLMs' inability to interpret cinematic grammar, such as shot composition, actor expressions, and camera motions. These intertwined limitations hinder realistic long-form synthesis and professional film-style generation. To address these limitations, we propose SkyReels-V2, an Infinite-length Film Generative Model, that synergizes Multi-modal Large Language Model (MLLM), Multi-stage Pretraining, Reinforcement Learning, and Diffusion Forcing Framework. Firstly, we design a comprehensive structural representation of video that combines the general descriptions by the Multi-modal LLM and the detailed shot language by sub-expert models. Aided with human annotation, we then train a unified Video Captioner, named SkyCaptioner-V1, to efficiently label the video data. Secondly, we establish progressive-resolution pretraining for the fundamental video generation, followed by a four-stage post-training enhancement: Initial concept-balanced Supervised Fine-Tuning (SFT) improves baseline quality; Motion-specific Reinforcement Learning (RL) training with human-annotated and synthetic distortion data addresses dynamic artifacts; Our diffusion forcing framework with non-decreasing noise schedules enables long-video synthesis in an efficient search space; Final high-quality SFT refines visual fidelity. All the code and models are available at [this https URL](https://github.com/SkyworkAI/SkyReels-V2).*
+
+You can find all the original SkyReels-V2 checkpoints under the [Skywork](https://huggingface.co/collections/Skywork/skyreels-v2-6801b1b93df627d441d0d0d9) organization.
+
+The following SkyReels-V2 models are supported in Diffusers:
+- [SkyReels-V2 DF 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers)
+- [SkyReels-V2 DF 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-DF-14B-540P-Diffusers)
+- [SkyReels-V2 DF 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-DF-14B-720P-Diffusers)
+- [SkyReels-V2 T2V 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-T2V-14B-540P-Diffusers)
+- [SkyReels-V2 T2V 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-T2V-14B-720P-Diffusers)
+- [SkyReels-V2 I2V 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers)
+- [SkyReels-V2 I2V 14B - 540P](https://huggingface.co/Skywork/SkyReels-V2-I2V-14B-540P-Diffusers)
+- [SkyReels-V2 I2V 14B - 720P](https://huggingface.co/Skywork/SkyReels-V2-I2V-14B-720P-Diffusers)
+- [SkyReels-V2 FLF2V 1.3B - 540P](https://huggingface.co/Skywork/SkyReels-V2-FLF2V-1.3B-540P-Diffusers)
+
+> [!TIP]
+> Click on the SkyReels-V2 models in the right sidebar for more examples of video generation.
+
+### A _Visual_ Demonstration
+
+        An example with these parameters:
+        base_num_frames=97, num_frames=97, num_inference_steps=30, ar_step=5, causal_block_size=5
+
+        vae_scale_factor_temporal -> 4
+        num_latent_frames: (97-1)//vae_scale_factor_temporal+1 = 25 frames -> 5 blocks of 5 frames each
+
+        base_num_latent_frames = (97-1)//vae_scale_factor_temporal+1 = 25 → blocks = 25//5 = 5 blocks
+        This 5 blocks means the maximum context length of the model is 25 frames in the latent space.
+
+        Asynchronous Processing Timeline:
+        ┌─────────────────────────────────────────────────────────────────┐
+        │ Steps:    1    6   11   16   21   26   31   36   41   46   50   │
+        │ Block 1: [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]                       │
+        │ Block 2:      [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]                  │
+        │ Block 3:           [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]             │
+        │ Block 4:                [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]        │
+        │ Block 5:                     [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■]   │
+        └─────────────────────────────────────────────────────────────────┘
+
+        For Long Videos (num_frames > base_num_frames):
+        base_num_frames acts as the "sliding window size" for processing long videos.
+
+        Example: 257-frame video with base_num_frames=97, overlap_history=17
+        ┌──── Iteration 1 (frames 1-97) ────┐
+        │ Processing window: 97 frames      │ → 5 blocks, async processing
+        │ Generates: frames 1-97            │
+        └───────────────────────────────────┘
+                    ┌────── Iteration 2 (frames 81-177) ──────┐
+                    │ Processing window: 97 frames            │
+                    │ Overlap: 17 frames (81-97) from prev    │ → 5 blocks, async processing
+                    │ Generates: frames 98-177                │
+                    └─────────────────────────────────────────┘
+                                ┌────── Iteration 3 (frames 161-257) ──────┐
+                                │ Processing window: 97 frames             │
+                                │ Overlap: 17 frames (161-177) from prev   │ → 5 blocks, async processing
+                                │ Generates: frames 178-257                │
+                                └──────────────────────────────────────────┘
+
+        Each iteration independently runs the asynchronous processing with its own 5 blocks.
+        base_num_frames controls:
+        1. Memory usage (larger window = more VRAM)
+        2. Model context length (must match training constraints)
+        3. Number of blocks per iteration (base_num_latent_frames // causal_block_size)
+
+        Each block takes 30 steps to complete denoising.
+        Block N starts at step: 1 + (N-1) x ar_step
+        Total steps: 30 + (5-1) x 5 = 50 steps
+
+
+        Synchronous mode (ar_step=0) would process all blocks/frames simultaneously:
+        ┌──────────────────────────────────────────────┐
+        │ Steps:       1            ...            30  │
+        │ All blocks: [■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■] │
+        └──────────────────────────────────────────────┘
+        Total steps: 30 steps
+
+
+        An example on how the step matrix is constructed for asynchronous processing:
+        Given the parameters: (num_inference_steps=30, flow_shift=8, num_frames=97, ar_step=5, causal_block_size=5)
+        - num_latent_frames = (97 frames - 1) // (4 temporal downsampling) + 1 = 25
+        - step_template = [999, 995, 991, 986, 980, 975, 969, 963, 956, 948,
+                           941, 932, 922, 912, 901, 888, 874, 859, 841, 822,
+                           799, 773, 743, 708, 666, 615, 551, 470, 363, 216]
+
+        The algorithm creates a 50x25 step_matrix where:
+        - Row 1:  [999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - Row 2:  [995, 995, 995, 995, 995, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - Row 3:  [991, 991, 991, 991, 991, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - ...
+        - Row 7:  [969, 969, 969, 969, 969, 995, 995, 995, 995, 995, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999, 999]
+        - ...
+        - Row 21: [799, 799, 799, 799, 799, 888, 888, 888, 888, 888, 941, 941, 941, 941, 941, 975, 975, 975, 975, 975, 999, 999, 999, 999, 999]
+        - ...
+        - Row 35: [  0,   0,   0,   0,   0, 216, 216, 216, 216, 216, 666, 666, 666, 666, 666, 822, 822, 822, 822, 822, 901, 901, 901, 901, 901]
+        - ...
+        - Row 42: [  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0, 551, 551, 551, 551, 551, 773, 773, 773, 773, 773]
+        - ...
+        - Row 50: [  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0, 216, 216, 216, 216, 216]
+
+        Detailed Row 6 Analysis:
+        - step_matrix[5]:       [ 975, 975, 975, 975, 975, 999, 999, 999, 999, 999, 999,  ...,  999]
+        - step_index[5]:        [   6,   6,   6,   6,   6,   1,   1,   1,   1,   1,   0,  ...,    0]
+        - step_update_mask[5]:  [True,True,True,True,True,True,True,True,True,True,False, ...,False]
+        - valid_interval[5]:    (0, 25)
+
+        Key Pattern: Block i lags behind Block i-1 by exactly ar_step=5 timesteps, creating the
+        staggered "diffusion forcing" effect where later blocks condition on cleaner earlier blocks.
+
+### Text-to-Video Generation
+
+The example below demonstrates how to generate a video from text.
+
+<hfoptions id="T2V usage">
+<hfoption id="T2V memory">
+
+Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
+
+From the original repo:
+>You can use --ar_step 5 to enable asynchronous inference. When asynchronous inference, --causal_block_size 5 is recommended while it is not supposed to be set for synchronous generation... Asynchronous inference will take more steps to diffuse the whole sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous inference may improve the instruction following and visual consistent performance.
+
+```py
+# pip install ftfy
+import torch
+from diffusers import AutoModel, SkyReelsV2DiffusionForcingPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video
+
+vae = AutoModel.from_pretrained("Skywork/SkyReels-V2-DF-14B-540P-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+transformer = AutoModel.from_pretrained("Skywork/SkyReels-V2-DF-14B-540P-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+
+pipeline = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
+    "Skywork/SkyReels-V2-DF-14B-540P-Diffusers",
+    vae=vae,
+    transformer=transformer,
+    torch_dtype=torch.bfloat16
+)
+flow_shift = 8.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline = pipeline.to("cuda")
+
+prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+
+output = pipeline(
+    prompt=prompt,
+    num_inference_steps=30,
+    height=544,  # 720 for 720P
+    width=960,   # 1280 for 720P
+    num_frames=97,
+    base_num_frames=97,  # 121 for 720P
+    ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
+    causal_block_size=5,  # Number of frames in each block for asynchronous processing
+    overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos; 17 for long video generations
+    addnoise_condition=20,  # Improves consistency in long video generation
+).frames[0]
+export_to_video(output, "T2V.mp4", fps=24, quality=8)
+```
+
+</hfoption>
+</hfoptions>
+
+### First-Last-Frame-to-Video Generation
+
+The example below demonstrates how to use the image-to-video pipeline to generate a video using a text description, a starting frame, and an ending frame.
+
+<hfoptions id="FLF2V usage">
+<hfoption id="usage">
+
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingImageToVideoPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video, load_image
+
+
+model_id = "Skywork/SkyReels-V2-DF-14B-720P-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipeline = SkyReelsV2DiffusionForcingImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 5.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline.to("cuda")
+
+first_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png")
+last_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png")
+
+def aspect_ratio_resize(image, pipeline, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipeline.vae_scale_factor_spatial * pipeline.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+
+def center_crop_resize(image, height, width):
+    # Calculate resize ratio to match first frame dimensions
+    resize_ratio = max(width / image.width, height / image.height)
+
+    # Resize the image
+    width = round(image.width * resize_ratio)
+    height = round(image.height * resize_ratio)
+    size = [width, height]
+    image = TF.center_crop(image, size)
+
+    return image, height, width
+
+first_frame, height, width = aspect_ratio_resize(first_frame, pipeline)
+if last_frame.size != first_frame.size:
+    last_frame, _, _ = center_crop_resize(last_frame, height, width)
+
+prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+
+output = pipeline(
+    image=first_frame, last_image=last_frame, prompt=prompt, height=height, width=width, guidance_scale=5.0
+).frames[0]
+export_to_video(output, "output.mp4", fps=24, quality=8)
+```
+
+</hfoption>
+</hfoptions>
+
+
+### Video-to-Video Generation
+
+<hfoptions id="V2V usage">
+<hfoption id="usage">
+
+`SkyReelsV2DiffusionForcingVideoToVideoPipeline` extends a given video.
+
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingVideoToVideoPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video, load_video
+
+
+model_id = "Skywork/SkyReels-V2-DF-14B-540P-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipeline = SkyReelsV2DiffusionForcingVideoToVideoPipeline.from_pretrained(
+    model_id, vae=vae, torch_dtype=torch.bfloat16
+)
+flow_shift = 5.0  # 8.0 for T2V, 5.0 for I2V
+pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, flow_shift=flow_shift)
+pipeline.to("cuda")
+
+video = load_video("input_video.mp4")
+
+prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+
+output = pipeline(
+    video=video, prompt=prompt, height=544, width=960, guidance_scale=5.0,
+    num_inference_steps=30, num_frames=257, base_num_frames=97#, ar_step=5, causal_block_size=5,
+).frames[0]
+export_to_video(output, "output.mp4", fps=24, quality=8)
+# Total frames will be the number of frames of given video + 257
+```
+
+</hfoption>
+</hfoptions>
+
+
+## Notes
+
+- SkyReels-V2 supports LoRAs with [`~loaders.SkyReelsV2LoraLoaderMixin.load_lora_weights`].
+
+  <details>
+  <summary>Show example code</summary>
+
+  ```py
+  # pip install ftfy
+  import torch
+  from diffusers import AutoModel, SkyReelsV2DiffusionForcingPipeline
+  from diffusers.utils import export_to_video
+
+  vae = AutoModel.from_pretrained(
+      "Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", subfolder="vae", torch_dtype=torch.float32
+  )
+  pipeline = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
+      "Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers", vae=vae, torch_dtype=torch.bfloat16
+  )
+  pipeline.to("cuda")
+
+  pipeline.load_lora_weights("benjamin-paine/steamboat-willie-1.3b", adapter_name="steamboat-willie")
+  pipeline.set_adapters("steamboat-willie")
+
+  pipeline.enable_model_cpu_offload()
+
+  # use "steamboat willie style" to trigger the LoRA
+  prompt = """
+  steamboat willie style, golden era animation, The camera rushes from far to near in a low-angle shot,
+  revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in
+  for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground.
+  Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic
+  shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
+  """
+
+  output = pipeline(
+      prompt=prompt,
+      num_frames=97,
+      guidance_scale=6.0,
+  ).frames[0]
+  export_to_video(output, "output.mp4", fps=24)
+  ```
+
+  </details>
+
+
+## SkyReelsV2DiffusionForcingPipeline
+
+[[autodoc]] SkyReelsV2DiffusionForcingPipeline
+  - all
+  - __call__
+
+## SkyReelsV2DiffusionForcingImageToVideoPipeline
+
+[[autodoc]] SkyReelsV2DiffusionForcingImageToVideoPipeline
+  - all
+  - __call__
+
+## SkyReelsV2DiffusionForcingVideoToVideoPipeline
+
+[[autodoc]] SkyReelsV2DiffusionForcingVideoToVideoPipeline
+  - all
+  - __call__
+
+## SkyReelsV2Pipeline
+
+[[autodoc]] SkyReelsV2Pipeline
+  - all
+  - __call__
+
+## SkyReelsV2ImageToVideoPipeline
+
+[[autodoc]] SkyReelsV2ImageToVideoPipeline
+  - all
+  - __call__
+
+## SkyReelsV2PipelineOutput
+
+[[autodoc]] pipelines.skyreels_v2.pipeline_output.SkyReelsV2PipelineOutput
@@ -239,6 +239,12 @@ The `step()` function is [called](https://github.com/huggingface/diffusers/blob/

 In general, the `sigmas` should [stay on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240) to avoid the communication sync and latency.

+<Tip>
+
+Refer to the [torch.compile and Diffusers: A Hands-On Guide to Peak Performance](https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/) blog post for maximizing performance with `torch.compile` for diffusion models.
+
+</Tip>
+
 ### Benchmarks

 Refer to the [diffusers/benchmarks](https://huggingface.co/datasets/diffusers/benchmarks) dataset to see inference latency and memory usage data for compiled pipelines.
@@ -298,4 +304,6 @@ pipeline.fuse_qkv_projections()

 - Read the [Presenting Flux Fast: Making Flux go brrr on H100s](https://pytorch.org/blog/presenting-flux-fast-making-flux-go-brrr-on-h100s/) blog post to learn more about how you can combine all of these optimizations with [TorchInductor](https://docs.pytorch.org/docs/stable/torch.compiler.html) and [AOTInductor](https://docs.pytorch.org/docs/stable/torch.compiler_aot_inductor.html) for a ~2.5x speedup using recipes from [flux-fast](https://github.com/huggingface/flux-fast).

-    These recipes support AMD hardware and [Flux.1 Kontext Dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).
+    These recipes support AMD hardware and [Flux.1 Kontext Dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).
+- Read the [torch.compile and Diffusers: A Hands-On Guide to Peak Performance](https://pytorch.org/blog/torch-compile-and-diffusers-a-hands-on-guide-to-peak-performance/) blog post
+to maximize performance when using `torch.compile`.
@@ -1,23 +0,0 @@
-<!--Copyright 2025 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
-->
-
-# Overview
-
-Welcome to 🧨 Diffusers! If you're new to diffusion models and generative AI, and want to learn more, then you've come to the right place. These beginner-friendly tutorials are designed to provide a gentle introduction to diffusion models and help you understand the library fundamentals - the core components and how 🧨 Diffusers is meant to be used.
-
-You'll learn how to use a pipeline for inference to rapidly generate things, and then deconstruct that pipeline to really understand how to use the library as a modular toolbox for building your own diffusion systems. In the next lesson, you'll learn how to train your own diffusion model to generate what you want.
-
-After completing the tutorials, you'll have gained the necessary skills to start exploring the library on your own and see how to use it for your own projects and applications.
-
-Feel free to join our community on [Discord](https://discord.com/invite/JfAtkvEtRb) or the [forums](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) to connect and collaborate with other users and developers!
-
-Let's start diffusing! 🧨
@@ -1,18 +0,0 @@
-<!--Copyright 2025 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
-->
-
-# Overview
-
-The inference pipeline supports and enables a wide range of techniques that are divided into two categories:
-
-* Pipeline functionality: these techniques modify the pipeline or extend it for other applications. For example, pipeline callbacks add new features to a pipeline and a pipeline can also be extended for distributed inference.
-* Improve inference quality: these techniques increase the visual quality of the generated images. For example, you can enhance your prompts with GPT2 to create better images with lower effort.
@@ -971,6 +971,7 @@ class DreamBoothDataset(Dataset):

    def __init__(
        self,
+        args,
        instance_data_root,
        instance_prompt,
        class_prompt,
@@ -980,10 +981,8 @@ class DreamBoothDataset(Dataset):
        class_num=None,
        size=1024,
        repeats=1,
-        center_crop=False,
    ):
        self.size = size
-        self.center_crop = center_crop

        self.instance_prompt = instance_prompt
        self.custom_instance_prompts = None
@@ -1058,7 +1057,7 @@ class DreamBoothDataset(Dataset):
        if interpolation is None:
            raise ValueError(f"Unsupported interpolation mode {interpolation=}.")
        train_resize = transforms.Resize(size, interpolation=interpolation)
-        train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
+        train_crop = transforms.CenterCrop(size) if args.center_crop else transforms.RandomCrop(size)
        train_flip = transforms.RandomHorizontalFlip(p=1.0)
        train_transforms = transforms.Compose(
            [
@@ -1075,11 +1074,11 @@ class DreamBoothDataset(Dataset):
                # flip
                image = train_flip(image)
            if args.center_crop:
-                y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
-                x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
+                y1 = max(0, int(round((image.height - self.size) / 2.0)))
+                x1 = max(0, int(round((image.width - self.size) / 2.0)))
                image = train_crop(image)
            else:
-                y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
+                y1, x1, h, w = train_crop.get_params(image, (self.size, self.size))
                image = crop(image, y1, x1, h, w)
            image = train_transforms(image)
            self.pixel_values.append(image)
@@ -1102,7 +1101,7 @@ class DreamBoothDataset(Dataset):
        self.image_transforms = transforms.Compose(
            [
                transforms.Resize(size, interpolation=interpolation),
-                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
+                transforms.CenterCrop(size) if args.center_crop else transforms.RandomCrop(size),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
@@ -1827,6 +1826,7 @@ def main(args):

    # Dataset and DataLoaders creation:
    train_dataset = DreamBoothDataset(
+        args=args,
        instance_data_root=args.instance_data_dir,
        instance_prompt=args.instance_prompt,
        train_text_encoder_ti=args.train_text_encoder_ti,
@@ -1836,7 +1836,6 @@ def main(args):
        class_num=args.num_class_images,
        size=args.resolution,
        repeats=args.repeats,
-        center_crop=args.center_crop,
    )

    train_dataloader = torch.utils.data.DataLoader(
@@ -1614,7 +1614,7 @@ def main(args):
    )
    if args.cond_image_column is not None:
        logger.info("I2I fine-tuning enabled.")
-    batch_sampler = BucketBatchSampler(train_dataset, batch_size=args.train_batch_size, drop_last=False)
+    batch_sampler = BucketBatchSampler(train_dataset, batch_size=args.train_batch_size, drop_last=True)
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        batch_sampler=batch_sampler,
@@ -58,6 +58,7 @@ from diffusers.training_utils import (
    compute_density_for_timestep_sampling,
    compute_loss_weighting_for_sd3,
    free_memory,
+    offload_models,
 )
 from diffusers.utils import (
    check_min_version,
@@ -1364,43 +1365,34 @@ def main(args):
    # provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
    # the redundant encoding.
    if not train_dataset.custom_instance_prompts:
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to(accelerator.device)
-        (
-            instance_prompt_hidden_states_t5,
-            instance_prompt_hidden_states_llama3,
-            instance_pooled_prompt_embeds,
-            _,
-            _,
-            _,
-        ) = compute_text_embeddings(args.instance_prompt, text_encoding_pipeline)
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to("cpu")
+        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
+            (
+                instance_prompt_hidden_states_t5,
+                instance_prompt_hidden_states_llama3,
+                instance_pooled_prompt_embeds,
+                _,
+                _,
+                _,
+            ) = compute_text_embeddings(args.instance_prompt, text_encoding_pipeline)

    # Handle class prompt for prior-preservation.
    if args.with_prior_preservation:
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to(accelerator.device)
-        (class_prompt_hidden_states_t5, class_prompt_hidden_states_llama3, class_pooled_prompt_embeds, _, _, _) = (
-            compute_text_embeddings(args.class_prompt, text_encoding_pipeline)
-        )
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to("cpu")
+        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
+            (class_prompt_hidden_states_t5, class_prompt_hidden_states_llama3, class_pooled_prompt_embeds, _, _, _) = (
+                compute_text_embeddings(args.class_prompt, text_encoding_pipeline)
+            )

    validation_embeddings = {}
    if args.validation_prompt is not None:
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to(accelerator.device)
-        (
-            validation_embeddings["prompt_embeds_t5"],
-            validation_embeddings["prompt_embeds_llama3"],
-            validation_embeddings["pooled_prompt_embeds"],
-            validation_embeddings["negative_prompt_embeds_t5"],
-            validation_embeddings["negative_prompt_embeds_llama3"],
-            validation_embeddings["negative_pooled_prompt_embeds"],
-        ) = compute_text_embeddings(args.validation_prompt, text_encoding_pipeline)
-        if args.offload:
-            text_encoding_pipeline = text_encoding_pipeline.to("cpu")
+        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
+            (
+                validation_embeddings["prompt_embeds_t5"],
+                validation_embeddings["prompt_embeds_llama3"],
+                validation_embeddings["pooled_prompt_embeds"],
+                validation_embeddings["negative_prompt_embeds_t5"],
+                validation_embeddings["negative_prompt_embeds_llama3"],
+                validation_embeddings["negative_pooled_prompt_embeds"],
+            ) = compute_text_embeddings(args.validation_prompt, text_encoding_pipeline)

    # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
    # pack the statically computed variables appropriately here. This is so that we don't
@@ -1581,12 +1573,10 @@ def main(args):
                if args.cache_latents:
                    model_input = latents_cache[step].sample()
                else:
-                    if args.offload:
-                        vae = vae.to(accelerator.device)
-                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
+                    with offload_models(vae, device=accelerator.device, offload=args.offload):
+                        pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                    model_input = vae.encode(pixel_values).latent_dist.sample()
-                    if args.offload:
-                        vae = vae.to("cpu")
+
                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                model_input = model_input.to(dtype=weight_dtype)

@@ -0,0 +1,637 @@
+import argparse
+import os
+import pathlib
+from typing import Any, Dict
+
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+from transformers import AutoProcessor, AutoTokenizer, CLIPVisionModelWithProjection, UMT5EncoderModel
+
+from diffusers import (
+    AutoencoderKLWan,
+    SkyReelsV2DiffusionForcingPipeline,
+    SkyReelsV2ImageToVideoPipeline,
+    SkyReelsV2Pipeline,
+    SkyReelsV2Transformer3DModel,
+    UniPCMultistepScheduler,
+)
+
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "time_embedding.0": "condition_embedder.time_embedder.linear_1",
+    "time_embedding.2": "condition_embedder.time_embedder.linear_2",
+    "text_embedding.0": "condition_embedder.text_embedder.linear_1",
+    "text_embedding.2": "condition_embedder.text_embedder.linear_2",
+    "time_projection.1": "condition_embedder.time_proj",
+    "head.modulation": "scale_shift_table",
+    "head.head": "proj_out",
+    "modulation": "scale_shift_table",
+    "ffn.0": "ffn.net.0.proj",
+    "ffn.2": "ffn.net.2",
+    "fps_projection.0": "fps_projection.net.0.proj",
+    "fps_projection.2": "fps_projection.net.2",
+    # Hack to swap the layer names
+    # The original model calls the norms in following order: norm1, norm3, norm2
+    # We convert it to: norm1, norm2, norm3
+    "norm2": "norm__placeholder",
+    "norm3": "norm2",
+    "norm__placeholder": "norm3",
+    # For the I2V model
+    "img_emb.proj.0": "condition_embedder.image_embedder.norm1",
+    "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
+    "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
+    "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+    # for the FLF2V model
+    "img_emb.emb_pos": "condition_embedder.image_embedder.pos_embed",
+    # Add attention component mappings
+    "self_attn.q": "attn1.to_q",
+    "self_attn.k": "attn1.to_k",
+    "self_attn.v": "attn1.to_v",
+    "self_attn.o": "attn1.to_out.0",
+    "self_attn.norm_q": "attn1.norm_q",
+    "self_attn.norm_k": "attn1.norm_k",
+    "cross_attn.q": "attn2.to_q",
+    "cross_attn.k": "attn2.to_k",
+    "cross_attn.v": "attn2.to_v",
+    "cross_attn.o": "attn2.to_out.0",
+    "cross_attn.norm_q": "attn2.norm_q",
+    "cross_attn.norm_k": "attn2.norm_k",
+    "attn2.to_k_img": "attn2.add_k_proj",
+    "attn2.to_v_img": "attn2.add_v_proj",
+    "attn2.norm_k_img": "attn2.norm_added_k",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+
+
+def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def load_sharded_safetensors(dir: pathlib.Path):
+    if "720P" in str(dir):
+        file_paths = list(dir.glob("diffusion_pytorch_model*.safetensors"))
+    else:
+        file_paths = list(dir.glob("model*.safetensors"))
+    state_dict = {}
+    for path in file_paths:
+        state_dict.update(load_file(path))
+    return state_dict
+
+
+def get_transformer_config(model_type: str) -> Dict[str, Any]:
+    if model_type == "SkyReels-V2-DF-1.3B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-DF-1.3B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 12,
+                "inject_sample_info": True,
+                "num_layers": 30,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "SkyReels-V2-DF-14B-720P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-DF-14B-720P",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "SkyReels-V2-DF-14B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-DF-14B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "SkyReels-V2-T2V-14B-720P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-T2V-14B-720P",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "SkyReels-V2-T2V-14B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-T2V-14B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": None,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 16,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+            },
+        }
+    elif model_type == "SkyReels-V2-I2V-1.3B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-1.3B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 1536,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 12,
+                "inject_sample_info": False,
+                "num_layers": 30,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+            },
+        }
+    elif model_type == "SkyReels-V2-I2V-14B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-14B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+            },
+        }
+    elif model_type == "SkyReels-V2-I2V-14B-720P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-14B-720P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+            },
+        }
+    elif model_type == "SkyReels-V2-FLF2V-1.3B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-1.3B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 1536,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 12,
+                "inject_sample_info": False,
+                "num_layers": 30,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+                "pos_embed_seq_len": 514,
+            },
+        }
+    elif model_type == "SkyReels-V2-FLF2V-14B-540P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-14B-540P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+                "pos_embed_seq_len": 514,
+            },
+        }
+    elif model_type == "SkyReels-V2-FLF2V-14B-720P":
+        config = {
+            "model_id": "Skywork/SkyReels-V2-I2V-14B-720P",
+            "diffusers_config": {
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "inject_sample_info": False,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "image_dim": 1280,
+                "pos_embed_seq_len": 514,
+            },
+        }
+    return config
+
+
+def convert_transformer(model_type: str):
+    config = get_transformer_config(model_type)
+    diffusers_config = config["diffusers_config"]
+    model_id = config["model_id"]
+
+    if "1.3B" in model_type:
+        original_state_dict = load_file(hf_hub_download(model_id, "model.safetensors"))
+    else:
+        os.makedirs(model_type, exist_ok=True)
+        model_dir = pathlib.Path(model_type)
+        if "720P" in model_type:
+            top_shard = 7 if "I2V" in model_type else 6
+            zeros = "0" * (4 if "I2V" or "T2V" in model_type else 3)
+            model_name = "diffusion_pytorch_model"
+        elif "540P" in model_type:
+            top_shard = 14 if "I2V" in model_type else 12
+            model_name = "model"
+
+        for i in range(1, top_shard + 1):
+            shard_path = f"{model_name}-{i:05d}-of-{zeros}{top_shard}.safetensors"
+            hf_hub_download(model_id, shard_path, local_dir=model_dir)
+        original_state_dict = load_sharded_safetensors(model_dir)
+
+    with init_empty_weights():
+        transformer = SkyReelsV2Transformer3DModel.from_config(diffusers_config)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    if "FLF2V" in model_type:
+        if (
+            hasattr(transformer.condition_embedder, "image_embedder")
+            and hasattr(transformer.condition_embedder.image_embedder, "pos_embed")
+            and transformer.condition_embedder.image_embedder.pos_embed is not None
+        ):
+            pos_embed_shape = transformer.condition_embedder.image_embedder.pos_embed.shape
+            original_state_dict["condition_embedder.image_embedder.pos_embed"] = torch.zeros(pos_embed_shape)
+
+    transformer.load_state_dict(original_state_dict, strict=True, assign=True)
+    return transformer
+
+
+def convert_vae():
+    vae_ckpt_path = hf_hub_download("Wan-AI/Wan2.1-T2V-14B", "Wan2.1_VAE.pth")
+    old_state_dict = torch.load(vae_ckpt_path, weights_only=True)
+    new_state_dict = {}
+
+    # Create mappings for specific components
+    middle_key_mapping = {
+        # Encoder middle block
+        "encoder.middle.0.residual.0.gamma": "encoder.mid_block.resnets.0.norm1.gamma",
+        "encoder.middle.0.residual.2.bias": "encoder.mid_block.resnets.0.conv1.bias",
+        "encoder.middle.0.residual.2.weight": "encoder.mid_block.resnets.0.conv1.weight",
+        "encoder.middle.0.residual.3.gamma": "encoder.mid_block.resnets.0.norm2.gamma",
+        "encoder.middle.0.residual.6.bias": "encoder.mid_block.resnets.0.conv2.bias",
+        "encoder.middle.0.residual.6.weight": "encoder.mid_block.resnets.0.conv2.weight",
+        "encoder.middle.2.residual.0.gamma": "encoder.mid_block.resnets.1.norm1.gamma",
+        "encoder.middle.2.residual.2.bias": "encoder.mid_block.resnets.1.conv1.bias",
+        "encoder.middle.2.residual.2.weight": "encoder.mid_block.resnets.1.conv1.weight",
+        "encoder.middle.2.residual.3.gamma": "encoder.mid_block.resnets.1.norm2.gamma",
+        "encoder.middle.2.residual.6.bias": "encoder.mid_block.resnets.1.conv2.bias",
+        "encoder.middle.2.residual.6.weight": "encoder.mid_block.resnets.1.conv2.weight",
+        # Decoder middle block
+        "decoder.middle.0.residual.0.gamma": "decoder.mid_block.resnets.0.norm1.gamma",
+        "decoder.middle.0.residual.2.bias": "decoder.mid_block.resnets.0.conv1.bias",
+        "decoder.middle.0.residual.2.weight": "decoder.mid_block.resnets.0.conv1.weight",
+        "decoder.middle.0.residual.3.gamma": "decoder.mid_block.resnets.0.norm2.gamma",
+        "decoder.middle.0.residual.6.bias": "decoder.mid_block.resnets.0.conv2.bias",
+        "decoder.middle.0.residual.6.weight": "decoder.mid_block.resnets.0.conv2.weight",
+        "decoder.middle.2.residual.0.gamma": "decoder.mid_block.resnets.1.norm1.gamma",
+        "decoder.middle.2.residual.2.bias": "decoder.mid_block.resnets.1.conv1.bias",
+        "decoder.middle.2.residual.2.weight": "decoder.mid_block.resnets.1.conv1.weight",
+        "decoder.middle.2.residual.3.gamma": "decoder.mid_block.resnets.1.norm2.gamma",
+        "decoder.middle.2.residual.6.bias": "decoder.mid_block.resnets.1.conv2.bias",
+        "decoder.middle.2.residual.6.weight": "decoder.mid_block.resnets.1.conv2.weight",
+    }
+
+    # Create a mapping for attention blocks
+    attention_mapping = {
+        # Encoder middle attention
+        "encoder.middle.1.norm.gamma": "encoder.mid_block.attentions.0.norm.gamma",
+        "encoder.middle.1.to_qkv.weight": "encoder.mid_block.attentions.0.to_qkv.weight",
+        "encoder.middle.1.to_qkv.bias": "encoder.mid_block.attentions.0.to_qkv.bias",
+        "encoder.middle.1.proj.weight": "encoder.mid_block.attentions.0.proj.weight",
+        "encoder.middle.1.proj.bias": "encoder.mid_block.attentions.0.proj.bias",
+        # Decoder middle attention
+        "decoder.middle.1.norm.gamma": "decoder.mid_block.attentions.0.norm.gamma",
+        "decoder.middle.1.to_qkv.weight": "decoder.mid_block.attentions.0.to_qkv.weight",
+        "decoder.middle.1.to_qkv.bias": "decoder.mid_block.attentions.0.to_qkv.bias",
+        "decoder.middle.1.proj.weight": "decoder.mid_block.attentions.0.proj.weight",
+        "decoder.middle.1.proj.bias": "decoder.mid_block.attentions.0.proj.bias",
+    }
+
+    # Create a mapping for the head components
+    head_mapping = {
+        # Encoder head
+        "encoder.head.0.gamma": "encoder.norm_out.gamma",
+        "encoder.head.2.bias": "encoder.conv_out.bias",
+        "encoder.head.2.weight": "encoder.conv_out.weight",
+        # Decoder head
+        "decoder.head.0.gamma": "decoder.norm_out.gamma",
+        "decoder.head.2.bias": "decoder.conv_out.bias",
+        "decoder.head.2.weight": "decoder.conv_out.weight",
+    }
+
+    # Create a mapping for the quant components
+    quant_mapping = {
+        "conv1.weight": "quant_conv.weight",
+        "conv1.bias": "quant_conv.bias",
+        "conv2.weight": "post_quant_conv.weight",
+        "conv2.bias": "post_quant_conv.bias",
+    }
+
+    # Process each key in the state dict
+    for key, value in old_state_dict.items():
+        # Handle middle block keys using the mapping
+        if key in middle_key_mapping:
+            new_key = middle_key_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle attention blocks using the mapping
+        elif key in attention_mapping:
+            new_key = attention_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle head keys using the mapping
+        elif key in head_mapping:
+            new_key = head_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle quant keys using the mapping
+        elif key in quant_mapping:
+            new_key = quant_mapping[key]
+            new_state_dict[new_key] = value
+        # Handle encoder conv1
+        elif key == "encoder.conv1.weight":
+            new_state_dict["encoder.conv_in.weight"] = value
+        elif key == "encoder.conv1.bias":
+            new_state_dict["encoder.conv_in.bias"] = value
+        # Handle decoder conv1
+        elif key == "decoder.conv1.weight":
+            new_state_dict["decoder.conv_in.weight"] = value
+        elif key == "decoder.conv1.bias":
+            new_state_dict["decoder.conv_in.bias"] = value
+        # Handle encoder downsamples
+        elif key.startswith("encoder.downsamples."):
+            # Convert to down_blocks
+            new_key = key.replace("encoder.downsamples.", "encoder.down_blocks.")
+
+            # Convert residual block naming but keep the original structure
+            if ".residual.0.gamma" in new_key:
+                new_key = new_key.replace(".residual.0.gamma", ".norm1.gamma")
+            elif ".residual.2.bias" in new_key:
+                new_key = new_key.replace(".residual.2.bias", ".conv1.bias")
+            elif ".residual.2.weight" in new_key:
+                new_key = new_key.replace(".residual.2.weight", ".conv1.weight")
+            elif ".residual.3.gamma" in new_key:
+                new_key = new_key.replace(".residual.3.gamma", ".norm2.gamma")
+            elif ".residual.6.bias" in new_key:
+                new_key = new_key.replace(".residual.6.bias", ".conv2.bias")
+            elif ".residual.6.weight" in new_key:
+                new_key = new_key.replace(".residual.6.weight", ".conv2.weight")
+            elif ".shortcut.bias" in new_key:
+                new_key = new_key.replace(".shortcut.bias", ".conv_shortcut.bias")
+            elif ".shortcut.weight" in new_key:
+                new_key = new_key.replace(".shortcut.weight", ".conv_shortcut.weight")
+
+            new_state_dict[new_key] = value
+
+        # Handle decoder upsamples
+        elif key.startswith("decoder.upsamples."):
+            # Convert to up_blocks
+            parts = key.split(".")
+            block_idx = int(parts[2])
+
+            # Group residual blocks
+            if "residual" in key:
+                if block_idx in [0, 1, 2]:
+                    new_block_idx = 0
+                    resnet_idx = block_idx
+                elif block_idx in [4, 5, 6]:
+                    new_block_idx = 1
+                    resnet_idx = block_idx - 4
+                elif block_idx in [8, 9, 10]:
+                    new_block_idx = 2
+                    resnet_idx = block_idx - 8
+                elif block_idx in [12, 13, 14]:
+                    new_block_idx = 3
+                    resnet_idx = block_idx - 12
+                else:
+                    # Keep as is for other blocks
+                    new_state_dict[key] = value
+                    continue
+
+                # Convert residual block naming
+                if ".residual.0.gamma" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.norm1.gamma"
+                elif ".residual.2.bias" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv1.bias"
+                elif ".residual.2.weight" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv1.weight"
+                elif ".residual.3.gamma" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.norm2.gamma"
+                elif ".residual.6.bias" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv2.bias"
+                elif ".residual.6.weight" in key:
+                    new_key = f"decoder.up_blocks.{new_block_idx}.resnets.{resnet_idx}.conv2.weight"
+                else:
+                    new_key = key
+
+                new_state_dict[new_key] = value
+
+            # Handle shortcut connections
+            elif ".shortcut." in key:
+                if block_idx == 4:
+                    new_key = key.replace(".shortcut.", ".resnets.0.conv_shortcut.")
+                    new_key = new_key.replace("decoder.upsamples.4", "decoder.up_blocks.1")
+                else:
+                    new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+                    new_key = new_key.replace(".shortcut.", ".conv_shortcut.")
+
+                new_state_dict[new_key] = value
+
+            # Handle upsamplers
+            elif ".resample." in key or ".time_conv." in key:
+                if block_idx == 3:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.0.upsamplers.0")
+                elif block_idx == 7:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.1.upsamplers.0")
+                elif block_idx == 11:
+                    new_key = key.replace(f"decoder.upsamples.{block_idx}", "decoder.up_blocks.2.upsamplers.0")
+                else:
+                    new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+
+                new_state_dict[new_key] = value
+            else:
+                new_key = key.replace("decoder.upsamples.", "decoder.up_blocks.")
+                new_state_dict[new_key] = value
+        else:
+            # Keep other keys unchanged
+            new_state_dict[key] = value
+
+    with init_empty_weights():
+        vae = AutoencoderKLWan()
+    vae.load_state_dict(new_state_dict, strict=True, assign=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_type", type=str, default=None)
+    parser.add_argument("--output_path", type=str, required=True)
+    parser.add_argument("--dtype", default="fp32")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    dtype = DTYPE_MAPPING[args.dtype]
+
+    transformer = convert_transformer(args.model_type).to(dtype=dtype)
+    vae = convert_vae()
+    text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl")
+    tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
+    scheduler = UniPCMultistepScheduler(
+        prediction_type="flow_prediction",
+        num_train_timesteps=1000,
+        use_flow_sigmas=True,
+    )
+
+    if "I2V" in args.model_type or "FLF2V" in args.model_type:
+        image_encoder = CLIPVisionModelWithProjection.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+        image_processor = AutoProcessor.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+        pipe = SkyReelsV2ImageToVideoPipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            image_processor=image_processor,
+        )
+    elif "T2V" in args.model_type:
+        pipe = SkyReelsV2Pipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+        )
+    elif "DF" in args.model_type:
+        pipe = SkyReelsV2DiffusionForcingPipeline(
+            transformer=transformer,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+    pipe.save_pretrained(
+        args.output_path,
+        safe_serialization=True,
+        max_shard_size="5GB",
+        # push_to_hub=True,
+        # repo_id=f"<place_holder>/{args.model_type}-Diffusers",
+    )
@@ -163,6 +163,7 @@ else:
        [
            "AllegroTransformer3DModel",
            "AsymmetricAutoencoderKL",
+            "AttentionBackendName",
            "AuraFlowTransformer2DModel",
            "AutoencoderDC",
            "AutoencoderKL",
@@ -219,6 +220,7 @@ else:
            "SD3ControlNetModel",
            "SD3MultiControlNetModel",
            "SD3Transformer2DModel",
+            "SkyReelsV2Transformer3DModel",
            "SparseControlNetModel",
            "StableAudioDiTModel",
            "StableCascadeUNet",
@@ -237,6 +239,7 @@ else:
            "VQModel",
            "WanTransformer3DModel",
            "WanVACETransformer3DModel",
+            "attention_backend",
        ]
    )
    _import_structure["modular_pipelines"].extend(
@@ -363,6 +366,8 @@ else:
        [
            "StableDiffusionXLAutoBlocks",
            "StableDiffusionXLModularPipeline",
+            "WanAutoBlocks",
+            "WanModularPipeline",
        ]
    )
    _import_structure["pipelines"].extend(
@@ -488,6 +493,11 @@ else:
            "SemanticStableDiffusionPipeline",
            "ShapEImg2ImgPipeline",
            "ShapEPipeline",
+            "SkyReelsV2DiffusionForcingImageToVideoPipeline",
+            "SkyReelsV2DiffusionForcingPipeline",
+            "SkyReelsV2DiffusionForcingVideoToVideoPipeline",
+            "SkyReelsV2ImageToVideoPipeline",
+            "SkyReelsV2Pipeline",
            "StableAudioPipeline",
            "StableAudioProjectionModel",
            "StableCascadeCombinedPipeline",
@@ -809,6 +819,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
        from .models import (
            AllegroTransformer3DModel,
            AsymmetricAutoencoderKL,
+            AttentionBackendName,
            AuraFlowTransformer2DModel,
            AutoencoderDC,
            AutoencoderKL,
@@ -865,6 +876,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            SD3ControlNetModel,
            SD3MultiControlNetModel,
            SD3Transformer2DModel,
+            SkyReelsV2Transformer3DModel,
            SparseControlNetModel,
            StableAudioDiTModel,
            T2IAdapter,
@@ -882,6 +894,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            VQModel,
            WanTransformer3DModel,
            WanVACETransformer3DModel,
+            attention_backend,
        )
        from .modular_pipelines import (
            ComponentsManager,
@@ -988,6 +1001,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
        from .modular_pipelines import (
            StableDiffusionXLAutoBlocks,
            StableDiffusionXLModularPipeline,
+            WanAutoBlocks,
+            WanModularPipeline,
        )
        from .pipelines import (
            AllegroPipeline,
@@ -1109,6 +1124,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            SemanticStableDiffusionPipeline,
            ShapEImg2ImgPipeline,
            ShapEPipeline,
+            SkyReelsV2DiffusionForcingImageToVideoPipeline,
+            SkyReelsV2DiffusionForcingPipeline,
+            SkyReelsV2DiffusionForcingVideoToVideoPipeline,
+            SkyReelsV2ImageToVideoPipeline,
+            SkyReelsV2Pipeline,
            StableAudioPipeline,
            StableAudioProjectionModel,
            StableCascadeCombinedPipeline,
@@ -763,4 +763,7 @@ class LegacyConfigMixin(ConfigMixin):
        # resolve remapping
        remapped_class = _fetch_remapped_cls_from_config(config, cls)

-        return remapped_class.from_config(config, return_unused_kwargs, **kwargs)
+        if remapped_class is cls:
+            return super(LegacyConfigMixin, remapped_class).from_config(config, return_unused_kwargs, **kwargs)
+        else:
+            return remapped_class.from_config(config, return_unused_kwargs, **kwargs)
@@ -107,6 +107,7 @@ class TransformerBlockRegistry:
 def _register_attention_processors_metadata():
    from ..models.attention_processor import AttnProcessor2_0
    from ..models.transformers.transformer_cogview4 import CogView4AttnProcessor
+    from ..models.transformers.transformer_wan import WanAttnProcessor2_0

    # AttnProcessor2_0
    AttentionProcessorRegistry.register(
@@ -124,6 +125,14 @@ def _register_attention_processors_metadata():
        ),
    )

+    # WanAttnProcessor2_0
+    AttentionProcessorRegistry.register(
+        model_class=WanAttnProcessor2_0,
+        metadata=AttentionProcessorMetadata(
+            skip_processor_output_fn=_skip_proc_output_fn_Attention_WanAttnProcessor2_0,
+        ),
+    )
+

 def _register_transformer_blocks_metadata():
    from ..models.attention import BasicTransformerBlock
@@ -261,4 +270,5 @@ def _skip_attention___ret___hidden_states___encoder_hidden_states(self, *args, *

 _skip_proc_output_fn_Attention_AttnProcessor2_0 = _skip_attention___ret___hidden_states
 _skip_proc_output_fn_Attention_CogView4AttnProcessor = _skip_attention___ret___hidden_states___encoder_hidden_states
+_skip_proc_output_fn_Attention_WanAttnProcessor2_0 = _skip_attention___ret___hidden_states
 # fmt: on
@@ -18,6 +18,7 @@ from typing import Any, Callable, List, Optional, Tuple

 import torch

+from ..models.attention import AttentionModuleMixin
 from ..models.attention_processor import Attention, MochiAttention
 from ..models.modeling_outputs import Transformer2DModelOutput
 from ..utils import logging
@@ -567,7 +568,7 @@ def apply_faster_cache(module: torch.nn.Module, config: FasterCacheConfig) -> No
    _apply_faster_cache_on_denoiser(module, config)

    for name, submodule in module.named_modules():
-        if not isinstance(submodule, _ATTENTION_CLASSES):
+        if not isinstance(submodule, (*_ATTENTION_CLASSES, AttentionModuleMixin)):
            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)
@@ -91,10 +91,19 @@ class AttentionScoreSkipFunctionMode(torch.overrides.TorchFunctionMode):
        if kwargs is None:
            kwargs = {}
        if func is torch.nn.functional.scaled_dot_product_attention:
+            query = kwargs.get("query", None)
+            key = kwargs.get("key", None)
            value = kwargs.get("value", None)
-            if value is None:
-                value = args[2]
-            return value
+            query = query if query is not None else args[0]
+            key = key if key is not None else args[1]
+            value = value if value is not None else args[2]
+            # If the Q sequence length does not match KV sequence length, methods like
+            # Perturbed Attention Guidance cannot be used (because the caller expects
+            # the same sequence length as Q, but if we return V here, it will not match).
+            # When Q.shape[2] != V.shape[2], PAG will essentially not be applied and
+            # the overall effect would that be of normal CFG with a scale of (guidance_scale + perturbed_guidance_scale).
+            if query.shape[2] == value.shape[2]:
+                return value
        return func(*args, **kwargs)


@@ -18,6 +18,7 @@ from typing import Any, Callable, Optional, Tuple, Union

 import torch

+from ..models.attention import AttentionModuleMixin
 from ..models.attention_processor import Attention, MochiAttention
 from ..utils import logging
 from .hooks import HookRegistry, ModelHook
@@ -227,7 +228,7 @@ def apply_pyramid_attention_broadcast(module: torch.nn.Module, config: PyramidAt
        config.spatial_attention_block_skip_range = 2

    for name, submodule in module.named_modules():
-        if not isinstance(submodule, _ATTENTION_CLASSES):
+        if not isinstance(submodule, (*_ATTENTION_CLASSES, AttentionModuleMixin)):
            # PAB has been implemented specific to Diffusers' Attention classes. However, this does not mean that PAB
            # cannot be applied to this layer. For custom layers, users can extend this functionality and implement
            # their own PAB logic similar to `_apply_pyramid_attention_broadcast_on_attention_class`.
@@ -78,6 +78,7 @@ if is_torch_available():
            "Lumina2LoraLoaderMixin",
            "WanLoraLoaderMixin",
            "HiDreamImageLoraLoaderMixin",
+            "SkyReelsV2LoraLoaderMixin",
        ]
        _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
        _import_structure["ip_adapter"] = [
@@ -119,6 +120,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                Mochi1LoraLoaderMixin,
                SanaLoraLoaderMixin,
                SD3LoraLoaderMixin,
+                SkyReelsV2LoraLoaderMixin,
                StableDiffusionLoraLoaderMixin,
                StableDiffusionXLLoraLoaderMixin,
                WanLoraLoaderMixin,
@@ -40,8 +40,6 @@ if is_transformers_available():
 from ..models.attention_processor import (
    AttnProcessor,
    AttnProcessor2_0,
-    FluxAttnProcessor2_0,
-    FluxIPAdapterJointAttnProcessor2_0,
    IPAdapterAttnProcessor,
    IPAdapterAttnProcessor2_0,
    IPAdapterXFormersAttnProcessor,
@@ -867,6 +865,9 @@ class FluxIPAdapterMixin:
        >>> ...
        ```
        """
+        # TODO: once the 1.0.0 deprecations are in, we can move the imports to top-level
+        from ..models.transformers.transformer_flux import FluxAttnProcessor, FluxIPAdapterAttnProcessor
+
        # remove CLIP image encoder
        if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is not None:
            self.image_encoder = None
@@ -886,9 +887,9 @@ class FluxIPAdapterMixin:
        # restore original Transformer attention processors layers
        attn_procs = {}
        for name, value in self.transformer.attn_processors.items():
-            attn_processor_class = FluxAttnProcessor2_0()
+            attn_processor_class = FluxAttnProcessor()
            attn_procs[name] = (
-                attn_processor_class if isinstance(value, (FluxIPAdapterJointAttnProcessor2_0)) else value.__class__()
+                attn_processor_class if isinstance(value, FluxIPAdapterAttnProcessor) else value.__class__()
            )
        self.transformer.set_attn_processor(attn_procs)

@@ -5454,6 +5454,404 @@ class WanLoraLoaderMixin(LoraBaseMixin):
        super().unfuse_lora(components=components, **kwargs)


+class SkyReelsV2LoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`SkyReelsV2Transformer3DModel`].
+    """
+
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+
+    @classmethod
+    @validate_hf_hub_args
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.lora_state_dict
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            return_lora_metadata (`bool`, *optional*, defaults to False):
+                When enabled, additionally return the LoRA adapter metadata, typically found in the state dict.
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        return_lora_metadata = kwargs.pop("return_lora_metadata", False)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {"file_type": "attn_procs_weights", "framework": "pytorch"}
+
+        state_dict, metadata = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+        if any(k.startswith("diffusion_model.") for k in state_dict):
+            state_dict = _convert_non_diffusers_wan_lora_to_diffusers(state_dict)
+        elif any(k.startswith("lora_unet_") for k in state_dict):
+            state_dict = _convert_musubi_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:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+
+        out = (state_dict, metadata) if return_lora_metadata else state_dict
+        return out
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin._maybe_expand_t2v_lora_for_i2v
+    def _maybe_expand_t2v_lora_for_i2v(
+        cls,
+        transformer: torch.nn.Module,
+        state_dict,
+    ):
+        if transformer.config.image_dim is None:
+            return state_dict
+
+        target_device = transformer.device
+
+        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 if "blocks." in k})
+            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)
+            has_bias = any(".lora_B.bias" 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"]):
+                    # These keys should exist if the block `i` was part of the T2V LoRA.
+                    ref_key_lora_A = f"transformer.blocks.{i}.attn2.to_k.lora_A.weight"
+                    ref_key_lora_B = f"transformer.blocks.{i}.attn2.to_k.lora_B.weight"
+
+                    if ref_key_lora_A not in state_dict or ref_key_lora_B not in state_dict:
+                        continue
+
+                    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"], device=target_device
+                    )
+                    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"], device=target_device
+                    )
+
+                    # If the original LoRA had biases (indicated by has_bias)
+                    # AND the specific reference bias key exists for this block.
+
+                    ref_key_lora_B_bias = f"transformer.blocks.{i}.attn2.to_k.lora_B.bias"
+                    if has_bias and ref_key_lora_B_bias in state_dict:
+                        ref_lora_B_bias_tensor = state_dict[ref_key_lora_B_bias]
+                        state_dict[f"transformer.blocks.{i}.attn2.{c}.lora_B.bias"] = torch.zeros_like(
+                            ref_lora_B_bias_tensor,
+                            device=target_device,
+                        )
+
+        return state_dict
+
+    # Copied from diffusers.loaders.lora_pipeline.WanLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        adapter_name: Optional[str] = None,
+        hotswap: bool = False,
+        **kwargs,
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
+        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
+        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
+        dict is loaded into `self.transformer`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            hotswap (`bool`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
+            kwargs (`dict`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        kwargs["return_lora_metadata"] = True
+        state_dict, metadata = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+        # 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.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->SkyReelsV2Transformer3DModel
+    def load_lora_into_transformer(
+        cls,
+        state_dict,
+        transformer,
+        adapter_name=None,
+        _pipeline=None,
+        low_cpu_mem_usage=False,
+        hotswap: bool = False,
+        metadata=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`SkyReelsV2Transformer3DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            hotswap (`bool`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`].
+            metadata (`dict`):
+                Optional LoRA adapter metadata. When supplied, the `LoraConfig` arguments of `peft` won't be derived
+                from the state dict.
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            metadata=metadata,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        transformer_lora_adapter_metadata: Optional[dict] = None,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the transformer.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+            transformer_lora_adapter_metadata:
+                LoRA adapter metadata associated with the transformer to be serialized with the state dict.
+        """
+        state_dict = {}
+        lora_adapter_metadata = {}
+
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+
+        state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        if transformer_lora_adapter_metadata is not None:
+            lora_adapter_metadata.update(
+                _pack_dict_with_prefix(transformer_lora_adapter_metadata, cls.transformer_name)
+            )
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+            lora_adapter_metadata=lora_adapter_metadata,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.fuse_lora
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+        Example:
+
+        ```py
+        from diffusers import DiffusionPipeline
+        import torch
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        super().fuse_lora(
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.unfuse_lora
+    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            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, **kwargs)
+
+
 class CogView4LoraLoaderMixin(LoraBaseMixin):
    r"""
    Load LoRA layers into [`WanTransformer3DModel`]. Specific to [`CogView4Pipeline`].
@@ -24,7 +24,7 @@ from typing_extensions import Self
 from .. import __version__
 from ..quantizers import DiffusersAutoQuantizer
 from ..utils import deprecate, is_accelerate_available, logging
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache
 from .single_file_utils import (
    SingleFileComponentError,
    convert_animatediff_checkpoint_to_diffusers,
@@ -431,10 +431,7 @@ class FromOriginalModelMixin:
                keep_in_fp32_modules=keep_in_fp32_modules,
                unexpected_keys=unexpected_keys,
            )
-            # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-            # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
            empty_device_cache()
-            device_synchronize()
        else:
            _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)

@@ -46,7 +46,7 @@ from ..utils import (
 )
 from ..utils.constants import DIFFUSERS_REQUEST_TIMEOUT
 from ..utils.hub_utils import _get_model_file
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache


 if is_transformers_available():
@@ -1690,10 +1690,7 @@ def create_diffusers_clip_model_from_ldm(

    if is_accelerate_available():
        load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
        empty_device_cache()
-        device_synchronize()
    else:
        model.load_state_dict(diffusers_format_checkpoint, strict=False)

@@ -2153,10 +2150,7 @@ def create_diffusers_t5_model_from_checkpoint(

    if is_accelerate_available():
        load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
        empty_device_cache()
-        device_synchronize()
    else:
        model.load_state_dict(diffusers_format_checkpoint)

@@ -19,7 +19,7 @@ from ..models.embeddings import (
 )
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import is_accelerate_available, is_torch_version, logging
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache


 if is_accelerate_available():
@@ -82,14 +82,11 @@ class FluxTransformer2DLoadersMixin:
            device_map = {"": self.device}
            load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
            empty_device_cache()
-            device_synchronize()

        return image_projection

    def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):
-        from ..models.attention_processor import (
-            FluxIPAdapterJointAttnProcessor2_0,
-        )
+        from ..models.transformers.transformer_flux import FluxIPAdapterAttnProcessor

        if low_cpu_mem_usage:
            if is_accelerate_available():
@@ -121,7 +118,7 @@ class FluxTransformer2DLoadersMixin:
            else:
                cross_attention_dim = self.config.joint_attention_dim
                hidden_size = self.inner_dim
-                attn_processor_class = FluxIPAdapterJointAttnProcessor2_0
+                attn_processor_class = FluxIPAdapterAttnProcessor
                num_image_text_embeds = []
                for state_dict in state_dicts:
                    if "proj.weight" in state_dict["image_proj"]:
@@ -158,7 +155,6 @@ class FluxTransformer2DLoadersMixin:
                key_id += 1

        empty_device_cache()
-        device_synchronize()

        return attn_procs

@@ -18,7 +18,7 @@ from ..models.attention_processor import SD3IPAdapterJointAttnProcessor2_0
 from ..models.embeddings import IPAdapterTimeImageProjection
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import is_accelerate_available, is_torch_version, logging
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache


 logger = logging.get_logger(__name__)
@@ -82,7 +82,6 @@ class SD3Transformer2DLoadersMixin:
                )

        empty_device_cache()
-        device_synchronize()

        return attn_procs

@@ -152,7 +151,6 @@ class SD3Transformer2DLoadersMixin:
            device_map = {"": self.device}
            load_model_dict_into_meta(image_proj, updated_state_dict, device_map=device_map, dtype=self.dtype)
            empty_device_cache()
-            device_synchronize()

        return image_proj

@@ -43,7 +43,7 @@ from ..utils import (
    is_torch_version,
    logging,
 )
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache
 from .lora_base import _func_optionally_disable_offloading
 from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE, TEXT_ENCODER_NAME, UNET_NAME
 from .utils import AttnProcsLayers
@@ -755,7 +755,6 @@ class UNet2DConditionLoadersMixin:
            device_map = {"": self.device}
            load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
            empty_device_cache()
-            device_synchronize()

        return image_projection

@@ -854,7 +853,6 @@ class UNet2DConditionLoadersMixin:
                key_id += 2

        empty_device_cache()
-        device_synchronize()

        return attn_procs

@@ -26,6 +26,7 @@ _import_structure = {}

 if is_torch_available():
    _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
+    _import_structure["attention_dispatch"] = ["AttentionBackendName", "attention_backend"]
    _import_structure["auto_model"] = ["AutoModel"]
    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
    _import_structure["autoencoders.autoencoder_dc"] = ["AutoencoderDC"]
@@ -88,6 +89,7 @@ if is_torch_available():
    _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
    _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
    _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
+    _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
    _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
    _import_structure["transformers.transformer_wan"] = ["WanTransformer3DModel"]
    _import_structure["transformers.transformer_wan_vace"] = ["WanVACETransformer3DModel"]
@@ -111,6 +113,7 @@ if is_flax_available():
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
    if is_torch_available():
        from .adapter import MultiAdapter, T2IAdapter
+        from .attention_dispatch import AttentionBackendName, attention_backend
        from .auto_model import AutoModel
        from .autoencoders import (
            AsymmetricAutoencoderKL,
@@ -176,6 +179,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            PriorTransformer,
            SanaTransformer2DModel,
            SD3Transformer2DModel,
+            SkyReelsV2Transformer3DModel,
            StableAudioDiTModel,
            T5FilmDecoder,
            Transformer2DModel,
@@ -11,23 +11,504 @@
 # 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.
-from typing import Any, Dict, List, Optional, Tuple
+
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union

 import torch
+import torch.nn as nn
 import torch.nn.functional as F
-from torch import nn

 from ..utils import deprecate, logging
+from ..utils.import_utils import is_torch_npu_available, is_torch_xla_available, is_xformers_available
 from ..utils.torch_utils import maybe_allow_in_graph
 from .activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, LinearActivation, SwiGLU
-from .attention_processor import Attention, JointAttnProcessor2_0
+from .attention_processor import Attention, AttentionProcessor, JointAttnProcessor2_0
 from .embeddings import SinusoidalPositionalEmbedding
 from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm, SD35AdaLayerNormZeroX


+if is_xformers_available():
+    import xformers as xops
+else:
+    xops = None
+
+
 logger = logging.get_logger(__name__)


+class AttentionMixin:
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+        """
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        for module in self.modules():
+            if isinstance(module, AttentionModuleMixin):
+                module.fuse_projections()
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        for module in self.modules():
+            if isinstance(module, AttentionModuleMixin):
+                module.unfuse_projections()
+
+
+class AttentionModuleMixin:
+    _default_processor_cls = None
+    _available_processors = []
+    fused_projections = False
+
+    def set_processor(self, processor: AttentionProcessor) -> None:
+        """
+        Set the attention processor to use.
+
+        Args:
+            processor (`AttnProcessor`):
+                The attention processor to use.
+        """
+        # if current processor is in `self._modules` and if passed `processor` is not, we need to
+        # pop `processor` from `self._modules`
+        if (
+            hasattr(self, "processor")
+            and isinstance(self.processor, torch.nn.Module)
+            and not isinstance(processor, torch.nn.Module)
+        ):
+            logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
+            self._modules.pop("processor")
+
+        self.processor = processor
+
+    def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
+        """
+        Get the attention processor in use.
+
+        Args:
+            return_deprecated_lora (`bool`, *optional*, defaults to `False`):
+                Set to `True` to return the deprecated LoRA attention processor.
+
+        Returns:
+            "AttentionProcessor": The attention processor in use.
+        """
+        if not return_deprecated_lora:
+            return self.processor
+
+    def set_attention_backend(self, backend: str):
+        from .attention_dispatch import AttentionBackendName
+
+        available_backends = {x.value for x in AttentionBackendName.__members__.values()}
+        if backend not in available_backends:
+            raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
+
+        backend = AttentionBackendName(backend.lower())
+        self.processor._attention_backend = backend
+
+    def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
+        """
+        Set whether to use NPU flash attention from `torch_npu` or not.
+
+        Args:
+            use_npu_flash_attention (`bool`): Whether to use NPU flash attention or not.
+        """
+
+        if use_npu_flash_attention:
+            if not is_torch_npu_available():
+                raise ImportError("torch_npu is not available")
+
+        self.set_attention_backend("_native_npu")
+
+    def set_use_xla_flash_attention(
+        self,
+        use_xla_flash_attention: bool,
+        partition_spec: Optional[Tuple[Optional[str], ...]] = None,
+        is_flux=False,
+    ) -> None:
+        """
+        Set whether to use XLA flash attention from `torch_xla` or not.
+
+        Args:
+            use_xla_flash_attention (`bool`):
+                Whether to use pallas flash attention kernel from `torch_xla` or not.
+            partition_spec (`Tuple[]`, *optional*):
+                Specify the partition specification if using SPMD. Otherwise None.
+            is_flux (`bool`, *optional*, defaults to `False`):
+                Whether the model is a Flux model.
+        """
+        if use_xla_flash_attention:
+            if not is_torch_xla_available():
+                raise ImportError("torch_xla is not available")
+
+        self.set_attention_backend("_native_xla")
+
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
+    ) -> None:
+        """
+        Set whether to use memory efficient attention from `xformers` or not.
+
+        Args:
+            use_memory_efficient_attention_xformers (`bool`):
+                Whether to use memory efficient attention from `xformers` or not.
+            attention_op (`Callable`, *optional*):
+                The attention operation to use. Defaults to `None` which uses the default attention operation from
+                `xformers`.
+        """
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    "Refer to https://github.com/facebookresearch/xformers for more information on how to install xformers",
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    if is_xformers_available():
+                        dtype = None
+                        if attention_op is not None:
+                            op_fw, op_bw = attention_op
+                            dtype, *_ = op_fw.SUPPORTED_DTYPES
+                        q = torch.randn((1, 2, 40), device="cuda", dtype=dtype)
+                        _ = xops.memory_efficient_attention(q, q, q)
+                except Exception as e:
+                    raise e
+
+                self.set_attention_backend("xformers")
+
+    @torch.no_grad()
+    def fuse_projections(self):
+        """
+        Fuse the query, key, and value projections into a single projection for efficiency.
+        """
+        # Skip if already fused
+        if getattr(self, "fused_projections", False):
+            return
+
+        device = self.to_q.weight.data.device
+        dtype = self.to_q.weight.data.dtype
+
+        if hasattr(self, "is_cross_attention") and self.is_cross_attention:
+            # Fuse cross-attention key-value projections
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_kv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+            self.to_kv.weight.copy_(concatenated_weights)
+            if hasattr(self, "use_bias") and self.use_bias:
+                concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+                self.to_kv.bias.copy_(concatenated_bias)
+        else:
+            # Fuse self-attention projections
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_qkv = nn.Linear(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+            self.to_qkv.weight.copy_(concatenated_weights)
+            if hasattr(self, "use_bias") and self.use_bias:
+                concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+                self.to_qkv.bias.copy_(concatenated_bias)
+
+        # Handle added projections for models like SD3, Flux, etc.
+        if (
+            getattr(self, "add_q_proj", None) is not None
+            and getattr(self, "add_k_proj", None) is not None
+            and getattr(self, "add_v_proj", None) is not None
+        ):
+            concatenated_weights = torch.cat(
+                [self.add_q_proj.weight.data, self.add_k_proj.weight.data, self.add_v_proj.weight.data]
+            )
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_added_qkv = nn.Linear(
+                in_features, out_features, bias=self.added_proj_bias, device=device, dtype=dtype
+            )
+            self.to_added_qkv.weight.copy_(concatenated_weights)
+            if self.added_proj_bias:
+                concatenated_bias = torch.cat(
+                    [self.add_q_proj.bias.data, self.add_k_proj.bias.data, self.add_v_proj.bias.data]
+                )
+                self.to_added_qkv.bias.copy_(concatenated_bias)
+
+        self.fused_projections = True
+
+    @torch.no_grad()
+    def unfuse_projections(self):
+        """
+        Unfuse the query, key, and value projections back to separate projections.
+        """
+        # Skip if not fused
+        if not getattr(self, "fused_projections", False):
+            return
+
+        # Remove fused projection layers
+        if hasattr(self, "to_qkv"):
+            delattr(self, "to_qkv")
+
+        if hasattr(self, "to_kv"):
+            delattr(self, "to_kv")
+
+        if hasattr(self, "to_added_qkv"):
+            delattr(self, "to_added_qkv")
+
+        self.fused_projections = False
+
+    def set_attention_slice(self, slice_size: int) -> None:
+        """
+        Set the slice size for attention computation.
+
+        Args:
+            slice_size (`int`):
+                The slice size for attention computation.
+        """
+        if hasattr(self, "sliceable_head_dim") and slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        processor = None
+
+        # Try to get a compatible processor for sliced attention
+        if slice_size is not None:
+            processor = self._get_compatible_processor("sliced")
+
+        # If no processor was found or slice_size is None, use default processor
+        if processor is None:
+            processor = self.default_processor_cls()
+
+        self.set_processor(processor)
+
+    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        batch_size, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
+        """
+        Reshape the tensor for multi-head attention processing.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        if tensor.ndim == 3:
+            batch_size, seq_len, dim = tensor.shape
+            extra_dim = 1
+        else:
+            batch_size, extra_dim, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size, seq_len * extra_dim, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3)
+
+        if out_dim == 3:
+            tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)
+
+        return tensor
+
+    def get_attention_scores(
+        self, query: torch.Tensor, key: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        """
+        Compute the attention scores.
+
+        Args:
+            query (`torch.Tensor`): The query tensor.
+            key (`torch.Tensor`): The key tensor.
+            attention_mask (`torch.Tensor`, *optional*): The attention mask to use.
+
+        Returns:
+            `torch.Tensor`: The attention probabilities/scores.
+        """
+        dtype = query.dtype
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        if attention_mask is None:
+            baddbmm_input = torch.empty(
+                query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device
+            )
+            beta = 0
+        else:
+            baddbmm_input = attention_mask
+            beta = 1
+
+        attention_scores = torch.baddbmm(
+            baddbmm_input,
+            query,
+            key.transpose(-1, -2),
+            beta=beta,
+            alpha=self.scale,
+        )
+        del baddbmm_input
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+        del attention_scores
+
+        attention_probs = attention_probs.to(dtype)
+
+        return attention_probs
+
+    def prepare_attention_mask(
+        self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3
+    ) -> torch.Tensor:
+        """
+        Prepare the attention mask for the attention computation.
+
+        Args:
+            attention_mask (`torch.Tensor`): The attention mask to prepare.
+            target_length (`int`): The target length of the attention mask.
+            batch_size (`int`): The batch size for repeating the attention mask.
+            out_dim (`int`, *optional*, defaults to `3`): Output dimension.
+
+        Returns:
+            `torch.Tensor`: The prepared attention mask.
+        """
+        head_size = self.heads
+        if attention_mask is None:
+            return attention_mask
+
+        current_length: int = attention_mask.shape[-1]
+        if current_length != target_length:
+            if attention_mask.device.type == "mps":
+                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
+                # Instead, we can manually construct the padding tensor.
+                padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)
+                padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)
+                attention_mask = torch.cat([attention_mask, padding], dim=2)
+            else:
+                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
+                #       we want to instead pad by (0, remaining_length), where remaining_length is:
+                #       remaining_length: int = target_length - current_length
+                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+
+        if out_dim == 3:
+            if attention_mask.shape[0] < batch_size * head_size:
+                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        elif out_dim == 4:
+            attention_mask = attention_mask.unsqueeze(1)
+            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
+        return attention_mask
+
+    def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Normalize the encoder hidden states.
+
+        Args:
+            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.
+
+        Returns:
+            `torch.Tensor`: The normalized encoder hidden states.
+        """
+        assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"
+        if isinstance(self.norm_cross, nn.LayerNorm):
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+        elif isinstance(self.norm_cross, nn.GroupNorm):
+            # Group norm norms along the channels dimension and expects
+            # input to be in the shape of (N, C, *). In this case, we want
+            # to norm along the hidden dimension, so we need to move
+            # (batch_size, sequence_length, hidden_size) ->
+            # (batch_size, hidden_size, sequence_length)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+        else:
+            assert False
+
+        return encoder_hidden_states
+
+
 def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
    # "feed_forward_chunk_size" can be used to save memory
    if hidden_states.shape[chunk_dim] % chunk_size != 0:
@@ -2272,558 +2272,6 @@ class FusedAuraFlowAttnProcessor2_0:
            return hidden_states


-class FluxAttnProcessor2_0:
-    """Attention processor used typically in processing the SD3-like self-attention projections."""
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError("FluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        query = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        if encoder_hidden_states is not None:
-            # `context` projections.
-            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
-            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
-            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, 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)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
-class FluxAttnProcessor2_0_NPU:
-    """Attention processor used typically in processing the SD3-like self-attention projections."""
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                "FluxAttnProcessor2_0_NPU requires PyTorch 2.0 and torch NPU, to use it, please upgrade PyTorch to 2.0 and install torch NPU"
-            )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        query = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        if encoder_hidden_states is not None:
-            # `context` projections.
-            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
-            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
-            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        if query.dtype in (torch.float16, torch.bfloat16):
-            hidden_states = torch_npu.npu_fusion_attention(
-                query,
-                key,
-                value,
-                attn.heads,
-                input_layout="BNSD",
-                pse=None,
-                scale=1.0 / math.sqrt(query.shape[-1]),
-                pre_tockens=65536,
-                next_tockens=65536,
-                keep_prob=1.0,
-                sync=False,
-                inner_precise=0,
-            )[0]
-        else:
-            hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
-class FusedFluxAttnProcessor2_0:
-    """Attention processor used typically in processing the SD3-like self-attention projections."""
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                "FusedFluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
-            )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        qkv = attn.to_qkv(hidden_states)
-        split_size = qkv.shape[-1] // 3
-        query, key, value = torch.split(qkv, split_size, dim=-1)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        # `context` projections.
-        if encoder_hidden_states is not None:
-            encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
-            split_size = encoder_qkv.shape[-1] // 3
-            (
-                encoder_hidden_states_query_proj,
-                encoder_hidden_states_key_proj,
-                encoder_hidden_states_value_proj,
-            ) = torch.split(encoder_qkv, split_size, dim=-1)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, 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)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
-class FusedFluxAttnProcessor2_0_NPU:
-    """Attention processor used typically in processing the SD3-like self-attention projections."""
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                "FluxAttnProcessor2_0_NPU requires PyTorch 2.0 and torch NPU, to use it, please upgrade PyTorch to 2.0, and install torch NPU"
-            )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        qkv = attn.to_qkv(hidden_states)
-        split_size = qkv.shape[-1] // 3
-        query, key, value = torch.split(qkv, split_size, dim=-1)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        # `context` projections.
-        if encoder_hidden_states is not None:
-            encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
-            split_size = encoder_qkv.shape[-1] // 3
-            (
-                encoder_hidden_states_query_proj,
-                encoder_hidden_states_key_proj,
-                encoder_hidden_states_value_proj,
-            ) = torch.split(encoder_qkv, split_size, dim=-1)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        if query.dtype in (torch.float16, torch.bfloat16):
-            hidden_states = torch_npu.npu_fusion_attention(
-                query,
-                key,
-                value,
-                attn.heads,
-                input_layout="BNSD",
-                pse=None,
-                scale=1.0 / math.sqrt(query.shape[-1]),
-                pre_tockens=65536,
-                next_tockens=65536,
-                keep_prob=1.0,
-                sync=False,
-                inner_precise=0,
-            )[0]
-        else:
-            hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
-class FluxIPAdapterJointAttnProcessor2_0(torch.nn.Module):
-    """Flux Attention processor for IP-Adapter."""
-
-    def __init__(
-        self, hidden_size: int, cross_attention_dim: int, num_tokens=(4,), scale=1.0, device=None, dtype=None
-    ):
-        super().__init__()
-
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
-            )
-
-        self.hidden_size = hidden_size
-        self.cross_attention_dim = cross_attention_dim
-
-        if not isinstance(num_tokens, (tuple, list)):
-            num_tokens = [num_tokens]
-
-        if not isinstance(scale, list):
-            scale = [scale] * len(num_tokens)
-        if len(scale) != len(num_tokens):
-            raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
-        self.scale = scale
-
-        self.to_k_ip = nn.ModuleList(
-            [
-                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
-                for _ in range(len(num_tokens))
-            ]
-        )
-        self.to_v_ip = nn.ModuleList(
-            [
-                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
-                for _ in range(len(num_tokens))
-            ]
-        )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-        ip_hidden_states: Optional[List[torch.Tensor]] = None,
-        ip_adapter_masks: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        hidden_states_query_proj = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        hidden_states_query_proj = hidden_states_query_proj.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            hidden_states_query_proj = attn.norm_q(hidden_states_query_proj)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        if encoder_hidden_states is not None:
-            # `context` projections.
-            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
-            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
-            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, hidden_states_query_proj], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, 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)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            # IP-adapter
-            ip_query = hidden_states_query_proj
-            ip_attn_output = torch.zeros_like(hidden_states)
-
-            for current_ip_hidden_states, scale, to_k_ip, to_v_ip in zip(
-                ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip
-            ):
-                ip_key = to_k_ip(current_ip_hidden_states)
-                ip_value = to_v_ip(current_ip_hidden_states)
-
-                ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-                ip_value = ip_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
-                current_ip_hidden_states = F.scaled_dot_product_attention(
-                    ip_query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
-                )
-                current_ip_hidden_states = current_ip_hidden_states.transpose(1, 2).reshape(
-                    batch_size, -1, attn.heads * head_dim
-                )
-                current_ip_hidden_states = current_ip_hidden_states.to(ip_query.dtype)
-                ip_attn_output += scale * current_ip_hidden_states
-
-            return hidden_states, encoder_hidden_states, ip_attn_output
-        else:
-            return hidden_states
-
-
 class CogVideoXAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
@@ -3453,106 +2901,6 @@ class XLAFlashAttnProcessor2_0:
        return hidden_states


-class XLAFluxFlashAttnProcessor2_0:
-    r"""
-    Processor for implementing scaled dot-product attention with pallas flash attention kernel if using `torch_xla`.
-    """
-
-    def __init__(self, partition_spec: Optional[Tuple[Optional[str], ...]] = None):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(
-                "XLAFlashAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
-            )
-        if is_torch_xla_version("<", "2.3"):
-            raise ImportError("XLA flash attention requires torch_xla version >= 2.3.")
-        if is_spmd() and is_torch_xla_version("<", "2.4"):
-            raise ImportError("SPMD support for XLA flash attention needs torch_xla version >= 2.4.")
-        self.partition_spec = partition_spec
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.FloatTensor,
-        encoder_hidden_states: torch.FloatTensor = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.FloatTensor:
-        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-
-        # `sample` projections.
-        query = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
-        if encoder_hidden_states is not None:
-            # `context` projections.
-            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
-            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
-            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
-
-            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
-                batch_size, -1, attn.heads, head_dim
-            ).transpose(1, 2)
-
-            if attn.norm_added_q is not None:
-                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
-            if attn.norm_added_k is not None:
-                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
-
-            # attention
-            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
-            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
-            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
-
-        if image_rotary_emb is not None:
-            from .embeddings import apply_rotary_emb
-
-            query = apply_rotary_emb(query, image_rotary_emb)
-            key = apply_rotary_emb(key, image_rotary_emb)
-
-        query /= math.sqrt(head_dim)
-        hidden_states = flash_attention(query, key, value, causal=False)
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-        hidden_states = hidden_states.to(query.dtype)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = (
-                hidden_states[:, : encoder_hidden_states.shape[1]],
-                hidden_states[:, encoder_hidden_states.shape[1] :],
-            )
-
-            # linear proj
-            hidden_states = attn.to_out[0](hidden_states)
-            # dropout
-            hidden_states = attn.to_out[1](hidden_states)
-
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
 class MochiVaeAttnProcessor2_0:
    r"""
    Attention processor used in Mochi VAE.
@@ -5992,17 +5340,6 @@ class LoRAAttnAddedKVProcessor:
        pass


-class FluxSingleAttnProcessor2_0(FluxAttnProcessor2_0):
-    r"""
-    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
-    """
-
-    def __init__(self):
-        deprecation_message = "`FluxSingleAttnProcessor2_0` is deprecated and will be removed in a future version. Please use `FluxAttnProcessor2_0` instead."
-        deprecate("FluxSingleAttnProcessor2_0", "0.32.0", deprecation_message)
-        super().__init__()
-
-
 class SanaLinearAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product linear attention.
@@ -6167,6 +5504,111 @@ class PAGIdentitySanaLinearAttnProcessor2_0:
        return hidden_states


+class FluxAttnProcessor2_0:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = "`FluxAttnProcessor2_0` is deprecated and this will be removed in a future version. Please use `FluxAttnProcessor`"
+        deprecate("FluxAttnProcessor2_0", "1.0.0", deprecation_message)
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        return FluxAttnProcessor(*args, **kwargs)
+
+
+class FluxSingleAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = "`FluxSingleAttnProcessor` is deprecated and will be removed in a future version. Please use `FluxAttnProcessorSDPA` instead."
+        deprecate("FluxSingleAttnProcessor2_0", "1.0.0", deprecation_message)
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        return FluxAttnProcessor(*args, **kwargs)
+
+
+class FusedFluxAttnProcessor2_0:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = "`FusedFluxAttnProcessor2_0` is deprecated and this will be removed in a future version. Please use `FluxAttnProcessor`"
+        deprecate("FusedFluxAttnProcessor2_0", "1.0.0", deprecation_message)
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        return FluxAttnProcessor(*args, **kwargs)
+
+
+class FluxIPAdapterJointAttnProcessor2_0:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = "`FluxIPAdapterJointAttnProcessor2_0` is deprecated and this will be removed in a future version. Please use `FluxIPAdapterAttnProcessor`"
+        deprecate("FluxIPAdapterJointAttnProcessor2_0", "1.0.0", deprecation_message)
+
+        from .transformers.transformer_flux import FluxIPAdapterAttnProcessor
+
+        return FluxIPAdapterAttnProcessor(*args, **kwargs)
+
+
+class FluxAttnProcessor2_0_NPU:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = (
+            "FluxAttnProcessor2_0_NPU is deprecated and will be removed in a future version. An "
+            "alternative solution to use NPU Flash Attention will be provided in the future."
+        )
+        deprecate("FluxAttnProcessor2_0_NPU", "1.0.0", deprecation_message, standard_warn=False)
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        processor = FluxAttnProcessor()
+        processor._attention_backend = "_native_npu"
+        return processor
+
+
+class FusedFluxAttnProcessor2_0_NPU:
+    def __new__(self):
+        deprecation_message = (
+            "FusedFluxAttnProcessor2_0_NPU is deprecated and will be removed in a future version. An "
+            "alternative solution to use NPU Flash Attention will be provided in the future."
+        )
+        deprecate("FusedFluxAttnProcessor2_0_NPU", "1.0.0", deprecation_message, standard_warn=False)
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        processor = FluxAttnProcessor()
+        processor._attention_backend = "_fused_npu"
+        return processor
+
+
+class XLAFluxFlashAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention with pallas flash attention kernel if using `torch_xla`.
+    """
+
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = (
+            "XLAFluxFlashAttnProcessor2_0 is deprecated and will be removed in diffusers 1.0.0. An "
+            "alternative solution to using XLA Flash Attention will be provided in the future."
+        )
+        deprecate("XLAFluxFlashAttnProcessor2_0", "1.0.0", deprecation_message, standard_warn=False)
+
+        if is_torch_xla_version("<", "2.3"):
+            raise ImportError("XLA flash attention requires torch_xla version >= 2.3.")
+        if is_spmd() and is_torch_xla_version("<", "2.4"):
+            raise ImportError("SPMD support for XLA flash attention needs torch_xla version >= 2.4.")
+
+        from .transformers.transformer_flux import FluxAttnProcessor
+
+        if len(args) > 0 or kwargs.get("partition_spec", None) is not None:
+            deprecation_message = (
+                "partition_spec was not used in the processor implementation when it was added. Passing it "
+                "is a no-op and support for it will be removed."
+            )
+            deprecate("partition_spec", "1.0.0", deprecation_message)
+
+        processor = FluxAttnProcessor(*args, **kwargs)
+        processor._attention_backend = "_native_xla"
+        return processor
+
+
 ADDED_KV_ATTENTION_PROCESSORS = (
    AttnAddedKVProcessor,
    SlicedAttnAddedKVProcessor,
@@ -319,7 +319,7 @@ def get_2d_sincos_pos_embed_from_grid(embed_dim, grid, output_type="np"):
    return emb


-def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np", flip_sin_to_cos=False):
    """
    This function generates 1D positional embeddings from a grid.

@@ -352,6 +352,11 @@ def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
    emb_cos = torch.cos(out)  # (M, D/2)

    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+
+    # flip sine and cosine embeddings
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, embed_dim // 2 :], emb[:, : embed_dim // 2]], dim=1)
+
    return emb


@@ -1176,6 +1181,7 @@ def apply_rotary_emb(
    freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
    use_real: bool = True,
    use_real_unbind_dim: int = -1,
+    sequence_dim: int = 2,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
@@ -1193,8 +1199,15 @@ def apply_rotary_emb(
    """
    if use_real:
        cos, sin = freqs_cis  # [S, D]
-        cos = cos[None, None]
-        sin = sin[None, None]
+        if sequence_dim == 2:
+            cos = cos[None, None, :, :]
+            sin = sin[None, None, :, :]
+        elif sequence_dim == 1:
+            cos = cos[None, :, None, :]
+            sin = sin[None, :, None, :]
+        else:
+            raise ValueError(f"`sequence_dim={sequence_dim}` but should be 1 or 2.")
+
        cos, sin = cos.to(x.device), sin.to(x.device)

        if use_real_unbind_dim == -1:
@@ -1238,37 +1251,6 @@ def apply_rotary_emb_allegro(x: torch.Tensor, freqs_cis, positions):
    return x


-class FluxPosEmbed(nn.Module):
-    # modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11
-    def __init__(self, theta: int, axes_dim: List[int]):
-        super().__init__()
-        self.theta = theta
-        self.axes_dim = axes_dim
-
-    def forward(self, ids: torch.Tensor) -> torch.Tensor:
-        n_axes = ids.shape[-1]
-        cos_out = []
-        sin_out = []
-        pos = ids.float()
-        is_mps = ids.device.type == "mps"
-        is_npu = ids.device.type == "npu"
-        freqs_dtype = torch.float32 if (is_mps or is_npu) else torch.float64
-        for i in range(n_axes):
-            cos, sin = get_1d_rotary_pos_embed(
-                self.axes_dim[i],
-                pos[:, i],
-                theta=self.theta,
-                repeat_interleave_real=True,
-                use_real=True,
-                freqs_dtype=freqs_dtype,
-            )
-            cos_out.append(cos)
-            sin_out.append(sin)
-        freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
-        freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
-        return freqs_cos, freqs_sin
-
-
 class TimestepEmbedding(nn.Module):
    def __init__(
        self,
@@ -2619,3 +2601,13 @@ class MultiIPAdapterImageProjection(nn.Module):
            projected_image_embeds.append(image_embed)

        return projected_image_embeds
+
+
+class FluxPosEmbed(nn.Module):
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = "Importing and using `FluxPosEmbed` from `diffusers.models.embeddings` is deprecated. Please import it from `diffusers.models.transformers.transformer_flux`."
+        deprecate("FluxPosEmbed", "1.0.0", deprecation_message)
+
+        from .transformers.transformer_flux import FluxPosEmbed
+
+        return FluxPosEmbed(*args, **kwargs)
@@ -62,7 +62,7 @@ from ..utils.hub_utils import (
    load_or_create_model_card,
    populate_model_card,
 )
-from ..utils.torch_utils import device_synchronize, empty_device_cache
+from ..utils.torch_utils import empty_device_cache
 from .model_loading_utils import (
    _caching_allocator_warmup,
    _determine_device_map,
@@ -172,7 +172,11 @@ def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:

    for name, param in parameter.named_parameters():
        last_dtype = param.dtype
-        if parameter._keep_in_fp32_modules and any(m in name for m in parameter._keep_in_fp32_modules):
+        if (
+            hasattr(parameter, "_keep_in_fp32_modules")
+            and parameter._keep_in_fp32_modules
+            and any(m in name for m in parameter._keep_in_fp32_modules)
+        ):
            continue

        if param.is_floating_point():
@@ -606,6 +610,60 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
            offload_to_disk_path=offload_to_disk_path,
        )

+    def set_attention_backend(self, backend: str) -> None:
+        """
+        Set the attention backend for the model.
+
+        Args:
+            backend (`str`):
+                The name of the backend to set. Must be one of the available backends defined in
+                `AttentionBackendName`. Available backends can be found in
+                `diffusers.attention_dispatch.AttentionBackendName`. Defaults to torch native scaled dot product
+                attention as backend.
+        """
+        from .attention import AttentionModuleMixin
+        from .attention_dispatch import AttentionBackendName, _check_attention_backend_requirements
+
+        # TODO: the following will not be required when everything is refactored to AttentionModuleMixin
+        from .attention_processor import Attention, MochiAttention
+
+        logger.warning("Attention backends are an experimental feature and the API may be subject to change.")
+
+        backend = backend.lower()
+        available_backends = {x.value for x in AttentionBackendName.__members__.values()}
+        if backend not in available_backends:
+            raise ValueError(f"`{backend=}` must be one of the following: " + ", ".join(available_backends))
+        backend = AttentionBackendName(backend)
+        _check_attention_backend_requirements(backend)
+
+        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
+        for module in self.modules():
+            if not isinstance(module, attention_classes):
+                continue
+            processor = module.processor
+            if processor is None or not hasattr(processor, "_attention_backend"):
+                continue
+            processor._attention_backend = backend
+
+    def reset_attention_backend(self) -> None:
+        """
+        Resets the attention backend for the model. Following calls to `forward` will use the environment default or
+        the torch native scaled dot product attention.
+        """
+        from .attention import AttentionModuleMixin
+        from .attention_processor import Attention, MochiAttention
+
+        logger.warning("Attention backends are an experimental feature and the API may be subject to change.")
+
+        attention_classes = (Attention, MochiAttention, AttentionModuleMixin)
+        for module in self.modules():
+            if not isinstance(module, attention_classes):
+                continue
+            processor = module.processor
+            if processor is None or not hasattr(processor, "_attention_backend"):
+                continue
+            processor._attention_backend = None
+
    def save_pretrained(
        self,
        save_directory: Union[str, os.PathLike],
@@ -1540,10 +1598,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                    assign_to_params_buffers = check_support_param_buffer_assignment(model, state_dict)
                error_msgs += _load_state_dict_into_model(model, state_dict, assign_to_params_buffers)

-        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
-        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
        empty_device_cache()
-        device_synchronize()

        if offload_index is not None and len(offload_index) > 0:
            save_offload_index(offload_index, offload_folder)
@@ -1880,4 +1935,9 @@ class LegacyModelMixin(ModelMixin):
        # resolve remapping
        remapped_class = _fetch_remapped_cls_from_config(config, cls)

-        return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs_copy)
+        if remapped_class is cls:
+            return super(LegacyModelMixin, remapped_class).from_pretrained(
+                pretrained_model_name_or_path, **kwargs_copy
+            )
+        else:
+            return remapped_class.from_pretrained(pretrained_model_name_or_path, **kwargs_copy)
@@ -31,6 +31,7 @@ if is_torch_available():
    from .transformer_mochi import MochiTransformer3DModel
    from .transformer_omnigen import OmniGenTransformer2DModel
    from .transformer_sd3 import SD3Transformer2DModel
+    from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
    from .transformer_temporal import TransformerTemporalModel
    from .transformer_wan import WanTransformer3DModel
    from .transformer_wan_vace import WanVACETransformer3DModel
@@ -24,19 +24,13 @@ from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, Pe
 from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.import_utils import is_torch_npu_available
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import FeedForward
-from ..attention_processor import (
-    Attention,
-    AttentionProcessor,
-    FluxAttnProcessor2_0,
-    FluxAttnProcessor2_0_NPU,
-    FusedFluxAttnProcessor2_0,
-)
+from ..attention import AttentionMixin, FeedForward
 from ..cache_utils import CacheMixin
 from ..embeddings import FluxPosEmbed, PixArtAlphaTextProjection, Timesteps, get_timestep_embedding
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import CombinedTimestepLabelEmbeddings, FP32LayerNorm, RMSNorm
+from .transformer_flux import FluxAttention, FluxAttnProcessor


 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -223,6 +217,8 @@ class ChromaSingleTransformerBlock(nn.Module):
        self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)

        if is_torch_npu_available():
+            from ..attention_processor import FluxAttnProcessor2_0_NPU
+
            deprecation_message = (
                "Defaulting to FluxAttnProcessor2_0_NPU for NPU devices will be removed. Attention processors "
                "should be set explicitly using the `set_attn_processor` method."
@@ -230,17 +226,15 @@ class ChromaSingleTransformerBlock(nn.Module):
            deprecate("npu_processor", "0.34.0", deprecation_message)
            processor = FluxAttnProcessor2_0_NPU()
        else:
-            processor = FluxAttnProcessor2_0()
+            processor = FluxAttnProcessor()

-        self.attn = Attention(
+        self.attn = FluxAttention(
            query_dim=dim,
-            cross_attention_dim=None,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            bias=True,
            processor=processor,
-            qk_norm="rms_norm",
            eps=1e-6,
            pre_only=True,
        )
@@ -292,17 +286,15 @@ class ChromaTransformerBlock(nn.Module):
        self.norm1 = ChromaAdaLayerNormZeroPruned(dim)
        self.norm1_context = ChromaAdaLayerNormZeroPruned(dim)

-        self.attn = Attention(
+        self.attn = FluxAttention(
            query_dim=dim,
-            cross_attention_dim=None,
            added_kv_proj_dim=dim,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            context_pre_only=False,
            bias=True,
-            processor=FluxAttnProcessor2_0(),
-            qk_norm=qk_norm,
+            processor=FluxAttnProcessor(),
            eps=eps,
        )

@@ -376,7 +368,13 @@ class ChromaTransformerBlock(nn.Module):


 class ChromaTransformer2DModel(
-    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin, CacheMixin
+    ModelMixin,
+    ConfigMixin,
+    PeftAdapterMixin,
+    FromOriginalModelMixin,
+    FluxTransformer2DLoadersMixin,
+    CacheMixin,
+    AttentionMixin,
 ):
    """
    The Transformer model introduced in Flux, modified for Chroma.
@@ -475,106 +473,6 @@ class ChromaTransformer2DModel(

        self.gradient_checkpointing = False

-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
-        are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-        self.set_attn_processor(FusedFluxAttnProcessor2_0())
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
-
    def forward(
        self,
        hidden_states: torch.Tensor,
@@ -12,28 +12,28 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-
-from typing import Any, Dict, Optional, Tuple, Union
+import inspect
+from typing import Any, Dict, List, Optional, Tuple, Union

 import numpy as np
 import torch
 import torch.nn as nn
+import torch.nn.functional as F

 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.import_utils import is_torch_npu_available
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..attention import FeedForward
-from ..attention_processor import (
-    Attention,
-    AttentionProcessor,
-    FluxAttnProcessor2_0,
-    FluxAttnProcessor2_0_NPU,
-    FusedFluxAttnProcessor2_0,
-)
+from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
+from ..attention_dispatch import dispatch_attention_fn
 from ..cache_utils import CacheMixin
-from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
+from ..embeddings import (
+    CombinedTimestepGuidanceTextProjEmbeddings,
+    CombinedTimestepTextProjEmbeddings,
+    apply_rotary_emb,
+    get_1d_rotary_pos_embed,
+)
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
@@ -42,6 +42,307 @@ from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNo
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


+def _get_projections(attn: "FluxAttention", hidden_states, encoder_hidden_states=None):
+    query = attn.to_q(hidden_states)
+    key = attn.to_k(hidden_states)
+    value = attn.to_v(hidden_states)
+
+    encoder_query = encoder_key = encoder_value = None
+    if encoder_hidden_states is not None and attn.added_kv_proj_dim is not None:
+        encoder_query = attn.add_q_proj(encoder_hidden_states)
+        encoder_key = attn.add_k_proj(encoder_hidden_states)
+        encoder_value = attn.add_v_proj(encoder_hidden_states)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_fused_projections(attn: "FluxAttention", hidden_states, encoder_hidden_states=None):
+    query, key, value = attn.to_qkv(hidden_states).chunk(3, dim=-1)
+
+    encoder_query = encoder_key = encoder_value = (None,)
+    if encoder_hidden_states is not None and hasattr(attn, "to_added_qkv"):
+        encoder_query, encoder_key, encoder_value = attn.to_added_qkv(encoder_hidden_states).chunk(3, dim=-1)
+
+    return query, key, value, encoder_query, encoder_key, encoder_value
+
+
+def _get_qkv_projections(attn: "FluxAttention", hidden_states, encoder_hidden_states=None):
+    if attn.fused_projections:
+        return _get_fused_projections(attn, hidden_states, encoder_hidden_states)
+    return _get_projections(attn, hidden_states, encoder_hidden_states)
+
+
+class FluxAttnProcessor:
+    _attention_backend = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
+
+    def __call__(
+        self,
+        attn: "FluxAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
+            attn, hidden_states, encoder_hidden_states
+        )
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if attn.added_kv_proj_dim is not None:
+            encoder_query = encoder_query.unflatten(-1, (attn.heads, -1))
+            encoder_key = encoder_key.unflatten(-1, (attn.heads, -1))
+            encoder_value = encoder_value.unflatten(-1, (attn.heads, -1))
+
+            encoder_query = attn.norm_added_q(encoder_query)
+            encoder_key = attn.norm_added_k(encoder_key)
+
+            query = torch.cat([encoder_query, query], dim=1)
+            key = torch.cat([encoder_key, key], dim=1)
+            value = torch.cat([encoder_value, value], dim=1)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        hidden_states = dispatch_attention_fn(
+            query, key, value, attn_mask=attention_mask, backend=self._attention_backend
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+            hidden_states = attn.to_out[0](hidden_states)
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
+class FluxIPAdapterAttnProcessor(torch.nn.Module):
+    """Flux Attention processor for IP-Adapter."""
+
+    _attention_backend = None
+
+    def __init__(
+        self, hidden_size: int, cross_attention_dim: int, num_tokens=(4,), scale=1.0, device=None, dtype=None
+    ):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+
+        if not isinstance(num_tokens, (tuple, list)):
+            num_tokens = [num_tokens]
+
+        if not isinstance(scale, list):
+            scale = [scale] * len(num_tokens)
+        if len(scale) != len(num_tokens):
+            raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
+        self.scale = scale
+
+        self.to_k_ip = nn.ModuleList(
+            [
+                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
+                for _ in range(len(num_tokens))
+            ]
+        )
+        self.to_v_ip = nn.ModuleList(
+            [
+                nn.Linear(cross_attention_dim, hidden_size, bias=True, device=device, dtype=dtype)
+                for _ in range(len(num_tokens))
+            ]
+        )
+
+    def __call__(
+        self,
+        attn: "FluxAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        ip_hidden_states: Optional[List[torch.Tensor]] = None,
+        ip_adapter_masks: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size = hidden_states.shape[0]
+
+        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
+            attn, hidden_states, encoder_hidden_states
+        )
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+        ip_query = query
+
+        if encoder_hidden_states is not None:
+            encoder_query = encoder_query.unflatten(-1, (attn.heads, -1))
+            encoder_key = encoder_key.unflatten(-1, (attn.heads, -1))
+            encoder_value = encoder_value.unflatten(-1, (attn.heads, -1))
+
+            encoder_query = attn.norm_added_q(encoder_query)
+            encoder_key = attn.norm_added_k(encoder_key)
+
+            query = torch.cat([encoder_query, query], dim=1)
+            key = torch.cat([encoder_key, key], dim=1)
+            value = torch.cat([encoder_value, value], dim=1)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+            hidden_states = attn.to_out[0](hidden_states)
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            # IP-adapter
+            ip_attn_output = torch.zeros_like(hidden_states)
+
+            for current_ip_hidden_states, scale, to_k_ip, to_v_ip in zip(
+                ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip
+            ):
+                ip_key = to_k_ip(current_ip_hidden_states)
+                ip_value = to_v_ip(current_ip_hidden_states)
+
+                ip_key = ip_key.view(batch_size, -1, attn.heads, attn.head_dim)
+                ip_value = ip_value.view(batch_size, -1, attn.heads, attn.head_dim)
+
+                current_ip_hidden_states = dispatch_attention_fn(
+                    ip_query,
+                    ip_key,
+                    ip_value,
+                    attn_mask=None,
+                    dropout_p=0.0,
+                    is_causal=False,
+                    backend=self._attention_backend,
+                )
+                current_ip_hidden_states = current_ip_hidden_states.reshape(batch_size, -1, attn.heads * attn.head_dim)
+                current_ip_hidden_states = current_ip_hidden_states.to(ip_query.dtype)
+                ip_attn_output += scale * current_ip_hidden_states
+
+            return hidden_states, encoder_hidden_states, ip_attn_output
+        else:
+            return hidden_states
+
+
+class FluxAttention(torch.nn.Module, AttentionModuleMixin):
+    _default_processor_cls = FluxAttnProcessor
+    _available_processors = [
+        FluxAttnProcessor,
+        FluxIPAdapterAttnProcessor,
+    ]
+
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        out_bias: bool = True,
+        eps: float = 1e-5,
+        out_dim: int = None,
+        context_pre_only: Optional[bool] = None,
+        pre_only: bool = False,
+        elementwise_affine: bool = True,
+        processor=None,
+    ):
+        super().__init__()
+
+        self.head_dim = dim_head
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.dropout = dropout
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.context_pre_only = context_pre_only
+        self.pre_only = pre_only
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.added_proj_bias = added_proj_bias
+
+        self.norm_q = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.norm_k = torch.nn.RMSNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine)
+        self.to_q = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_k = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+        self.to_v = torch.nn.Linear(query_dim, self.inner_dim, bias=bias)
+
+        if not self.pre_only:
+            self.to_out = torch.nn.ModuleList([])
+            self.to_out.append(torch.nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
+            self.to_out.append(torch.nn.Dropout(dropout))
+
+        if added_kv_proj_dim is not None:
+            self.norm_added_q = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.norm_added_k = torch.nn.RMSNorm(dim_head, eps=eps)
+            self.add_q_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+            self.to_add_out = torch.nn.Linear(self.inner_dim, query_dim, bias=out_bias)
+
+        if processor is None:
+            processor = self._default_processor_cls()
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"}
+        unused_kwargs = [k for k, _ in kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"joint_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
+        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs)
+
+
@maybe_allow_in_graph
 class FluxSingleTransformerBlock(nn.Module):
    def __init__(self, dim: int, num_attention_heads: int, attention_head_dim: int, mlp_ratio: float = 4.0):
@@ -54,6 +355,8 @@ class FluxSingleTransformerBlock(nn.Module):
        self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)

        if is_torch_npu_available():
+            from ..attention_processor import FluxAttnProcessor2_0_NPU
+
            deprecation_message = (
                "Defaulting to FluxAttnProcessor2_0_NPU for NPU devices will be removed. Attention processors "
                "should be set explicitly using the `set_attn_processor` method."
@@ -61,17 +364,15 @@ class FluxSingleTransformerBlock(nn.Module):
            deprecate("npu_processor", "0.34.0", deprecation_message)
            processor = FluxAttnProcessor2_0_NPU()
        else:
-            processor = FluxAttnProcessor2_0()
+            processor = FluxAttnProcessor()

-        self.attn = Attention(
+        self.attn = FluxAttention(
            query_dim=dim,
-            cross_attention_dim=None,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            bias=True,
            processor=processor,
-            qk_norm="rms_norm",
            eps=1e-6,
            pre_only=True,
        )
@@ -118,17 +419,15 @@ class FluxTransformerBlock(nn.Module):
        self.norm1 = AdaLayerNormZero(dim)
        self.norm1_context = AdaLayerNormZero(dim)

-        self.attn = Attention(
+        self.attn = FluxAttention(
            query_dim=dim,
-            cross_attention_dim=None,
            added_kv_proj_dim=dim,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            context_pre_only=False,
            bias=True,
-            processor=FluxAttnProcessor2_0(),
-            qk_norm=qk_norm,
+            processor=FluxAttnProcessor(),
            eps=eps,
        )

@@ -152,6 +451,7 @@ class FluxTransformerBlock(nn.Module):
            encoder_hidden_states, emb=temb
        )
        joint_attention_kwargs = joint_attention_kwargs or {}
+
        # Attention.
        attention_outputs = self.attn(
            hidden_states=norm_hidden_states,
@@ -180,7 +480,6 @@ class FluxTransformerBlock(nn.Module):
            hidden_states = hidden_states + ip_attn_output

        # Process attention outputs for the `encoder_hidden_states`.
-
        context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
        encoder_hidden_states = encoder_hidden_states + context_attn_output

@@ -195,8 +494,45 @@ class FluxTransformerBlock(nn.Module):
        return encoder_hidden_states, hidden_states


+class FluxPosEmbed(nn.Module):
+    # modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        cos_out = []
+        sin_out = []
+        pos = ids.float()
+        is_mps = ids.device.type == "mps"
+        is_npu = ids.device.type == "npu"
+        freqs_dtype = torch.float32 if (is_mps or is_npu) else torch.float64
+        for i in range(n_axes):
+            cos, sin = get_1d_rotary_pos_embed(
+                self.axes_dim[i],
+                pos[:, i],
+                theta=self.theta,
+                repeat_interleave_real=True,
+                use_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            cos_out.append(cos)
+            sin_out.append(sin)
+        freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
+        freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
+        return freqs_cos, freqs_sin
+
+
 class FluxTransformer2DModel(
-    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin, CacheMixin
+    ModelMixin,
+    ConfigMixin,
+    PeftAdapterMixin,
+    FromOriginalModelMixin,
+    FluxTransformer2DLoadersMixin,
+    CacheMixin,
+    AttentionMixin,
 ):
    """
    The Transformer model introduced in Flux.
@@ -292,106 +628,6 @@ class FluxTransformer2DModel(

        self.gradient_checkpointing = False

-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
-        are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-        self.set_attn_processor(FusedFluxAttnProcessor2_0())
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
-
    def forward(
        self,
        hidden_states: torch.Tensor,
@@ -0,0 +1,607 @@
+# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ..attention import FeedForward
+from ..attention_processor import Attention
+from ..cache_utils import CacheMixin
+from ..embeddings import (
+    PixArtAlphaTextProjection,
+    TimestepEmbedding,
+    get_1d_rotary_pos_embed,
+    get_1d_sincos_pos_embed_from_grid,
+)
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin, get_parameter_dtype
+from ..normalization import FP32LayerNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class SkyReelsV2AttnProcessor2_0:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "SkyReelsV2AttnProcessor2_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,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            # 512 is the context length of the text encoder, hardcoded for now
+            image_context_length = encoder_hidden_states.shape[1] - 512
+            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
+            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(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)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
+                x_rotated = torch.view_as_complex(hidden_states.to(torch.float32).unflatten(3, (-1, 2)))
+                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
+                return x_out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, rotary_emb)
+            key = apply_rotary_emb(key, rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img = attn.add_k_proj(encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+            value_img = attn.add_v_proj(encoder_hidden_states_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            hidden_states_img = F.scaled_dot_product_attention(
+                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        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).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanImageEmbedding with WanImageEmbedding -> SkyReelsV2ImageEmbedding
+class SkyReelsV2ImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int, pos_embed_seq_len=None):
+        super().__init__()
+
+        self.norm1 = FP32LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = FP32LayerNorm(out_features)
+        if pos_embed_seq_len is not None:
+            self.pos_embed = nn.Parameter(torch.zeros(1, pos_embed_seq_len, in_features))
+        else:
+            self.pos_embed = None
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        if self.pos_embed is not None:
+            batch_size, seq_len, embed_dim = encoder_hidden_states_image.shape
+            encoder_hidden_states_image = encoder_hidden_states_image.view(-1, 2 * seq_len, embed_dim)
+            encoder_hidden_states_image = encoder_hidden_states_image + self.pos_embed
+
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+class SkyReelsV2Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool, output_type: str = "pt"):
+        super().__init__()
+        self.num_channels = num_channels
+        self.output_type = output_type
+        self.flip_sin_to_cos = flip_sin_to_cos
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        original_shape = timesteps.shape
+        t_emb = get_1d_sincos_pos_embed_from_grid(
+            self.num_channels,
+            timesteps,
+            output_type=self.output_type,
+            flip_sin_to_cos=self.flip_sin_to_cos,
+        )
+        # Reshape back to maintain batch structure
+        if len(original_shape) > 1:
+            t_emb = t_emb.reshape(*original_shape, self.num_channels)
+        return t_emb
+
+
+class SkyReelsV2TimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+        pos_embed_seq_len: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = SkyReelsV2Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = SkyReelsV2ImageEmbedding(image_embed_dim, dim, pos_embed_seq_len=pos_embed_seq_len)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+
+        time_embedder_dtype = get_parameter_dtype(self.time_embedder)
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class SkyReelsV2RotaryPosEmbed(nn.Module):
+    def __init__(
+        self, attention_head_dim: int, patch_size: Tuple[int, int, int], max_seq_len: int, theta: float = 10000.0
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+
+        freqs = []
+        for dim in [t_dim, h_dim, w_dim]:
+            freq = get_1d_rotary_pos_embed(
+                dim, max_seq_len, theta, use_real=False, repeat_interleave_real=False, freqs_dtype=torch.float32
+            )
+            freqs.append(freq)
+        self.freqs = torch.cat(freqs, dim=1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        freqs = self.freqs.to(hidden_states.device)
+        freqs = freqs.split_with_sizes(
+            [
+                self.attention_head_dim // 2 - 2 * (self.attention_head_dim // 6),
+                self.attention_head_dim // 6,
+                self.attention_head_dim // 6,
+            ],
+            dim=1,
+        )
+
+        freqs_f = freqs[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_h = freqs[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_w = freqs[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+        freqs = torch.cat([freqs_f, freqs_h, freqs_w], dim=-1).reshape(1, 1, ppf * pph * ppw, -1)
+        return freqs
+
+
+class SkyReelsV2TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            processor=SkyReelsV2AttnProcessor2_0(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = Attention(
+            query_dim=dim,
+            heads=num_heads,
+            kv_heads=num_heads,
+            dim_head=dim // num_heads,
+            qk_norm=qk_norm,
+            eps=eps,
+            bias=True,
+            cross_attention_dim=None,
+            out_bias=True,
+            added_kv_proj_dim=added_kv_proj_dim,
+            added_proj_bias=True,
+            processor=SkyReelsV2AttnProcessor2_0(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+        attention_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        if temb.dim() == 3:
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table + temb.float()
+            ).chunk(6, dim=1)
+        elif temb.dim() == 4:
+            # For 4D temb in Diffusion Forcing framework, we assume the shape is  (b, 6, f * pp_h * pp_w, inner_dim)
+            e = (self.scale_shift_table.unsqueeze(2) + temb.float()).chunk(6, dim=1)
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = [ei.squeeze(1) for ei in e]
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(
+            hidden_states=norm_hidden_states, rotary_emb=rotary_emb, attention_mask=attention_mask
+        )
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+        return hidden_states
+
+
+class SkyReelsV2Transformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    A Transformer model for video-like data used in the Wan-based SkyReels-V2 model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `16`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `4096`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `8192`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `32`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`str`, *optional*, defaults to `"rms_norm_across_heads"`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        inject_sample_info (`bool`, defaults to `False`):
+            Whether to inject sample information into the model.
+        image_dim (`int`, *optional*):
+            The dimension of the image embeddings.
+        added_kv_proj_dim (`int`, *optional*):
+            The dimension of the added key/value projection.
+        rope_max_seq_len (`int`, defaults to `1024`):
+            The maximum sequence length for the rotary embeddings.
+        pos_embed_seq_len (`int`, *optional*):
+            The sequence length for the positional embeddings.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["SkyReelsV2TransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 8192,
+        num_layers: int = 32,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        pos_embed_seq_len: Optional[int] = None,
+        inject_sample_info: bool = False,
+        num_frame_per_block: int = 1,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = SkyReelsV2RotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = SkyReelsV2TimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+            pos_embed_seq_len=pos_embed_seq_len,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                SkyReelsV2TransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        if inject_sample_info:
+            self.fps_embedding = nn.Embedding(2, inner_dim)
+            self.fps_projection = FeedForward(inner_dim, inner_dim * 6, mult=1, activation_fn="linear-silu")
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        enable_diffusion_forcing: bool = False,
+        fps: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        causal_mask = None
+        if self.config.num_frame_per_block > 1:
+            block_num = post_patch_num_frames // self.config.num_frame_per_block
+            range_tensor = torch.arange(block_num, device=hidden_states.device).repeat_interleave(
+                self.config.num_frame_per_block
+            )
+            causal_mask = range_tensor.unsqueeze(0) <= range_tensor.unsqueeze(1)  # f, f
+            causal_mask = causal_mask.view(post_patch_num_frames, 1, 1, post_patch_num_frames, 1, 1)
+            causal_mask = causal_mask.repeat(
+                1, post_patch_height, post_patch_width, 1, post_patch_height, post_patch_width
+            )
+            causal_mask = causal_mask.reshape(
+                post_patch_num_frames * post_patch_height * post_patch_width,
+                post_patch_num_frames * post_patch_height * post_patch_width,
+            )
+            causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+
+        timestep_proj = timestep_proj.unflatten(-1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        if self.config.inject_sample_info:
+            fps = torch.tensor(fps, dtype=torch.long, device=hidden_states.device)
+
+            fps_emb = self.fps_embedding(fps)
+            if enable_diffusion_forcing:
+                timestep_proj = timestep_proj + self.fps_projection(fps_emb).unflatten(1, (6, -1)).repeat(
+                    timestep.shape[1], 1, 1
+                )
+            else:
+                timestep_proj = timestep_proj + self.fps_projection(fps_emb).unflatten(1, (6, -1))
+
+        if enable_diffusion_forcing:
+            b, f = timestep.shape
+            temb = temb.view(b, f, 1, 1, -1)
+            timestep_proj = timestep_proj.view(b, f, 1, 1, 6, -1)  # (b, f, 1, 1, 6, inner_dim)
+            temb = temb.repeat(1, 1, post_patch_height, post_patch_width, 1).flatten(1, 3)
+            timestep_proj = timestep_proj.repeat(1, 1, post_patch_height, post_patch_width, 1, 1).flatten(
+                1, 3
+            )  # (b, f, pp_h, pp_w, 6, inner_dim) -> (b, f * pp_h * pp_w, 6, inner_dim)
+            timestep_proj = timestep_proj.transpose(1, 2).contiguous()  # (b, 6, f * pp_h * pp_w, inner_dim)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    timestep_proj,
+                    rotary_emb,
+                    causal_mask,
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states,
+                    timestep_proj,
+                    rotary_emb,
+                    causal_mask,
+                )
+
+        if temb.dim() == 2:
+            # If temb is 2D, we assume it has time 1-D time embedding values for each batch.
+            # For models:
+            # - Skywork/SkyReels-V2-T2V-14B-540P-Diffusers
+            # - Skywork/SkyReels-V2-T2V-14B-720P-Diffusers
+            # - Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers
+            # - Skywork/SkyReels-V2-I2V-14B-540P-Diffusers
+            # - Skywork/SkyReels-V2-I2V-14B-720P-Diffusers
+            shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+        elif temb.dim() == 3:
+            # If temb is 3D, we assume it has 2-D time embedding values for each batch.
+            # Each time embedding tensor includes values for each latent frame; thus Diffusion Forcing.
+            # For models:
+            # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers
+            # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers
+            # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers
+            shift, scale = (self.scale_shift_table.unsqueeze(2) + temb.unsqueeze(1)).chunk(2, dim=1)
+            shift, scale = shift.squeeze(1), scale.squeeze(1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
+
+    def _set_ar_attention(self, causal_block_size: int):
+        self.register_to_config(num_frame_per_block=causal_block_size)
@@ -165,7 +165,7 @@ class UNet2DConditionModel(
    """

    _supports_gradient_checkpointing = True
-    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"]
    _skip_layerwise_casting_patterns = ["norm"]
    _repeated_blocks = ["BasicTransformerBlock"]

@@ -40,6 +40,7 @@ else:
        "InsertableDict",
    ]
    _import_structure["stable_diffusion_xl"] = ["StableDiffusionXLAutoBlocks", "StableDiffusionXLModularPipeline"]
+    _import_structure["wan"] = ["WanAutoBlocks", "WanModularPipeline"]
    _import_structure["components_manager"] = ["ComponentsManager"]

 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -71,6 +72,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
            StableDiffusionXLAutoBlocks,
            StableDiffusionXLModularPipeline,
        )
+        from .wan import WanAutoBlocks, WanModularPipeline
 else:
    import sys

@@ -386,6 +386,7 @@ class ComponentsManager:
                 id(component) is Python's built-in unique identifier for the object
        """
        component_id = f"{name}_{id(component)}"
+        is_new_component = True

        # check for duplicated components
        for comp_id, comp in self.components.items():
@@ -394,6 +395,7 @@ class ComponentsManager:
                if comp_name == name:
                    logger.warning(f"ComponentsManager: component '{name}' already exists as '{comp_id}'")
                    component_id = comp_id
+                    is_new_component = False
                    break
                else:
                    logger.warning(
@@ -426,7 +428,9 @@ class ComponentsManager:
                    logger.warning(
                        f"ComponentsManager: removing existing {name} from collection '{collection}': {comp_id}"
                    )
-                    self.remove(comp_id)
+                    # remove existing component from this collection (if it is not in any other collection, will be removed from ComponentsManager)
+                    self.remove_from_collection(comp_id, collection)
+
                self.collections[collection].add(component_id)
                logger.info(
                    f"ComponentsManager: added component '{name}' in collection '{collection}': {component_id}"
@@ -434,11 +438,29 @@ class ComponentsManager:
        else:
            logger.info(f"ComponentsManager: added component '{name}' as '{component_id}'")

-        if self._auto_offload_enabled:
+        if self._auto_offload_enabled and is_new_component:
            self.enable_auto_cpu_offload(self._auto_offload_device)

        return component_id

+    def remove_from_collection(self, component_id: str, collection: str):
+        """
+        Remove a component from a collection.
+        """
+        if collection not in self.collections:
+            logger.warning(f"Collection '{collection}' not found in ComponentsManager")
+            return
+        if component_id not in self.collections[collection]:
+            logger.warning(f"Component '{component_id}' not found in collection '{collection}'")
+            return
+        # remove from the collection
+        self.collections[collection].remove(component_id)
+        # check if this component is in any other collection
+        comp_colls = [coll for coll, comps in self.collections.items() if component_id in comps]
+        if not comp_colls:  # only if no other collection contains this component, remove it
+            logger.warning(f"ComponentsManager: removing component '{component_id}' from ComponentsManager")
+            self.remove(component_id)
+
    def remove(self, component_id: str = None):
        """
        Remove a component from the ComponentsManager.
@@ -60,12 +60,14 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 MODULAR_PIPELINE_MAPPING = OrderedDict(
    [
        ("stable-diffusion-xl", "StableDiffusionXLModularPipeline"),
+        ("wan", "WanModularPipeline"),
    ]
 )

 MODULAR_PIPELINE_BLOCKS_MAPPING = OrderedDict(
    [
        ("StableDiffusionXLModularPipeline", "StableDiffusionXLAutoBlocks"),
+        ("WanModularPipeline", "WanAutoBlocks"),
    ]
 )

@@ -322,9 +324,12 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
    </Tip>
    """

-    config_name = "config.json"
+    config_name = "modular_config.json"
    model_name = None

+    def __init__(self):
+        self.sub_blocks = InsertableDict()
+
    @classmethod
    def _get_signature_keys(cls, obj):
        parameters = inspect.signature(obj.__init__).parameters
@@ -342,6 +347,11 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
    def expected_configs(self) -> List[ConfigSpec]:
        return []

+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        """List of intermediate output parameters. Must be implemented by subclasses."""
+        return []
+
    @classmethod
    def from_pretrained(
        cls,
@@ -423,6 +433,60 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
        )
        return modular_pipeline

+    def get_block_state(self, state: PipelineState) -> dict:
+        """Get all inputs and intermediates in one dictionary"""
+        data = {}
+        state_inputs = self.inputs + self.intermediate_inputs
+
+        # Check inputs
+        for input_param in state_inputs:
+            if input_param.name:
+                value = state.get_input(input_param.name) or state.get_intermediate(input_param.name)
+                if input_param.required and value is None:
+                    raise ValueError(f"Required input '{input_param.name}' is missing")
+                elif value is not None or (value is None and input_param.name not in data):
+                    data[input_param.name] = value
+
+            elif input_param.kwargs_type:
+                # if kwargs_type is provided, get all inputs with matching kwargs_type
+                if input_param.kwargs_type not in data:
+                    data[input_param.kwargs_type] = {}
+                inputs_kwargs = state.get_inputs_kwargs(input_param.kwargs_type) or state.get_intermediate_kwargs(
+                    input_param.kwargs_type
+                )
+                if inputs_kwargs:
+                    for k, v in inputs_kwargs.items():
+                        if v is not None:
+                            data[k] = v
+                            data[input_param.kwargs_type][k] = v
+
+        return BlockState(**data)
+
+    def set_block_state(self, state: PipelineState, block_state: BlockState):
+        for output_param in self.intermediate_outputs:
+            if not hasattr(block_state, output_param.name):
+                raise ValueError(f"Intermediate output '{output_param.name}' is missing in block state")
+            param = getattr(block_state, output_param.name)
+            state.set_intermediate(output_param.name, param, output_param.kwargs_type)
+
+        for input_param in self.intermediate_inputs:
+            if input_param.name and hasattr(block_state, input_param.name):
+                param = getattr(block_state, input_param.name)
+                # Only add if the value is different from what's in the state
+                current_value = state.get_intermediate(input_param.name)
+                if current_value is not param:  # Using identity comparison to check if object was modified
+                    state.set_intermediate(input_param.name, param, input_param.kwargs_type)
+            elif input_param.kwargs_type:
+                # if it is a kwargs type, e.g. "guider_input_fields", it is likely to be a list of parameters
+                # we need to first find out which inputs are and loop through them.
+                intermediate_kwargs = state.get_intermediate_kwargs(input_param.kwargs_type)
+                for param_name, current_value in intermediate_kwargs.items():
+                    if not hasattr(block_state, param_name):
+                        continue
+                    param = getattr(block_state, param_name)
+                    if current_value is not param:  # Using identity comparison to check if object was modified
+                        state.set_intermediate(param_name, param, input_param.kwargs_type)
+
    @staticmethod
    def combine_inputs(*named_input_lists: List[Tuple[str, List[InputParam]]]) -> List[InputParam]:
        """
@@ -652,51 +716,6 @@ class PipelineBlock(ModularPipelineBlocks):
            expected_configs=self.expected_configs,
        )

-    # YiYi TODO: input and inteermediate inputs with same name? should warn?
-    def get_block_state(self, state: PipelineState) -> dict:
-        """Get all inputs and intermediates in one dictionary"""
-        data = {}
-
-        # Check inputs
-        for input_param in self.inputs:
-            if input_param.name:
-                value = state.get_input(input_param.name)
-                if input_param.required and value is None:
-                    raise ValueError(f"Required input '{input_param.name}' is missing")
-                elif value is not None or (value is None and input_param.name not in data):
-                    data[input_param.name] = value
-            elif input_param.kwargs_type:
-                # if kwargs_type is provided, get all inputs with matching kwargs_type
-                if input_param.kwargs_type not in data:
-                    data[input_param.kwargs_type] = {}
-                inputs_kwargs = state.get_inputs_kwargs(input_param.kwargs_type)
-                if inputs_kwargs:
-                    for k, v in inputs_kwargs.items():
-                        if v is not None:
-                            data[k] = v
-                            data[input_param.kwargs_type][k] = v
-
-        # Check intermediates
-        for input_param in self.intermediate_inputs:
-            if input_param.name:
-                value = state.get_intermediate(input_param.name)
-                if input_param.required and value is None:
-                    raise ValueError(f"Required intermediate input '{input_param.name}' is missing")
-                elif value is not None or (value is None and input_param.name not in data):
-                    data[input_param.name] = value
-            elif input_param.kwargs_type:
-                # if kwargs_type is provided, get all intermediates with matching kwargs_type
-                if input_param.kwargs_type not in data:
-                    data[input_param.kwargs_type] = {}
-                intermediate_kwargs = state.get_intermediate_kwargs(input_param.kwargs_type)
-                if intermediate_kwargs:
-                    for k, v in intermediate_kwargs.items():
-                        if v is not None:
-                            if k not in data:
-                                data[k] = v
-                            data[input_param.kwargs_type][k] = v
-        return BlockState(**data)
-
    def set_block_state(self, state: PipelineState, block_state: BlockState):
        for output_param in self.intermediate_outputs:
            if not hasattr(block_state, output_param.name):
@@ -1437,11 +1456,6 @@ class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
        """List of input parameters. Must be implemented by subclasses."""
        return []

-    @property
-    def loop_intermediate_inputs(self) -> List[InputParam]:
-        """List of intermediate input parameters. Must be implemented by subclasses."""
-        return []
-
    @property
    def loop_intermediate_outputs(self) -> List[OutputParam]:
        """List of intermediate output parameters. Must be implemented by subclasses."""
@@ -1455,14 +1469,6 @@ class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
                input_names.append(input_param.name)
        return input_names

-    @property
-    def loop_required_intermediate_inputs(self) -> List[str]:
-        input_names = []
-        for input_param in self.loop_intermediate_inputs:
-            if input_param.required:
-                input_names.append(input_param.name)
-        return input_names
-
    # modified from SequentialPipelineBlocks to include loop_expected_components
    @property
    def expected_components(self):
@@ -1633,75 +1639,6 @@ class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
    def __call__(self, components, state: PipelineState) -> PipelineState:
        raise NotImplementedError("`__call__` method needs to be implemented by the subclass")

-    def get_block_state(self, state: PipelineState) -> dict:
-        """Get all inputs and intermediates in one dictionary"""
-        data = {}
-
-        # Check inputs
-        for input_param in self.inputs:
-            if input_param.name:
-                value = state.get_input(input_param.name)
-                if input_param.required and value is None:
-                    raise ValueError(f"Required input '{input_param.name}' is missing")
-                elif value is not None or (value is None and input_param.name not in data):
-                    data[input_param.name] = value
-            elif input_param.kwargs_type:
-                # if kwargs_type is provided, get all inputs with matching kwargs_type
-                if input_param.kwargs_type not in data:
-                    data[input_param.kwargs_type] = {}
-                inputs_kwargs = state.get_inputs_kwargs(input_param.kwargs_type)
-                if inputs_kwargs:
-                    for k, v in inputs_kwargs.items():
-                        if v is not None:
-                            data[k] = v
-                            data[input_param.kwargs_type][k] = v
-
-        # Check intermediates
-        for input_param in self.intermediate_inputs:
-            if input_param.name:
-                value = state.get_intermediate(input_param.name)
-                if input_param.required and value is None:
-                    raise ValueError(f"Required intermediate input '{input_param.name}' is missing")
-                elif value is not None or (value is None and input_param.name not in data):
-                    data[input_param.name] = value
-            elif input_param.kwargs_type:
-                # if kwargs_type is provided, get all intermediates with matching kwargs_type
-                if input_param.kwargs_type not in data:
-                    data[input_param.kwargs_type] = {}
-                intermediate_kwargs = state.get_intermediate_kwargs(input_param.kwargs_type)
-                if intermediate_kwargs:
-                    for k, v in intermediate_kwargs.items():
-                        if v is not None:
-                            if k not in data:
-                                data[k] = v
-                            data[input_param.kwargs_type][k] = v
-        return BlockState(**data)
-
-    def set_block_state(self, state: PipelineState, block_state: BlockState):
-        for output_param in self.intermediate_outputs:
-            if not hasattr(block_state, output_param.name):
-                raise ValueError(f"Intermediate output '{output_param.name}' is missing in block state")
-            param = getattr(block_state, output_param.name)
-            state.set_intermediate(output_param.name, param, output_param.kwargs_type)
-
-        for input_param in self.intermediate_inputs:
-            if input_param.name and hasattr(block_state, input_param.name):
-                param = getattr(block_state, input_param.name)
-                # Only add if the value is different from what's in the state
-                current_value = state.get_intermediate(input_param.name)
-                if current_value is not param:  # Using identity comparison to check if object was modified
-                    state.set_intermediate(input_param.name, param, input_param.kwargs_type)
-            elif input_param.kwargs_type:
-                # if it is a kwargs type, e.g. "guider_input_fields", it is likely to be a list of parameters
-                # we need to first find out which inputs are and loop through them.
-                intermediate_kwargs = state.get_intermediate_kwargs(input_param.kwargs_type)
-                for param_name, current_value in intermediate_kwargs.items():
-                    if not hasattr(block_state, param_name):
-                        continue
-                    param = getattr(block_state, param_name)
-                    if current_value is not param:  # Using identity comparison to check if object was modified
-                        state.set_intermediate(param_name, param, input_param.kwargs_type)
-
    @property
    def doc(self):
        return make_doc_string(
@@ -1974,7 +1911,6 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):

        # Add inputs to state, using defaults if not provided in the kwargs or the state
        # if same input already in the state, will override it if provided in the kwargs
-
        intermediate_inputs = [inp.name for inp in self.blocks.intermediate_inputs]
        for expected_input_param in self.blocks.inputs:
            name = expected_input_param.name
@@ -185,6 +185,8 @@ class ComponentSpec:
        Unique identifier for this spec's pretrained load, composed of repo|subfolder|variant|revision (no empty
        segments).
        """
+        if self.default_creation_method == "from_config":
+            return "null"
        parts = [getattr(self, k) for k in self.loading_fields()]
        parts = ["null" if p is None else p for p in parts]
        return "|".join(p for p in parts if p)
@@ -27,7 +27,7 @@ from ...schedulers import EulerDiscreteScheduler
 from ...utils import logging
 from ...utils.torch_utils import randn_tensor, unwrap_module
 from ..modular_pipeline import (
-    PipelineBlock,
+    ModularPipelineBlocks,
    PipelineState,
 )
 from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
@@ -195,7 +195,7 @@ def prepare_latents_img2img(
    return latents


-class StableDiffusionXLInputStep(PipelineBlock):
+class StableDiffusionXLInputStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -394,7 +394,7 @@ class StableDiffusionXLInputStep(PipelineBlock):
        return components, state


-class StableDiffusionXLImg2ImgSetTimestepsStep(PipelineBlock):
+class StableDiffusionXLImg2ImgSetTimestepsStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -543,7 +543,7 @@ class StableDiffusionXLImg2ImgSetTimestepsStep(PipelineBlock):
        return components, state


-class StableDiffusionXLSetTimestepsStep(PipelineBlock):
+class StableDiffusionXLSetTimestepsStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -611,7 +611,7 @@ class StableDiffusionXLSetTimestepsStep(PipelineBlock):
        return components, state


-class StableDiffusionXLInpaintPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLInpaintPrepareLatentsStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -900,7 +900,7 @@ class StableDiffusionXLInpaintPrepareLatentsStep(PipelineBlock):
        return components, state


-class StableDiffusionXLImg2ImgPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLImg2ImgPrepareLatentsStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -981,7 +981,7 @@ class StableDiffusionXLImg2ImgPrepareLatentsStep(PipelineBlock):
        return components, state


-class StableDiffusionXLPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLPrepareLatentsStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -1092,7 +1092,7 @@ class StableDiffusionXLPrepareLatentsStep(PipelineBlock):
        return components, state


-class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(PipelineBlock):
+class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -1316,7 +1316,7 @@ class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(PipelineBlock):
        return components, state


-class StableDiffusionXLPrepareAdditionalConditioningStep(PipelineBlock):
+class StableDiffusionXLPrepareAdditionalConditioningStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -1499,7 +1499,7 @@ class StableDiffusionXLPrepareAdditionalConditioningStep(PipelineBlock):
        return components, state


-class StableDiffusionXLControlNetInputStep(PipelineBlock):
+class StableDiffusionXLControlNetInputStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -1718,7 +1718,7 @@ class StableDiffusionXLControlNetInputStep(PipelineBlock):
        return components, state


-class StableDiffusionXLControlNetUnionInputStep(PipelineBlock):
+class StableDiffusionXLControlNetUnionInputStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -23,17 +23,14 @@ from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL
 from ...models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from ...utils import logging
-from ..modular_pipeline import (
-    PipelineBlock,
-    PipelineState,
-)
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam


 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


-class StableDiffusionXLDecodeStep(PipelineBlock):
+class StableDiffusionXLDecodeStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -157,7 +154,7 @@ class StableDiffusionXLDecodeStep(PipelineBlock):
        return components, state


-class StableDiffusionXLInpaintOverlayMaskStep(PipelineBlock):
+class StableDiffusionXLInpaintOverlayMaskStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -25,7 +25,7 @@ from ...utils import logging
 from ..modular_pipeline import (
    BlockState,
    LoopSequentialPipelineBlocks,
-    PipelineBlock,
+    ModularPipelineBlocks,
    PipelineState,
 )
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
@@ -37,7 +37,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

 # YiYi experimenting composible denoise loop
 # loop step (1): prepare latent input for denoiser
-class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):
+class StableDiffusionXLLoopBeforeDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -55,7 +55,7 @@ class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):
        )

    @property
-    def intermediate_inputs(self) -> List[str]:
+    def inputs(self) -> List[str]:
        return [
            InputParam(
                "latents",
@@ -73,7 +73,7 @@ class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):


 # loop step (1): prepare latent input for denoiser (with inpainting)
-class StableDiffusionXLInpaintLoopBeforeDenoiser(PipelineBlock):
+class StableDiffusionXLInpaintLoopBeforeDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -144,7 +144,7 @@ class StableDiffusionXLInpaintLoopBeforeDenoiser(PipelineBlock):


 # loop step (2): denoise the latents with guidance
-class StableDiffusionXLLoopDenoiser(PipelineBlock):
+class StableDiffusionXLLoopDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -249,7 +249,7 @@ class StableDiffusionXLLoopDenoiser(PipelineBlock):


 # loop step (2): denoise the latents with guidance (with controlnet)
-class StableDiffusionXLControlNetLoopDenoiser(PipelineBlock):
+class StableDiffusionXLControlNetLoopDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -449,7 +449,7 @@ class StableDiffusionXLControlNetLoopDenoiser(PipelineBlock):


 # loop step (3): scheduler step to update latents
-class StableDiffusionXLLoopAfterDenoiser(PipelineBlock):
+class StableDiffusionXLLoopAfterDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -520,7 +520,7 @@ class StableDiffusionXLLoopAfterDenoiser(PipelineBlock):


 # loop step (3): scheduler step to update latents (with inpainting)
-class StableDiffusionXLInpaintLoopAfterDenoiser(PipelineBlock):
+class StableDiffusionXLInpaintLoopAfterDenoiser(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -660,7 +660,7 @@ class StableDiffusionXLDenoiseLoopWrapper(LoopSequentialPipelineBlocks):
        ]

    @property
-    def loop_intermediate_inputs(self) -> List[InputParam]:
+    def loop_inputs(self) -> List[InputParam]:
        return [
            InputParam(
                "timesteps",
@@ -35,7 +35,7 @@ from ...utils import (
    scale_lora_layers,
    unscale_lora_layers,
 )
-from ..modular_pipeline import PipelineBlock, PipelineState
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
 from .modular_pipeline import StableDiffusionXLModularPipeline

@@ -57,7 +57,7 @@ def retrieve_latents(
        raise AttributeError("Could not access latents of provided encoder_output")


-class StableDiffusionXLIPAdapterStep(PipelineBlock):
+class StableDiffusionXLIPAdapterStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -215,7 +215,7 @@ class StableDiffusionXLIPAdapterStep(PipelineBlock):
        return components, state


-class StableDiffusionXLTextEncoderStep(PipelineBlock):
+class StableDiffusionXLTextEncoderStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -576,7 +576,7 @@ class StableDiffusionXLTextEncoderStep(PipelineBlock):
        return components, state


-class StableDiffusionXLVaeEncoderStep(PipelineBlock):
+class StableDiffusionXLVaeEncoderStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -691,7 +691,7 @@ class StableDiffusionXLVaeEncoderStep(PipelineBlock):
        return components, state


-class StableDiffusionXLInpaintVaeEncoderStep(PipelineBlock):
+class StableDiffusionXLInpaintVaeEncoderStep(ModularPipelineBlocks):
    model_name = "stable-diffusion-xl"

    @property
@@ -0,0 +1,66 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["encoders"] = ["WanTextEncoderStep"]
+    _import_structure["modular_blocks"] = [
+        "ALL_BLOCKS",
+        "AUTO_BLOCKS",
+        "TEXT2VIDEO_BLOCKS",
+        "WanAutoBeforeDenoiseStep",
+        "WanAutoBlocks",
+        "WanAutoBlocks",
+        "WanAutoDecodeStep",
+        "WanAutoDenoiseStep",
+    ]
+    _import_structure["modular_pipeline"] = ["WanModularPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .encoders import WanTextEncoderStep
+        from .modular_blocks import (
+            ALL_BLOCKS,
+            AUTO_BLOCKS,
+            TEXT2VIDEO_BLOCKS,
+            WanAutoBeforeDenoiseStep,
+            WanAutoBlocks,
+            WanAutoDecodeStep,
+            WanAutoDenoiseStep,
+        )
+        from .modular_pipeline import WanModularPipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)
@@ -0,0 +1,365 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import List, Optional, Union
+
+import torch
+
+from ...schedulers import UniPCMultistepScheduler
+from ...utils import logging
+from ...utils.torch_utils import randn_tensor
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
+from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
+from .modular_pipeline import WanModularPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# TODO(yiyi, aryan): We need another step before text encoder to set the `num_inference_steps` attribute for guider so that
+# things like when to do guidance and how many conditions to be prepared can be determined. Currently, this is done by
+# always assuming you want to do guidance in the Guiders. So, negative embeddings are prepared regardless of what the
+# configuration of guider is.
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.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.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        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 sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_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 WanInputStep(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Input processing step that:\n"
+            "  1. Determines `batch_size` and `dtype` based on `prompt_embeds`\n"
+            "  2. Adjusts input tensor shapes based on `batch_size` (number of prompts) and `num_videos_per_prompt`\n\n"
+            "All input tensors are expected to have either batch_size=1 or match the batch_size\n"
+            "of prompt_embeds. The tensors will be duplicated across the batch dimension to\n"
+            "have a final batch_size of batch_size * num_videos_per_prompt."
+        )
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("num_videos_per_prompt", default=1),
+        ]
+
+    @property
+    def intermediate_inputs(self) -> List[str]:
+        return [
+            InputParam(
+                "prompt_embeds",
+                required=True,
+                type_hint=torch.Tensor,
+                description="Pre-generated text embeddings. Can be generated from text_encoder step.",
+            ),
+            InputParam(
+                "negative_prompt_embeds",
+                type_hint=torch.Tensor,
+                description="Pre-generated negative text embeddings. Can be generated from text_encoder step.",
+            ),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[str]:
+        return [
+            OutputParam(
+                "batch_size",
+                type_hint=int,
+                description="Number of prompts, the final batch size of model inputs should be batch_size * num_videos_per_prompt",
+            ),
+            OutputParam(
+                "dtype",
+                type_hint=torch.dtype,
+                description="Data type of model tensor inputs (determined by `prompt_embeds`)",
+            ),
+            OutputParam(
+                "prompt_embeds",
+                type_hint=torch.Tensor,
+                kwargs_type="guider_input_fields",  # already in intermedites state but declare here again for guider_input_fields
+                description="text embeddings used to guide the image generation",
+            ),
+            OutputParam(
+                "negative_prompt_embeds",
+                type_hint=torch.Tensor,
+                kwargs_type="guider_input_fields",  # already in intermedites state but declare here again for guider_input_fields
+                description="negative text embeddings used to guide the image generation",
+            ),
+        ]
+
+    def check_inputs(self, components, block_state):
+        if block_state.prompt_embeds is not None and block_state.negative_prompt_embeds is not None:
+            if block_state.prompt_embeds.shape != block_state.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` {block_state.prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {block_state.negative_prompt_embeds.shape}."
+                )
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+        self.check_inputs(components, block_state)
+
+        block_state.batch_size = block_state.prompt_embeds.shape[0]
+        block_state.dtype = block_state.prompt_embeds.dtype
+
+        _, seq_len, _ = block_state.prompt_embeds.shape
+        block_state.prompt_embeds = block_state.prompt_embeds.repeat(1, block_state.num_videos_per_prompt, 1)
+        block_state.prompt_embeds = block_state.prompt_embeds.view(
+            block_state.batch_size * block_state.num_videos_per_prompt, seq_len, -1
+        )
+
+        if block_state.negative_prompt_embeds is not None:
+            _, seq_len, _ = block_state.negative_prompt_embeds.shape
+            block_state.negative_prompt_embeds = block_state.negative_prompt_embeds.repeat(
+                1, block_state.num_videos_per_prompt, 1
+            )
+            block_state.negative_prompt_embeds = block_state.negative_prompt_embeds.view(
+                block_state.batch_size * block_state.num_videos_per_prompt, seq_len, -1
+            )
+
+        self.set_block_state(state, block_state)
+
+        return components, state
+
+
+class WanSetTimestepsStep(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("scheduler", UniPCMultistepScheduler),
+        ]
+
+    @property
+    def description(self) -> str:
+        return "Step that sets the scheduler's timesteps for inference"
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("num_inference_steps", default=50),
+            InputParam("timesteps"),
+            InputParam("sigmas"),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam("timesteps", type_hint=torch.Tensor, description="The timesteps to use for inference"),
+            OutputParam(
+                "num_inference_steps",
+                type_hint=int,
+                description="The number of denoising steps to perform at inference time",
+            ),
+        ]
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+        block_state.device = components._execution_device
+
+        block_state.timesteps, block_state.num_inference_steps = retrieve_timesteps(
+            components.scheduler,
+            block_state.num_inference_steps,
+            block_state.device,
+            block_state.timesteps,
+            block_state.sigmas,
+        )
+
+        self.set_block_state(state, block_state)
+        return components, state
+
+
+class WanPrepareLatentsStep(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return []
+
+    @property
+    def description(self) -> str:
+        return "Prepare latents step that prepares the latents for the text-to-video generation process"
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("height", type_hint=int),
+            InputParam("width", type_hint=int),
+            InputParam("num_frames", type_hint=int),
+            InputParam("latents", type_hint=Optional[torch.Tensor]),
+            InputParam("num_videos_per_prompt", type_hint=int, default=1),
+        ]
+
+    @property
+    def intermediate_inputs(self) -> List[InputParam]:
+        return [
+            InputParam("generator"),
+            InputParam(
+                "batch_size",
+                required=True,
+                type_hint=int,
+                description="Number of prompts, the final batch size of model inputs should be `batch_size * num_videos_per_prompt`. Can be generated in input step.",
+            ),
+            InputParam("dtype", type_hint=torch.dtype, description="The dtype of the model inputs"),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam(
+                "latents", type_hint=torch.Tensor, description="The initial latents to use for the denoising process"
+            )
+        ]
+
+    @staticmethod
+    def check_inputs(components, block_state):
+        if (block_state.height is not None and block_state.height % components.vae_scale_factor_spatial != 0) or (
+            block_state.width is not None and block_state.width % components.vae_scale_factor_spatial != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {components.vae_scale_factor_spatial} but are {block_state.height} and {block_state.width}."
+            )
+        if block_state.num_frames is not None and (
+            block_state.num_frames < 1 or (block_state.num_frames - 1) % components.vae_scale_factor_temporal != 0
+        ):
+            raise ValueError(
+                f"`num_frames` has to be greater than 0, and (num_frames - 1) must be divisible by {components.vae_scale_factor_temporal}, but got {block_state.num_frames}."
+            )
+
+    @staticmethod
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.prepare_latents with self->comp
+    def prepare_latents(
+        comp,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // comp.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // comp.vae_scale_factor_spatial,
+            int(width) // comp.vae_scale_factor_spatial,
+        )
+        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
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+
+        block_state.height = block_state.height or components.default_height
+        block_state.width = block_state.width or components.default_width
+        block_state.num_frames = block_state.num_frames or components.default_num_frames
+        block_state.device = components._execution_device
+        block_state.dtype = torch.float32  # Wan latents should be torch.float32 for best quality
+        block_state.num_channels_latents = components.num_channels_latents
+
+        self.check_inputs(components, block_state)
+
+        block_state.latents = self.prepare_latents(
+            components,
+            block_state.batch_size * block_state.num_videos_per_prompt,
+            block_state.num_channels_latents,
+            block_state.height,
+            block_state.width,
+            block_state.num_frames,
+            block_state.dtype,
+            block_state.device,
+            block_state.generator,
+            block_state.latents,
+        )
+
+        self.set_block_state(state, block_state)
+
+        return components, state
@@ -0,0 +1,105 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, List, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+
+from ...configuration_utils import FrozenDict
+from ...models import AutoencoderKLWan
+from ...utils import logging
+from ...video_processor import VideoProcessor
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
+from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanDecodeStep(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("vae", AutoencoderKLWan),
+            ComponentSpec(
+                "video_processor",
+                VideoProcessor,
+                config=FrozenDict({"vae_scale_factor": 8}),
+                default_creation_method="from_config",
+            ),
+        ]
+
+    @property
+    def description(self) -> str:
+        return "Step that decodes the denoised latents into images"
+
+    @property
+    def inputs(self) -> List[Tuple[str, Any]]:
+        return [
+            InputParam("output_type", default="pil"),
+        ]
+
+    @property
+    def intermediate_inputs(self) -> List[str]:
+        return [
+            InputParam(
+                "latents",
+                required=True,
+                type_hint=torch.Tensor,
+                description="The denoised latents from the denoising step",
+            )
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[str]:
+        return [
+            OutputParam(
+                "videos",
+                type_hint=Union[List[List[PIL.Image.Image]], List[torch.Tensor], List[np.ndarray]],
+                description="The generated videos, can be a PIL.Image.Image, torch.Tensor or a numpy array",
+            )
+        ]
+
+    @torch.no_grad()
+    def __call__(self, components, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+        vae_dtype = components.vae.dtype
+
+        if not block_state.output_type == "latent":
+            latents = block_state.latents
+            latents_mean = (
+                torch.tensor(components.vae.config.latents_mean)
+                .view(1, components.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(components.vae.config.latents_std).view(
+                1, components.vae.config.z_dim, 1, 1, 1
+            ).to(latents.device, latents.dtype)
+            latents = latents / latents_std + latents_mean
+            latents = latents.to(vae_dtype)
+            block_state.videos = components.vae.decode(latents, return_dict=False)[0]
+        else:
+            block_state.videos = block_state.latents
+
+        block_state.videos = components.video_processor.postprocess_video(
+            block_state.videos, output_type=block_state.output_type
+        )
+
+        self.set_block_state(state, block_state)
+
+        return components, state
@@ -0,0 +1,261 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, List, Tuple
+
+import torch
+
+from ...configuration_utils import FrozenDict
+from ...guiders import ClassifierFreeGuidance
+from ...models import WanTransformer3DModel
+from ...schedulers import UniPCMultistepScheduler
+from ...utils import logging
+from ..modular_pipeline import (
+    BlockState,
+    LoopSequentialPipelineBlocks,
+    ModularPipelineBlocks,
+    PipelineState,
+)
+from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
+from .modular_pipeline import WanModularPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanLoopDenoiser(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec(
+                "guider",
+                ClassifierFreeGuidance,
+                config=FrozenDict({"guidance_scale": 5.0}),
+                default_creation_method="from_config",
+            ),
+            ComponentSpec("transformer", WanTransformer3DModel),
+        ]
+
+    @property
+    def description(self) -> str:
+        return (
+            "Step within the denoising loop that denoise the latents with guidance. "
+            "This block should be used to compose the `sub_blocks` attribute of a `LoopSequentialPipelineBlocks` "
+            "object (e.g. `WanDenoiseLoopWrapper`)"
+        )
+
+    @property
+    def inputs(self) -> List[Tuple[str, Any]]:
+        return [
+            InputParam("attention_kwargs"),
+        ]
+
+    @property
+    def intermediate_inputs(self) -> List[str]:
+        return [
+            InputParam(
+                "latents",
+                required=True,
+                type_hint=torch.Tensor,
+                description="The initial latents to use for the denoising process. Can be generated in prepare_latent step.",
+            ),
+            InputParam(
+                "num_inference_steps",
+                required=True,
+                type_hint=int,
+                description="The number of inference steps to use for the denoising process. Can be generated in set_timesteps step.",
+            ),
+            InputParam(
+                kwargs_type="guider_input_fields",
+                description=(
+                    "All conditional model inputs that need to be prepared with guider. "
+                    "It should contain prompt_embeds/negative_prompt_embeds. "
+                    "Please add `kwargs_type=guider_input_fields` to their parameter spec (`OutputParam`) when they are created and added to the pipeline state"
+                ),
+            ),
+        ]
+
+    @torch.no_grad()
+    def __call__(
+        self, components: WanModularPipeline, block_state: BlockState, i: int, t: torch.Tensor
+    ) -> PipelineState:
+        #  Map the keys we'll see on each `guider_state_batch` (e.g. guider_state_batch.prompt_embeds)
+        #  to the corresponding (cond, uncond) fields on block_state. (e.g. block_state.prompt_embeds, block_state.negative_prompt_embeds)
+        guider_input_fields = {
+            "prompt_embeds": ("prompt_embeds", "negative_prompt_embeds"),
+        }
+        transformer_dtype = components.transformer.dtype
+
+        components.guider.set_state(step=i, num_inference_steps=block_state.num_inference_steps, timestep=t)
+
+        # Prepare mini‐batches according to guidance method and `guider_input_fields`
+        # Each guider_state_batch will have .prompt_embeds, .time_ids, text_embeds, image_embeds.
+        # e.g. for CFG, we prepare two batches: one for uncond, one for cond
+        # for first batch, guider_state_batch.prompt_embeds correspond to block_state.prompt_embeds
+        # for second batch, guider_state_batch.prompt_embeds correspond to block_state.negative_prompt_embeds
+        guider_state = components.guider.prepare_inputs(block_state, guider_input_fields)
+
+        # run the denoiser for each guidance batch
+        for guider_state_batch in guider_state:
+            components.guider.prepare_models(components.transformer)
+            cond_kwargs = guider_state_batch.as_dict()
+            cond_kwargs = {k: v for k, v in cond_kwargs.items() if k in guider_input_fields}
+            prompt_embeds = cond_kwargs.pop("prompt_embeds")
+
+            # Predict the noise residual
+            # store the noise_pred in guider_state_batch so that we can apply guidance across all batches
+            guider_state_batch.noise_pred = components.transformer(
+                hidden_states=block_state.latents.to(transformer_dtype),
+                timestep=t.flatten(),
+                encoder_hidden_states=prompt_embeds,
+                attention_kwargs=block_state.attention_kwargs,
+                return_dict=False,
+            )[0]
+            components.guider.cleanup_models(components.transformer)
+
+        # Perform guidance
+        block_state.noise_pred, block_state.scheduler_step_kwargs = components.guider(guider_state)
+
+        return components, block_state
+
+
+class WanLoopAfterDenoiser(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("scheduler", UniPCMultistepScheduler),
+        ]
+
+    @property
+    def description(self) -> str:
+        return (
+            "step within the denoising loop that update the latents. "
+            "This block should be used to compose the `sub_blocks` attribute of a `LoopSequentialPipelineBlocks` "
+            "object (e.g. `WanDenoiseLoopWrapper`)"
+        )
+
+    @property
+    def inputs(self) -> List[Tuple[str, Any]]:
+        return []
+
+    @property
+    def intermediate_inputs(self) -> List[str]:
+        return [
+            InputParam("generator"),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [OutputParam("latents", type_hint=torch.Tensor, description="The denoised latents")]
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, block_state: BlockState, i: int, t: torch.Tensor):
+        # Perform scheduler step using the predicted output
+        latents_dtype = block_state.latents.dtype
+        block_state.latents = components.scheduler.step(
+            block_state.noise_pred.float(),
+            t,
+            block_state.latents.float(),
+            **block_state.scheduler_step_kwargs,
+            return_dict=False,
+        )[0]
+
+        if block_state.latents.dtype != latents_dtype:
+            block_state.latents = block_state.latents.to(latents_dtype)
+
+        return components, block_state
+
+
+class WanDenoiseLoopWrapper(LoopSequentialPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Pipeline block that iteratively denoise the latents over `timesteps`. "
+            "The specific steps with each iteration can be customized with `sub_blocks` attributes"
+        )
+
+    @property
+    def loop_expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec(
+                "guider",
+                ClassifierFreeGuidance,
+                config=FrozenDict({"guidance_scale": 5.0}),
+                default_creation_method="from_config",
+            ),
+            ComponentSpec("scheduler", UniPCMultistepScheduler),
+            ComponentSpec("transformer", WanTransformer3DModel),
+        ]
+
+    @property
+    def loop_intermediate_inputs(self) -> List[InputParam]:
+        return [
+            InputParam(
+                "timesteps",
+                required=True,
+                type_hint=torch.Tensor,
+                description="The timesteps to use for the denoising process. Can be generated in set_timesteps step.",
+            ),
+            InputParam(
+                "num_inference_steps",
+                required=True,
+                type_hint=int,
+                description="The number of inference steps to use for the denoising process. Can be generated in set_timesteps step.",
+            ),
+        ]
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, state: PipelineState) -> PipelineState:
+        block_state = self.get_block_state(state)
+
+        block_state.num_warmup_steps = max(
+            len(block_state.timesteps) - block_state.num_inference_steps * components.scheduler.order, 0
+        )
+
+        with self.progress_bar(total=block_state.num_inference_steps) as progress_bar:
+            for i, t in enumerate(block_state.timesteps):
+                components, block_state = self.loop_step(components, block_state, i=i, t=t)
+                if i == len(block_state.timesteps) - 1 or (
+                    (i + 1) > block_state.num_warmup_steps and (i + 1) % components.scheduler.order == 0
+                ):
+                    progress_bar.update()
+
+        self.set_block_state(state, block_state)
+
+        return components, state
+
+
+class WanDenoiseStep(WanDenoiseLoopWrapper):
+    block_classes = [
+        WanLoopDenoiser,
+        WanLoopAfterDenoiser,
+    ]
+    block_names = ["before_denoiser", "denoiser", "after_denoiser"]
+
+    @property
+    def description(self) -> str:
+        return (
+            "Denoise step that iteratively denoise the latents. \n"
+            "Its loop logic is defined in `WanDenoiseLoopWrapper.__call__` method \n"
+            "At each iteration, it runs blocks defined in `sub_blocks` sequencially:\n"
+            " - `WanLoopDenoiser`\n"
+            " - `WanLoopAfterDenoiser`\n"
+            "This block supports both text2vid tasks."
+        )
@@ -0,0 +1,242 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import List, Optional, Union
+
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from ...configuration_utils import FrozenDict
+from ...guiders import ClassifierFreeGuidance
+from ...utils import is_ftfy_available, logging
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
+from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
+from .modular_pipeline import WanModularPipeline
+
+
+if is_ftfy_available():
+    import ftfy
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+class WanTextEncoderStep(ModularPipelineBlocks):
+    model_name = "wan"
+
+    @property
+    def description(self) -> str:
+        return "Text Encoder step that generate text_embeddings to guide the video generation"
+
+    @property
+    def expected_components(self) -> List[ComponentSpec]:
+        return [
+            ComponentSpec("text_encoder", UMT5EncoderModel),
+            ComponentSpec("tokenizer", AutoTokenizer),
+            ComponentSpec(
+                "guider",
+                ClassifierFreeGuidance,
+                config=FrozenDict({"guidance_scale": 5.0}),
+                default_creation_method="from_config",
+            ),
+        ]
+
+    @property
+    def expected_configs(self) -> List[ConfigSpec]:
+        return []
+
+    @property
+    def inputs(self) -> List[InputParam]:
+        return [
+            InputParam("prompt"),
+            InputParam("negative_prompt"),
+            InputParam("attention_kwargs"),
+        ]
+
+    @property
+    def intermediate_outputs(self) -> List[OutputParam]:
+        return [
+            OutputParam(
+                "prompt_embeds",
+                type_hint=torch.Tensor,
+                kwargs_type="guider_input_fields",
+                description="text embeddings used to guide the image generation",
+            ),
+            OutputParam(
+                "negative_prompt_embeds",
+                type_hint=torch.Tensor,
+                kwargs_type="guider_input_fields",
+                description="negative text embeddings used to guide the image generation",
+            ),
+        ]
+
+    @staticmethod
+    def check_inputs(block_state):
+        if block_state.prompt is not None and (
+            not isinstance(block_state.prompt, str) and not isinstance(block_state.prompt, list)
+        ):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(block_state.prompt)}")
+
+    @staticmethod
+    def _get_t5_prompt_embeds(
+        components,
+        prompt: Union[str, List[str]],
+        max_sequence_length: int,
+        device: torch.device,
+    ):
+        dtype = components.text_encoder.dtype
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_clean(u) for u in prompt]
+
+        text_inputs = components.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+        prompt_embeds = components.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        return prompt_embeds
+
+    @staticmethod
+    def encode_prompt(
+        components,
+        prompt: str,
+        device: Optional[torch.device] = None,
+        num_videos_per_prompt: int = 1,
+        prepare_unconditional_embeds: bool = True,
+        negative_prompt: Optional[str] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_videos_per_prompt (`int`):
+                number of videos that should be generated per prompt
+            prepare_unconditional_embeds (`bool`):
+                whether to use prepare unconditional embeddings or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.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.
+            max_sequence_length (`int`, defaults to `512`):
+                The maximum number of text tokens to be used for the generation process.
+        """
+        device = device or components._execution_device
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt) if prompt is not None else prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = WanTextEncoderStep._get_t5_prompt_embeds(components, prompt, max_sequence_length, device)
+
+        if prepare_unconditional_embeds 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 = WanTextEncoderStep._get_t5_prompt_embeds(
+                components, negative_prompt, max_sequence_length, device
+            )
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        if prepare_unconditional_embeds:
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    @torch.no_grad()
+    def __call__(self, components: WanModularPipeline, state: PipelineState) -> PipelineState:
+        # Get inputs and intermediates
+        block_state = self.get_block_state(state)
+        self.check_inputs(block_state)
+
+        block_state.prepare_unconditional_embeds = components.guider.num_conditions > 1
+        block_state.device = components._execution_device
+
+        # Encode input prompt
+        (
+            block_state.prompt_embeds,
+            block_state.negative_prompt_embeds,
+        ) = self.encode_prompt(
+            components,
+            block_state.prompt,
+            block_state.device,
+            1,
+            block_state.prepare_unconditional_embeds,
+            block_state.negative_prompt,
+            prompt_embeds=None,
+            negative_prompt_embeds=None,
+        )
+
+        # Add outputs
+        self.set_block_state(state, block_state)
+        return components, state
@@ -0,0 +1,144 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ...utils import logging
+from ..modular_pipeline import AutoPipelineBlocks, SequentialPipelineBlocks
+from ..modular_pipeline_utils import InsertableDict
+from .before_denoise import (
+    WanInputStep,
+    WanPrepareLatentsStep,
+    WanSetTimestepsStep,
+)
+from .decoders import WanDecodeStep
+from .denoise import WanDenoiseStep
+from .encoders import WanTextEncoderStep
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# before_denoise: text2vid
+class WanBeforeDenoiseStep(SequentialPipelineBlocks):
+    block_classes = [
+        WanInputStep,
+        WanSetTimestepsStep,
+        WanPrepareLatentsStep,
+    ]
+    block_names = ["input", "set_timesteps", "prepare_latents"]
+
+    @property
+    def description(self):
+        return (
+            "Before denoise step that prepare the inputs for the denoise step.\n"
+            + "This is a sequential pipeline blocks:\n"
+            + " - `WanInputStep` is used to adjust the batch size of the model inputs\n"
+            + " - `WanSetTimestepsStep` is used to set the timesteps\n"
+            + " - `WanPrepareLatentsStep` is used to prepare the latents\n"
+        )
+
+
+# before_denoise: all task (text2vid,)
+class WanAutoBeforeDenoiseStep(AutoPipelineBlocks):
+    block_classes = [
+        WanBeforeDenoiseStep,
+    ]
+    block_names = ["text2vid"]
+    block_trigger_inputs = [None]
+
+    @property
+    def description(self):
+        return (
+            "Before denoise step that prepare the inputs for the denoise step.\n"
+            + "This is an auto pipeline block that works for text2vid.\n"
+            + " - `WanBeforeDenoiseStep` (text2vid) is used.\n"
+        )
+
+
+# denoise: text2vid
+class WanAutoDenoiseStep(AutoPipelineBlocks):
+    block_classes = [
+        WanDenoiseStep,
+    ]
+    block_names = ["denoise"]
+    block_trigger_inputs = [None]
+
+    @property
+    def description(self) -> str:
+        return (
+            "Denoise step that iteratively denoise the latents. "
+            "This is a auto pipeline block that works for text2vid tasks.."
+            " - `WanDenoiseStep` (denoise) for text2vid tasks."
+        )
+
+
+# decode: all task (text2img, img2img, inpainting)
+class WanAutoDecodeStep(AutoPipelineBlocks):
+    block_classes = [WanDecodeStep]
+    block_names = ["non-inpaint"]
+    block_trigger_inputs = [None]
+
+    @property
+    def description(self):
+        return "Decode step that decode the denoised latents into videos outputs.\n - `WanDecodeStep`"
+
+
+# text2vid
+class WanAutoBlocks(SequentialPipelineBlocks):
+    block_classes = [
+        WanTextEncoderStep,
+        WanAutoBeforeDenoiseStep,
+        WanAutoDenoiseStep,
+        WanAutoDecodeStep,
+    ]
+    block_names = [
+        "text_encoder",
+        "before_denoise",
+        "denoise",
+        "decoder",
+    ]
+
+    @property
+    def description(self):
+        return (
+            "Auto Modular pipeline for text-to-video using Wan.\n"
+            + "- for text-to-video generation, all you need to provide is `prompt`"
+        )
+
+
+TEXT2VIDEO_BLOCKS = InsertableDict(
+    [
+        ("text_encoder", WanTextEncoderStep),
+        ("input", WanInputStep),
+        ("set_timesteps", WanSetTimestepsStep),
+        ("prepare_latents", WanPrepareLatentsStep),
+        ("denoise", WanDenoiseStep),
+        ("decode", WanDecodeStep),
+    ]
+)
+
+
+AUTO_BLOCKS = InsertableDict(
+    [
+        ("text_encoder", WanTextEncoderStep),
+        ("before_denoise", WanAutoBeforeDenoiseStep),
+        ("denoise", WanAutoDenoiseStep),
+        ("decode", WanAutoDecodeStep),
+    ]
+)
+
+
+ALL_BLOCKS = {
+    "text2video": TEXT2VIDEO_BLOCKS,
+    "auto": AUTO_BLOCKS,
+}
@@ -0,0 +1,90 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from ...loaders import WanLoraLoaderMixin
+from ...pipelines.pipeline_utils import StableDiffusionMixin
+from ...utils import logging
+from ..modular_pipeline import ModularPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class WanModularPipeline(
+    ModularPipeline,
+    StableDiffusionMixin,
+    WanLoraLoaderMixin,
+):
+    """
+    A ModularPipeline for Wan.
+
+    <Tip warning={true}>
+
+        This is an experimental feature and is likely to change in the future.
+
+    </Tip>
+    """
+
+    @property
+    def default_height(self):
+        return self.default_sample_height * self.vae_scale_factor_spatial
+
+    @property
+    def default_width(self):
+        return self.default_sample_width * self.vae_scale_factor_spatial
+
+    @property
+    def default_num_frames(self):
+        return (self.default_sample_num_frames - 1) * self.vae_scale_factor_temporal + 1
+
+    @property
+    def default_sample_height(self):
+        return 60
+
+    @property
+    def default_sample_width(self):
+        return 104
+
+    @property
+    def default_sample_num_frames(self):
+        return 21
+
+    @property
+    def vae_scale_factor_spatial(self):
+        vae_scale_factor = 8
+        if hasattr(self, "vae") and self.vae is not None:
+            vae_scale_factor = 2 ** len(self.vae.temperal_downsample)
+        return vae_scale_factor
+
+    @property
+    def vae_scale_factor_temporal(self):
+        vae_scale_factor = 4
+        if hasattr(self, "vae") and self.vae is not None:
+            vae_scale_factor = 2 ** sum(self.vae.temperal_downsample)
+        return vae_scale_factor
+
+    @property
+    def num_channels_transformer(self):
+        num_channels_transformer = 16
+        if hasattr(self, "transformer") and self.transformer is not None:
+            num_channels_transformer = self.transformer.config.in_channels
+        return num_channels_transformer
+
+    @property
+    def num_channels_latents(self):
+        num_channels_latents = 16
+        if hasattr(self, "vae") and self.vae is not None:
+            num_channels_latents = self.vae.config.z_dim
+        return num_channels_latents
@@ -380,6 +380,13 @@ else:
        "WuerstchenPriorPipeline",
    ]
    _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline", "WanVACEPipeline"]
+    _import_structure["skyreels_v2"] = [
+        "SkyReelsV2DiffusionForcingPipeline",
+        "SkyReelsV2DiffusionForcingImageToVideoPipeline",
+        "SkyReelsV2DiffusionForcingVideoToVideoPipeline",
+        "SkyReelsV2ImageToVideoPipeline",
+        "SkyReelsV2Pipeline",
+    ]
 try:
    if not is_onnx_available():
        raise OptionalDependencyNotAvailable()
@@ -851,6 +858,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                SpectrogramDiffusionPipeline,
            )

+        from .skyreels_v2 import (
+            SkyReelsV2DiffusionForcingImageToVideoPipeline,
+            SkyReelsV2DiffusionForcingPipeline,
+            SkyReelsV2DiffusionForcingVideoToVideoPipeline,
+            SkyReelsV2ImageToVideoPipeline,
+            SkyReelsV2Pipeline,
+        )
+
 else:
    import sys

@@ -663,11 +663,11 @@ class ChromaPipeline(
                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 3.5):
-                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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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*):
@@ -725,11 +725,11 @@ class ChromaImg2ImgPipeline(
                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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            strength (`float, *optional*, defaults to 0.9):
                Conceptually, indicates how much to transform the reference image. Must be between 0 and 1. image will
                be used as a starting point, adding more noise to it the larger the strength. The number of denoising
@@ -19,7 +19,6 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 import numpy as np
 import PIL.Image
 import torch
-import torch.nn.functional as F
 from transformers import (
    CLIPImageProcessor,
    CLIPTextModel,
@@ -38,7 +37,13 @@ from ...loaders import (
    StableDiffusionXLLoraLoaderMixin,
    TextualInversionLoaderMixin,
 )
-from ...models import AutoencoderKL, ControlNetModel, ControlNetUnionModel, ImageProjection, UNet2DConditionModel
+from ...models import (
+    AutoencoderKL,
+    ControlNetUnionModel,
+    ImageProjection,
+    MultiControlNetUnionModel,
+    UNet2DConditionModel,
+)
 from ...models.attention_processor import (
    AttnProcessor2_0,
    XFormersAttnProcessor,
@@ -262,7 +267,9 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
-        controlnet: ControlNetUnionModel,
+        controlnet: Union[
+            ControlNetUnionModel, List[ControlNetUnionModel], Tuple[ControlNetUnionModel], MultiControlNetUnionModel
+        ],
        scheduler: KarrasDiffusionSchedulers,
        requires_aesthetics_score: bool = False,
        force_zeros_for_empty_prompt: bool = True,
@@ -272,8 +279,8 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
    ):
        super().__init__()

-        if not isinstance(controlnet, ControlNetUnionModel):
-            raise ValueError("Expected `controlnet` to be of type `ControlNetUnionModel`.")
+        if isinstance(controlnet, (list, tuple)):
+            controlnet = MultiControlNetUnionModel(controlnet)

        self.register_modules(
            vae=vae,
@@ -649,6 +656,7 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        controlnet_conditioning_scale=1.0,
        control_guidance_start=0.0,
        control_guidance_end=1.0,
+        control_mode=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if strength < 0 or strength > 1:
@@ -722,28 +730,44 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )

+        # `prompt` needs more sophisticated handling when there are multiple
+        # conditionings.
+        if isinstance(self.controlnet, MultiControlNetUnionModel):
+            if isinstance(prompt, list):
+                logger.warning(
+                    f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+                    " prompts. The conditionings will be fixed across the prompts."
+                )
+
        # Check `image`
-        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
-            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
-        )
-        if (
-            isinstance(self.controlnet, ControlNetModel)
-            or is_compiled
-            and isinstance(self.controlnet._orig_mod, ControlNetModel)
-        ):
-            self.check_image(image, prompt, prompt_embeds)
-        elif (
-            isinstance(self.controlnet, ControlNetUnionModel)
-            or is_compiled
-            and isinstance(self.controlnet._orig_mod, ControlNetUnionModel)
-        ):
-            self.check_image(image, prompt, prompt_embeds)
-        else:
-            assert False
+        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+        if isinstance(controlnet, ControlNetUnionModel):
+            for image_ in image:
+                self.check_image(image_, prompt, prompt_embeds)
+        elif isinstance(controlnet, MultiControlNetUnionModel):
+            if not isinstance(image, list):
+                raise TypeError("For multiple controlnets: `image` must be type `list`")
+            elif not all(isinstance(i, list) for i in image):
+                raise ValueError("For multiple controlnets: elements of `image` must be list of conditionings.")
+            elif len(image) != len(self.controlnet.nets):
+                raise ValueError(
+                    f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+                )
+
+            for images_ in image:
+                for image_ in images_:
+                    self.check_image(image_, prompt, prompt_embeds)

        if not isinstance(control_guidance_start, (tuple, list)):
            control_guidance_start = [control_guidance_start]

+        if isinstance(controlnet, MultiControlNetUnionModel):
+            if len(control_guidance_start) != len(self.controlnet.nets):
+                raise ValueError(
+                    f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+                )
+
        if not isinstance(control_guidance_end, (tuple, list)):
            control_guidance_end = [control_guidance_end]

@@ -762,6 +786,15 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
            if end > 1.0:
                raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")

+        # Check `control_mode`
+        if isinstance(controlnet, ControlNetUnionModel):
+            if max(control_mode) >= controlnet.config.num_control_type:
+                raise ValueError(f"control_mode: must be lower than {controlnet.config.num_control_type}.")
+        elif isinstance(controlnet, MultiControlNetUnionModel):
+            for _control_mode, _controlnet in zip(control_mode, self.controlnet.nets):
+                if max(_control_mode) >= _controlnet.config.num_control_type:
+                    raise ValueError(f"control_mode: must be lower than {_controlnet.config.num_control_type}.")
+
        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
            raise ValueError(
                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
@@ -1049,7 +1082,7 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        image: PipelineImageInput = None,
-        control_image: PipelineImageInput = None,
+        control_image: Union[PipelineImageInput, List[PipelineImageInput]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        strength: float = 0.8,
@@ -1074,7 +1107,7 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        guess_mode: bool = False,
        control_guidance_start: Union[float, List[float]] = 0.0,
        control_guidance_end: Union[float, List[float]] = 1.0,
-        control_mode: Optional[Union[int, List[int]]] = None,
+        control_mode: Optional[Union[int, List[int], List[List[int]]]] = None,
        original_size: Tuple[int, int] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Tuple[int, int] = None,
@@ -1104,13 +1137,13 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
                    `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
                The initial image will be used as the starting point for the image generation process. Can also accept
                image latents as `image`, if passing latents directly, it will not be encoded again.
-            control_image (`PipelineImageInput`):
-                The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
-                the type is specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also
-                be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
-                and/or width are passed, `image` is resized according to them. If multiple ControlNets are specified in
-                init, images must be passed as a list such that each element of the list can be correctly batched for
-                input to a single controlnet.
+            control_image (`PipelineImageInput` or `List[PipelineImageInput]`, *optional*):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted
+                as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or
+                width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`,
+                images must be passed as a list such that each element of the list can be correctly batched for input
+                to a single ControlNet.
            height (`int`, *optional*, defaults to the size of control_image):
                The height in pixels of the generated image. Anything below 512 pixels won't work well for
                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
@@ -1184,16 +1217,21 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
-                The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
-                to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
-                corresponding scale as a list.
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
            guess_mode (`bool`, *optional*, defaults to `False`):
                In this mode, the ControlNet encoder will try best to recognize the content of the input image even if
                you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
-                The percentage of total steps at which the controlnet starts applying.
+                The percentage of total steps at which the ControlNet starts applying.
            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
-                The percentage of total steps at which the controlnet stops applying.
+                The percentage of total steps at which the ControlNet stops applying.
+            control_mode (`int` or `List[int]` or `List[List[int]], *optional*):
+                The control condition types for the ControlNet. See the ControlNet's model card forinformation on the
+                available control modes. If multiple ControlNets are specified in `init`, control_mode should be a list
+                where each ControlNet should have its corresponding control mode list. Should reflect the order of
+                conditions in control_image
            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
@@ -1273,12 +1311,6 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(

        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet

-        # align format for control guidance
-        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
-            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
-        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
-            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
-
        if not isinstance(control_image, list):
            control_image = [control_image]
        else:
@@ -1287,37 +1319,56 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        if not isinstance(control_mode, list):
            control_mode = [control_mode]

-        if len(control_image) != len(control_mode):
-            raise ValueError("Expected len(control_image) == len(control_type)")
+        if isinstance(controlnet, MultiControlNetUnionModel):
+            control_image = [[item] for item in control_image]
+            control_mode = [[item] for item in control_mode]

-        num_control_type = controlnet.config.num_control_type
-
-        # 1. Check inputs
-        control_type = [0 for _ in range(num_control_type)]
-        for _image, control_idx in zip(control_image, control_mode):
-            control_type[control_idx] = 1
-            self.check_inputs(
-                prompt,
-                prompt_2,
-                _image,
-                strength,
-                num_inference_steps,
-                callback_steps,
-                negative_prompt,
-                negative_prompt_2,
-                prompt_embeds,
-                negative_prompt_embeds,
-                pooled_prompt_embeds,
-                negative_pooled_prompt_embeds,
-                ip_adapter_image,
-                ip_adapter_image_embeds,
-                controlnet_conditioning_scale,
-                control_guidance_start,
-                control_guidance_end,
-                callback_on_step_end_tensor_inputs,
+        # align format for control guidance
+        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetUnionModel) else len(control_mode)
+            control_guidance_start, control_guidance_end = (
+                mult * [control_guidance_start],
+                mult * [control_guidance_end],
            )

-        control_type = torch.Tensor(control_type)
+        if isinstance(controlnet_conditioning_scale, float):
+            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetUnionModel) else len(control_mode)
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * mult
+
+        # 1. Check inputs
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            control_image,
+            strength,
+            num_inference_steps,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            controlnet_conditioning_scale,
+            control_guidance_start,
+            control_guidance_end,
+            control_mode,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        if isinstance(controlnet, ControlNetUnionModel):
+            control_type = torch.zeros(controlnet.config.num_control_type).scatter_(0, torch.tensor(control_mode), 1)
+        elif isinstance(controlnet, MultiControlNetUnionModel):
+            control_type = [
+                torch.zeros(controlnet_.config.num_control_type).scatter_(0, torch.tensor(control_mode_), 1)
+                for control_mode_, controlnet_ in zip(control_mode, self.controlnet.nets)
+            ]

        self._guidance_scale = guidance_scale
        self._clip_skip = clip_skip
@@ -1334,7 +1385,11 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(

        device = self._execution_device

-        global_pool_conditions = controlnet.config.global_pool_conditions
+        global_pool_conditions = (
+            controlnet.config.global_pool_conditions
+            if isinstance(controlnet, ControlNetUnionModel)
+            else controlnet.nets[0].config.global_pool_conditions
+        )
        guess_mode = guess_mode or global_pool_conditions

        # 3.1. Encode input prompt
@@ -1372,22 +1427,55 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
                self.do_classifier_free_guidance,
            )

-        # 4. Prepare image and controlnet_conditioning_image
+        # 4.1 Prepare image
        image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)

-        for idx, _ in enumerate(control_image):
-            control_image[idx] = self.prepare_control_image(
-                image=control_image[idx],
-                width=width,
-                height=height,
-                batch_size=batch_size * num_images_per_prompt,
-                num_images_per_prompt=num_images_per_prompt,
-                device=device,
-                dtype=controlnet.dtype,
-                do_classifier_free_guidance=self.do_classifier_free_guidance,
-                guess_mode=guess_mode,
-            )
-            height, width = control_image[idx].shape[-2:]
+        # 4.2 Prepare control images
+        if isinstance(controlnet, ControlNetUnionModel):
+            control_images = []
+
+            for image_ in control_image:
+                image_ = self.prepare_control_image(
+                    image=image_,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=controlnet.dtype,
+                    do_classifier_free_guidance=self.do_classifier_free_guidance,
+                    guess_mode=guess_mode,
+                )
+
+                control_images.append(image_)
+
+            control_image = control_images
+            height, width = control_image[0].shape[-2:]
+
+        elif isinstance(controlnet, MultiControlNetUnionModel):
+            control_images = []
+
+            for control_image_ in control_image:
+                images = []
+
+                for image_ in control_image_:
+                    image_ = self.prepare_control_image(
+                        image=image_,
+                        width=width,
+                        height=height,
+                        batch_size=batch_size * num_images_per_prompt,
+                        num_images_per_prompt=num_images_per_prompt,
+                        device=device,
+                        dtype=controlnet.dtype,
+                        do_classifier_free_guidance=self.do_classifier_free_guidance,
+                        guess_mode=guess_mode,
+                    )
+
+                    images.append(image_)
+                control_images.append(images)
+
+            control_image = control_images
+            height, width = control_image[0][0].shape[-2:]

        # 5. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
@@ -1414,10 +1502,11 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        # 7.1 Create tensor stating which controlnets to keep
        controlnet_keep = []
        for i in range(len(timesteps)):
-            controlnet_keep.append(
-                1.0
-                - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end)
-            )
+            keeps = [
+                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+                for s, e in zip(control_guidance_start, control_guidance_end)
+            ]
+            controlnet_keep.append(keeps)

        # 7.2 Prepare added time ids & embeddings
        original_size = original_size or (height, width)
@@ -1460,12 +1549,25 @@ class StableDiffusionXLControlNetUnionImg2ImgPipeline(
        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device)
-        control_type = (
-            control_type.reshape(1, -1)
-            .to(device, dtype=prompt_embeds.dtype)
-            .repeat(batch_size * num_images_per_prompt * 2, 1)
+
+        control_type_repeat_factor = (
+            batch_size * num_images_per_prompt * (2 if self.do_classifier_free_guidance else 1)
        )

+        if isinstance(controlnet, ControlNetUnionModel):
+            control_type = (
+                control_type.reshape(1, -1)
+                .to(self._execution_device, dtype=prompt_embeds.dtype)
+                .repeat(control_type_repeat_factor, 1)
+            )
+        elif isinstance(controlnet, MultiControlNetUnionModel):
+            control_type = [
+                _control_type.reshape(1, -1)
+                .to(self._execution_device, dtype=prompt_embeds.dtype)
+                .repeat(control_type_repeat_factor, 1)
+                for _control_type in control_type
+            ]
+
        # 8. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=num_inference_steps) as progress_bar:
@@ -46,7 +46,9 @@ class DiTPipeline(DiffusionPipeline):

    Parameters:
        transformer ([`DiTTransformer2DModel`]):
-            A class conditioned `DiTTransformer2DModel` to denoise the encoded image latents.
+            A class conditioned `DiTTransformer2DModel` to denoise the encoded image latents. Initially published as
+            [`Transformer2DModel`](https://huggingface.co/facebook/DiT-XL-2-256/blob/main/transformer/config.json#L2)
+            in the config, but the mismatch can be ignored.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
        scheduler ([`DDIMScheduler`]):
@@ -674,7 +674,8 @@ class FluxPipeline(
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
            true_cfg_scale (`float`, *optional*, defaults to 1.0):
-                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+                True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and
+                `negative_prompt` is provided.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
@@ -687,11 +688,11 @@ class FluxPipeline(
                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 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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*):
@@ -840,6 +841,8 @@ class FluxPipeline(

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
+        if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:
+            sigmas = None
        image_seq_len = latents.shape[1]
        mu = calculate_shift(
            image_seq_len,
@@ -661,11 +661,11 @@ class FluxControlPipeline(
                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 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with prompt at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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*):
@@ -795,11 +795,11 @@ class FluxKontextPipeline(
                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 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with prompt at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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*):
@@ -989,7 +989,8 @@ class FluxKontextInpaintPipeline(
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
            true_cfg_scale (`float`, *optional*, defaults to 1.0):
-                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+                True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and
+                `negative_prompt` is provided.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
@@ -1015,11 +1016,11 @@ class FluxKontextInpaintPipeline(
                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 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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*):
@@ -763,11 +763,11 @@ class HiDreamImagePipeline(DiffusionPipeline, HiDreamImageLoraLoaderMixin):
                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 3.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            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 `true_cfg_scale` is
@@ -529,15 +529,14 @@ class HunyuanVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            true_cfg_scale (`float`, *optional*, defaults to 1.0):
-                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+                True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and
+                `negative_prompt` is provided.
            guidance_scale (`float`, defaults to `6.0`):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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. Note that the only available
-                HunyuanVideo model is CFG-distilled, which means that traditional guidance between unconditional and
-                conditional latent is not applied.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            num_videos_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
@@ -256,7 +256,9 @@ class PixArtAlphaPipeline(DiffusionPipeline):
            Tokenizer of class
            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
        transformer ([`PixArtTransformer2DModel`]):
-            A text conditioned `PixArtTransformer2DModel` to denoise the encoded image latents.
+            A text conditioned `PixArtTransformer2DModel` to denoise the encoded image latents. Initially published as
+            [`Transformer2DModel`](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/blob/main/transformer/config.json#L2)
+            in the config, but the mismatch can be ignored.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
    """
@@ -185,6 +185,26 @@ def retrieve_timesteps(
 class PixArtSigmaPipeline(DiffusionPipeline):
    r"""
    Pipeline for text-to-image generation using PixArt-Sigma.
+
+    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 ([`T5EncoderModel`]):
+            Frozen text-encoder. PixArt-Alpha uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`PixArtTransformer2DModel`]):
+            A text conditioned `PixArtTransformer2DModel` to denoise the encoded image latents. Initially published as
+            [`Transformer2DModel`](https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/blob/main/transformer/config.json#L2)
+            in the config, but the mismatch can be ignored.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
    """

    bad_punct_regex = re.compile(
@@ -643,11 +643,11 @@ class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
                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.
            guidance_scale (`float`, *optional*, defaults to 4.5):
-                Guidance scale as defined in [Classifier-Free Diffusion
-                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
-                of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.
+                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
+                a model to generate images more aligned with `prompt` at the expense of lower image quality.
+
+                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
+                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            height (`int`, *optional*, defaults to self.unet.config.sample_size):
@@ -0,0 +1,59 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_skyreels_v2"] = ["SkyReelsV2Pipeline"]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing"] = ["SkyReelsV2DiffusionForcingPipeline"]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing_i2v"] = [
+        "SkyReelsV2DiffusionForcingImageToVideoPipeline"
+    ]
+    _import_structure["pipeline_skyreels_v2_diffusion_forcing_v2v"] = [
+        "SkyReelsV2DiffusionForcingVideoToVideoPipeline"
+    ]
+    _import_structure["pipeline_skyreels_v2_i2v"] = ["SkyReelsV2ImageToVideoPipeline"]
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_skyreels_v2 import SkyReelsV2Pipeline
+        from .pipeline_skyreels_v2_diffusion_forcing import SkyReelsV2DiffusionForcingPipeline
+        from .pipeline_skyreels_v2_diffusion_forcing_i2v import SkyReelsV2DiffusionForcingImageToVideoPipeline
+        from .pipeline_skyreels_v2_diffusion_forcing_v2v import SkyReelsV2DiffusionForcingVideoToVideoPipeline
+        from .pipeline_skyreels_v2_i2v import SkyReelsV2ImageToVideoPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)
@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+import torch
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class SkyReelsV2PipelineOutput(BaseOutput):
+    r"""
+    Output class for SkyReelsV2 pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    frames: torch.Tensor
@@ -0,0 +1,611 @@
+# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import regex as re
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...loaders import SkyReelsV2LoraLoaderMixin
+from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel
+from ...schedulers import UniPCMultistepScheduler
+from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import SkyReelsV2PipelineOutput
+
+
+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
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """\
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import (
+        ...     SkyReelsV2Pipeline,
+        ...     UniPCMultistepScheduler,
+        ...     AutoencoderKLWan,
+        ... )
+        >>> from diffusers.utils import export_to_video
+
+        >>> # Load the pipeline
+        >>> # Available models:
+        >>> # - Skywork/SkyReels-V2-T2V-14B-540P-Diffusers
+        >>> # - Skywork/SkyReels-V2-T2V-14B-720P-Diffusers
+        >>> vae = AutoencoderKLWan.from_pretrained(
+        ...     "Skywork/SkyReels-V2-T2V-14B-720P-Diffusers",
+        ...     subfolder="vae",
+        ...     torch_dtype=torch.float32,
+        ... )
+        >>> pipe = SkyReelsV2Pipeline.from_pretrained(
+        ...     "Skywork/SkyReels-V2-T2V-14B-720P-Diffusers",
+        ...     vae=vae,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> flow_shift = 8.0  # 8.0 for T2V, 5.0 for I2V
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     num_inference_steps=50,
+        ...     height=544,
+        ...     width=960,
+        ...     guidance_scale=6.0,  # 6.0 for T2V, 5.0 for I2V
+        ...     num_frames=97,
+        ... ).frames[0]
+        >>> export_to_video(output, "video.mp4", fps=24, quality=8)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+class SkyReelsV2Pipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin):
+    r"""
+    Pipeline for Text-to-Video (t2v) generation using SkyReels-V2.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`SkyReelsV2Transformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: SkyReelsV2Transformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: UniPCMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        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
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.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_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos 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
+        """
+        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_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        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_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=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 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`: {negative_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)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        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
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 544,
+        width: int = 960,
+        num_frames: int = 97,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        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 = 512,
+    ):
+        r"""
+        The call function to 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`.
+                instead.
+            height (`int`, defaults to `544`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `960`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `97`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `6.0`):
+                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_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`SkyReelsV2PipelineOutput`] 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`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. 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`, *optional*, defaults to `512`):
+                The maximum sequence length for the text encoder.
+
+        Examples:
+
+        Returns:
+            [`~SkyReelsV2PipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define 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]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        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 = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                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, t, 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)
+
+                # call the callback, if provided
+                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)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return SkyReelsV2PipelineOutput(frames=video)
@@ -0,0 +1,978 @@
+# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import math
+import re
+from copy import deepcopy
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import ftfy
+import torch
+from transformers import AutoTokenizer, UMT5EncoderModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...loaders import SkyReelsV2LoraLoaderMixin
+from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel
+from ...schedulers import UniPCMultistepScheduler
+from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import SkyReelsV2PipelineOutput
+
+
+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
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """\
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import (
+        ...     SkyReelsV2DiffusionForcingPipeline,
+        ...     UniPCMultistepScheduler,
+        ...     AutoencoderKLWan,
+        ... )
+        >>> from diffusers.utils import export_to_video
+
+        >>> # Load the pipeline
+        >>> # Available models:
+        >>> # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers
+        >>> # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers
+        >>> # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers
+        >>> vae = AutoencoderKLWan.from_pretrained(
+        ...     "Skywork/SkyReels-V2-DF-14B-720P-Diffusers",
+        ...     subfolder="vae",
+        ...     torch_dtype=torch.float32,
+        ... )
+        >>> pipe = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
+        ...     "Skywork/SkyReels-V2-DF-14B-720P-Diffusers",
+        ...     vae=vae,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> flow_shift = 8.0  # 8.0 for T2V, 5.0 for I2V
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+
+        >>> output = pipe(
+        ...     prompt=prompt,
+        ...     num_inference_steps=30,
+        ...     height=544,
+        ...     width=960,
+        ...     guidance_scale=6.0,  # 6.0 for T2V, 5.0 for I2V
+        ...     num_frames=97,
+        ...     ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
+        ...     causal_block_size=5,  # Number of frames processed together in a causal block
+        ...     overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos
+        ...     addnoise_condition=20,  # Improves consistency in long video generation
+        ... ).frames[0]
+        >>> export_to_video(output, "video.mp4", fps=24, quality=8)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class SkyReelsV2DiffusionForcingPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin):
+    """
+    Pipeline for Text-to-Video (t2v) generation using SkyReels-V2 with diffusion forcing.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a specific device, etc.).
+
+    Args:
+        tokenizer ([`AutoTokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`UMT5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        transformer ([`SkyReelsV2Transformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        transformer: SkyReelsV2Transformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: UniPCMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        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
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.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_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos 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
+        """
+        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_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        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_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        overlap_history=None,
+        num_frames=None,
+        base_num_frames=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 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`: {negative_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)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+        if num_frames > base_num_frames and overlap_history is None:
+            raise ValueError(
+                "`overlap_history` is required when `num_frames` exceeds `base_num_frames` to ensure smooth transitions in long video generation. "
+                "Please specify a value for `overlap_history`. Recommended values are 17 or 37."
+            )
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 97,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        base_latent_num_frames: Optional[int] = None,
+        video_latents: Optional[torch.Tensor] = None,
+        causal_block_size: Optional[int] = None,
+        overlap_history_latent_frames: Optional[int] = None,
+        long_video_iter: Optional[int] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height = height // self.vae_scale_factor_spatial
+        latent_width = width // self.vae_scale_factor_spatial
+
+        prefix_video_latents = None
+        prefix_video_latents_frames = 0
+
+        if video_latents is not None:  # long video generation at the iterations other than the first one
+            prefix_video_latents = video_latents[:, :, -overlap_history_latent_frames:]
+
+            if prefix_video_latents.shape[2] % causal_block_size != 0:
+                truncate_len_latents = prefix_video_latents.shape[2] % causal_block_size
+                logger.warning(
+                    f"The length of prefix video latents is truncated by {truncate_len_latents} frames for the causal block size alignment. "
+                    f"This truncation ensures compatibility with the causal block size, which is required for proper processing. "
+                    f"However, it may slightly affect the continuity of the generated video at the truncation boundary."
+                )
+                prefix_video_latents = prefix_video_latents[:, :, :-truncate_len_latents]
+            prefix_video_latents_frames = prefix_video_latents.shape[2]
+
+            finished_frame_num = (
+                long_video_iter * (base_latent_num_frames - overlap_history_latent_frames)
+                + overlap_history_latent_frames
+            )
+            left_frame_num = num_latent_frames - finished_frame_num
+            num_latent_frames = min(left_frame_num + overlap_history_latent_frames, base_latent_num_frames)
+        elif base_latent_num_frames is not None:  # long video generation at the first iteration
+            num_latent_frames = base_latent_num_frames
+        else:  # short video generation
+            num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_latent_frames,
+            latent_height,
+            latent_width,
+        )
+        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, num_latent_frames, prefix_video_latents, prefix_video_latents_frames
+
+    def generate_timestep_matrix(
+        self,
+        num_latent_frames: int,
+        step_template: torch.Tensor,
+        base_num_latent_frames: int,
+        ar_step: int = 5,
+        num_pre_ready: int = 0,
+        causal_block_size: int = 1,
+        shrink_interval_with_mask: bool = False,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[tuple]]:
+        """
+        This function implements the core diffusion forcing algorithm that creates a coordinated denoising schedule
+        across temporal frames. It supports both synchronous and asynchronous generation modes:
+
+        **Synchronous Mode** (ar_step=0, causal_block_size=1):
+        - All frames are denoised simultaneously at each timestep
+        - Each frame follows the same denoising trajectory: [1000, 800, 600, ..., 0]
+        - Simpler but may have less temporal consistency for long videos
+
+        **Asynchronous Mode** (ar_step>0, causal_block_size>1):
+        - Frames are grouped into causal blocks and processed block/chunk-wise
+        - Each block is denoised in a staggered pattern creating a "denoising wave"
+        - Earlier blocks are more denoised, later blocks lag behind by ar_step timesteps
+        - Creates stronger temporal dependencies and better consistency
+
+        Args:
+            num_latent_frames (int): Total number of latent frames to generate
+            step_template (torch.Tensor): Base timestep schedule (e.g., [1000, 800, 600, ..., 0])
+            base_num_latent_frames (int): Maximum frames the model can process in one forward pass
+            ar_step (int, optional): Autoregressive step size for temporal lag.
+                                   0 = synchronous, >0 = asynchronous. Defaults to 5.
+            num_pre_ready (int, optional):
+                                         Number of frames already denoised (e.g., from prefix in a video2video task).
+                                         Defaults to 0.
+            causal_block_size (int, optional): Number of frames processed as a causal block.
+                                             Defaults to 1.
+            shrink_interval_with_mask (bool, optional): Whether to optimize processing intervals.
+                                                      Defaults to False.
+
+        Returns:
+            tuple containing:
+                - step_matrix (torch.Tensor): Matrix of timesteps for each frame at each iteration Shape:
+                  [num_iterations, num_latent_frames]
+                - step_index (torch.Tensor): Index matrix for timestep lookup Shape: [num_iterations,
+                  num_latent_frames]
+                - step_update_mask (torch.Tensor): Boolean mask indicating which frames to update Shape:
+                  [num_iterations, num_latent_frames]
+                - valid_interval (list[tuple]): List of (start, end) intervals for each iteration
+
+        Raises:
+            ValueError: If ar_step is too small for the given configuration
+        """
+        # Initialize lists to store the scheduling matrices and metadata
+        step_matrix, step_index = [], []  # Will store timestep values and indices for each iteration
+        update_mask, valid_interval = [], []  # Will store update masks and processing intervals
+
+        # Calculate total number of denoising iterations (add 1 for initial noise state)
+        num_iterations = len(step_template) + 1
+
+        # Convert frame counts to block counts for causal processing
+        # Each block contains causal_block_size frames that are processed together
+        # E.g.: 25 frames ÷ 5 = 5 blocks total
+        num_blocks = num_latent_frames // causal_block_size
+        base_num_blocks = base_num_latent_frames // causal_block_size
+
+        # Validate ar_step is sufficient for the given configuration
+        # In asynchronous mode, we need enough timesteps to create the staggered pattern
+        if base_num_blocks < num_blocks:
+            min_ar_step = len(step_template) / base_num_blocks
+            if ar_step < min_ar_step:
+                raise ValueError(f"`ar_step` should be at least {math.ceil(min_ar_step)} in your setting")
+
+        # Extend step_template with boundary values for easier indexing
+        # 999: dummy value for counter starting from 1
+        # 0: final timestep (completely denoised)
+        step_template = torch.cat(
+            [
+                torch.tensor([999], dtype=torch.int64, device=step_template.device),
+                step_template.long(),
+                torch.tensor([0], dtype=torch.int64, device=step_template.device),
+            ]
+        )
+
+        # Initialize the previous row state (tracks denoising progress for each block)
+        # 0 means not started, num_iterations means fully denoised
+        pre_row = torch.zeros(num_blocks, dtype=torch.long)
+
+        # Mark pre-ready frames (e.g., from prefix video for a video2video task) as already at final denoising state
+        if num_pre_ready > 0:
+            pre_row[: num_pre_ready // causal_block_size] = num_iterations
+
+        # Main loop: Generate denoising schedule until all frames are fully denoised
+        while not torch.all(pre_row >= (num_iterations - 1)):
+            # Create new row representing the next denoising step
+            new_row = torch.zeros(num_blocks, dtype=torch.long)
+
+            # Apply diffusion forcing logic for each block
+            for i in range(num_blocks):
+                if i == 0 or pre_row[i - 1] >= (
+                    num_iterations - 1
+                ):  # the first frame or the last frame is completely denoised
+                    new_row[i] = pre_row[i] + 1
+                else:
+                    # Asynchronous mode: lag behind previous block by ar_step timesteps
+                    # This creates the "diffusion forcing" staggered pattern
+                    new_row[i] = new_row[i - 1] - ar_step
+
+            # Clamp values to valid range [0, num_iterations]
+            new_row = new_row.clamp(0, num_iterations)
+
+            # Create update mask: True for blocks that need denoising update at this iteration
+            # Exclude blocks that haven't started (new_row != pre_row) or are finished (new_row != num_iterations)
+            # Final state example: [False, ..., False, True, True, True, True, True]
+            # where first 20 frames are done (False) and last 5 frames still need updates (True)
+            update_mask.append((new_row != pre_row) & (new_row != num_iterations))
+
+            # Store the iteration state
+            step_index.append(new_row)  # Index into step_template
+            step_matrix.append(step_template[new_row])  # Actual timestep values
+            pre_row = new_row  # Update for next iteration
+
+        # For videos longer than model capacity, we process in sliding windows
+        terminal_flag = base_num_blocks
+
+        # Optional optimization: shrink interval based on first update mask
+        if shrink_interval_with_mask:
+            idx_sequence = torch.arange(num_blocks, dtype=torch.int64)
+            update_mask = update_mask[0]
+            update_mask_idx = idx_sequence[update_mask]
+            last_update_idx = update_mask_idx[-1].item()
+            terminal_flag = last_update_idx + 1
+
+        # Each interval defines which frames to process in the current forward pass
+        for curr_mask in update_mask:
+            # Extend terminal flag if current mask has updates beyond current terminal
+            if terminal_flag < num_blocks and curr_mask[terminal_flag]:
+                terminal_flag += 1
+            # Create interval: [start, end) where start ensures we don't exceed model capacity
+            valid_interval.append((max(terminal_flag - base_num_blocks, 0), terminal_flag))
+
+        # Convert lists to tensors for efficient processing
+        step_update_mask = torch.stack(update_mask, dim=0)
+        step_index = torch.stack(step_index, dim=0)
+        step_matrix = torch.stack(step_matrix, dim=0)
+
+        # Each block's schedule is replicated to all frames within that block
+        if causal_block_size > 1:
+            # Expand each block to causal_block_size frames
+            step_update_mask = step_update_mask.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous()
+            step_index = step_index.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous()
+            step_matrix = step_matrix.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous()
+            # Scale intervals from block-level to frame-level
+            valid_interval = [(s * causal_block_size, e * causal_block_size) for s, e in valid_interval]
+
+        return step_matrix, step_index, step_update_mask, valid_interval
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 544,
+        width: int = 960,
+        num_frames: int = 97,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        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 = 512,
+        overlap_history: Optional[int] = None,
+        addnoise_condition: float = 0,
+        base_num_frames: int = 97,
+        ar_step: int = 0,
+        causal_block_size: Optional[int] = None,
+        fps: int = 24,
+    ):
+        r"""
+        The call function to 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`.
+                instead.
+            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`, defaults to `544`):
+                The height of the generated video.
+            width (`int`, defaults to `960`):
+                The width of the generated video.
+            num_frames (`int`, defaults to `97`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `6.0`):
+                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. (**6.0 for T2V**, **5.0 for I2V**)
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`SkyReelsV2PipelineOutput`] 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`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. 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`, *optional*, defaults to `512`):
+                The maximum sequence length of the prompt.
+            overlap_history (`int`, *optional*, defaults to `None`):
+                Number of frames to overlap for smooth transitions in long videos. If `None`, the pipeline assumes
+                short video generation mode, and no overlap is applied. 17 and 37 are recommended to set.
+            addnoise_condition (`float`, *optional*, defaults to `0`):
+                This is used to help smooth the long video generation by adding some noise to the clean condition. Too
+                large noise can cause the inconsistency as well. 20 is a recommended value, and you may try larger
+                ones, but it is recommended to not exceed 50.
+            base_num_frames (`int`, *optional*, defaults to `97`):
+                97 or 121 | Base frame count (**97 for 540P**, **121 for 720P**)
+            ar_step (`int`, *optional*, defaults to `0`):
+                Controls asynchronous inference (0 for synchronous mode) You can set `ar_step=5` to enable asynchronous
+                inference. When asynchronous inference, `causal_block_size=5` is recommended while it is not supposed
+                to be set for synchronous generation. Asynchronous inference will take more steps to diffuse the whole
+                sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous
+                inference may improve the instruction following and visual consistent performance.
+            causal_block_size (`int`, *optional*, defaults to `None`):
+                The number of frames in each block/chunk. Recommended when using asynchronous inference (when ar_step >
+                0)
+            fps (`int`, *optional*, defaults to `24`):
+                Frame rate of the generated video
+
+        Examples:
+
+        Returns:
+            [`~SkyReelsV2PipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            overlap_history,
+            num_frames,
+            base_num_frames,
+        )
+
+        if addnoise_condition > 60:
+            logger.warning(
+                f"The value of 'addnoise_condition' is too large ({addnoise_condition}) and may cause inconsistencies in long video generation. A value of 20 is recommended."
+            )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define 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]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        if causal_block_size is None:
+            causal_block_size = self.transformer.config.num_frame_per_block
+        else:
+            self.transformer._set_ar_attention(causal_block_size)
+
+        fps_embeds = [fps] * prompt_embeds.shape[0]
+        fps_embeds = [0 if i == 16 else 1 for i in fps_embeds]
+
+        # Determine if we're doing long video generation
+        is_long_video = overlap_history is not None and base_num_frames is not None and num_frames > base_num_frames
+        # Initialize accumulated_latents to store all latents in one tensor
+        accumulated_latents = None
+        if is_long_video:
+            # Long video generation setup
+            overlap_history_latent_frames = (overlap_history - 1) // self.vae_scale_factor_temporal + 1
+            num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+            base_latent_num_frames = (
+                (base_num_frames - 1) // self.vae_scale_factor_temporal + 1
+                if base_num_frames is not None
+                else num_latent_frames
+            )
+            n_iter = (
+                1
+                + (num_latent_frames - base_latent_num_frames - 1)
+                // (base_latent_num_frames - overlap_history_latent_frames)
+                + 1
+            )
+        else:
+            # Short video generation setup
+            n_iter = 1
+            base_latent_num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+
+        # Loop through iterations (multiple iterations only for long videos)
+        for iter_idx in range(n_iter):
+            if is_long_video:
+                logger.debug(f"Processing iteration {iter_idx + 1}/{n_iter} for long video generation...")
+
+            # 5. Prepare latent variables
+            num_channels_latents = self.transformer.config.in_channels
+            latents, current_num_latent_frames, prefix_video_latents, prefix_video_latents_frames = (
+                self.prepare_latents(
+                    batch_size * num_videos_per_prompt,
+                    num_channels_latents,
+                    height,
+                    width,
+                    num_frames,
+                    torch.float32,
+                    device,
+                    generator,
+                    latents if iter_idx == 0 else None,
+                    video_latents=accumulated_latents,  # Pass latents directly instead of decoded video
+                    base_latent_num_frames=base_latent_num_frames if is_long_video else None,
+                    causal_block_size=causal_block_size,
+                    overlap_history_latent_frames=overlap_history_latent_frames if is_long_video else None,
+                    long_video_iter=iter_idx if is_long_video else None,
+                )
+            )
+
+            if prefix_video_latents_frames > 0:
+                latents[:, :, :prefix_video_latents_frames, :, :] = prefix_video_latents.to(transformer_dtype)
+
+            # 6. Prepare sample schedulers and timestep matrix
+            sample_schedulers = []
+            for _ in range(current_num_latent_frames):
+                sample_scheduler = deepcopy(self.scheduler)
+                sample_scheduler.set_timesteps(num_inference_steps, device=device)
+                sample_schedulers.append(sample_scheduler)
+
+            # Different matrix generation for short vs long video
+            step_matrix, _, step_update_mask, valid_interval = self.generate_timestep_matrix(
+                current_num_latent_frames,
+                timesteps,
+                current_num_latent_frames if is_long_video else base_latent_num_frames,
+                ar_step,
+                prefix_video_latents_frames,
+                causal_block_size,
+            )
+
+            # 7. Denoising loop
+            num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+            self._num_timesteps = len(step_matrix)
+
+            with self.progress_bar(total=len(step_matrix)) as progress_bar:
+                for i, t in enumerate(step_matrix):
+                    if self.interrupt:
+                        continue
+
+                    self._current_timestep = t
+                    valid_interval_start, valid_interval_end = valid_interval[i]
+                    latent_model_input = (
+                        latents[:, :, valid_interval_start:valid_interval_end, :, :].to(transformer_dtype).clone()
+                    )
+                    timestep = t.expand(latents.shape[0], -1)[:, valid_interval_start:valid_interval_end].clone()
+
+                    if addnoise_condition > 0 and valid_interval_start < prefix_video_latents_frames:
+                        noise_factor = 0.001 * addnoise_condition
+                        latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] = (
+                            latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :]
+                            * (1.0 - noise_factor)
+                            + torch.randn_like(
+                                latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :]
+                            )
+                            * noise_factor
+                        )
+                        timestep[:, valid_interval_start:prefix_video_latents_frames] = addnoise_condition
+
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=prompt_embeds,
+                        enable_diffusion_forcing=True,
+                        fps=fps_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    if self.do_classifier_free_guidance:
+                        noise_uncond = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            enable_diffusion_forcing=True,
+                            fps=fps_embeds,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                        noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                    update_mask_i = step_update_mask[i]
+                    for idx in range(valid_interval_start, valid_interval_end):
+                        if update_mask_i[idx].item():
+                            latents[:, :, idx, :, :] = sample_schedulers[idx].step(
+                                noise_pred[:, :, idx - valid_interval_start, :, :],
+                                t[idx],
+                                latents[:, :, idx, :, :],
+                                return_dict=False,
+                            )[0]
+
+                    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, t, 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)
+
+                    # call the callback, if provided
+                    if i == len(step_matrix) - 1 or (
+                        (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                    ):
+                        progress_bar.update()
+
+                    if XLA_AVAILABLE:
+                        xm.mark_step()
+
+            # Handle latent accumulation for long videos or use the current latents for short videos
+            if is_long_video:
+                if accumulated_latents is None:
+                    accumulated_latents = latents
+                else:
+                    # Keep overlap frames for conditioning but don't include them in final output
+                    accumulated_latents = torch.cat(
+                        [accumulated_latents, latents[:, :, overlap_history_latent_frames:]], dim=2
+                    )
+
+        if is_long_video:
+            latents = accumulated_latents
+
+        self._current_timestep = None
+
+        # Final decoding step - convert latents to pixels
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return SkyReelsV2PipelineOutput(frames=video)
@@ -0,0 +1,747 @@
+# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL
+import regex as re
+import torch
+from transformers import AutoTokenizer, CLIPProcessor, CLIPVisionModelWithProjection, UMT5EncoderModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...loaders import SkyReelsV2LoraLoaderMixin
+from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel
+from ...schedulers import UniPCMultistepScheduler
+from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import SkyReelsV2PipelineOutput
+
+
+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
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """\
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import (
+        ...     SkyReelsV2ImageToVideoPipeline,
+        ...     UniPCMultistepScheduler,
+        ...     AutoencoderKLWan,
+        ... )
+        >>> from diffusers.utils import export_to_video
+        >>> from PIL import Image
+
+        >>> # Load the pipeline
+        >>> # Available models:
+        >>> # - Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers
+        >>> # - Skywork/SkyReels-V2-I2V-14B-540P-Diffusers
+        >>> # - Skywork/SkyReels-V2-I2V-14B-720P-Diffusers
+        >>> vae = AutoencoderKLWan.from_pretrained(
+        ...     "Skywork/SkyReels-V2-I2V-14B-720P-Diffusers",
+        ...     subfolder="vae",
+        ...     torch_dtype=torch.float32,
+        ... )
+        >>> pipe = SkyReelsV2ImageToVideoPipeline.from_pretrained(
+        ...     "Skywork/SkyReels-V2-I2V-14B-720P-Diffusers",
+        ...     vae=vae,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> flow_shift = 5.0  # 8.0 for T2V, 5.0 for I2V
+        >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+        >>> image = Image.open("path/to/image.png")
+
+        >>> output = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     num_inference_steps=50,
+        ...     height=544,
+        ...     width=960,
+        ...     guidance_scale=5.0,  # 6.0 for T2V, 5.0 for I2V
+        ...     num_frames=97,
+        ... ).frames[0]
+        >>> export_to_video(output, "video.mp4", fps=24, quality=8)
+        ```
+"""
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def prompt_clean(text):
+    text = whitespace_clean(basic_clean(text))
+    return text
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class SkyReelsV2ImageToVideoPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin):
+    r"""
+    Pipeline for Image-to-Video (i2v) generation using SkyReels-V2.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        tokenizer ([`T5Tokenizer`]):
+            Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer),
+            specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant.
+        image_encoder ([`CLIPVisionModelWithProjection`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModelWithProjection),
+            specifically the
+            [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large)
+            variant.
+        transformer ([`SkyReelsV2Transformer3DModel`]):
+            Conditional Transformer to denoise the input latents.
+        scheduler ([`UniPCMultistepScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLWan`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+    """
+
+    model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: UMT5EncoderModel,
+        image_encoder: CLIPVisionModelWithProjection,
+        image_processor: CLIPProcessor,
+        transformer: SkyReelsV2Transformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: UniPCMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            image_encoder=image_encoder,
+            transformer=transformer,
+            scheduler=scheduler,
+            image_processor=image_processor,
+        )
+
+        self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+        self.image_processor = image_processor
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 512,
+        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
+        prompt = [prompt_clean(u) for u in prompt]
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_attention_mask=True,
+            return_tensors="pt",
+        )
+        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
+        seq_lens = mask.gt(0).sum(dim=1).long()
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
+        )
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.encode_image
+    def encode_image(
+        self,
+        image: PipelineImageInput,
+        device: Optional[torch.device] = None,
+    ):
+        device = device or self._execution_device
+        image = self.image_processor(images=image, return_tensors="pt").to(device)
+        image_embeds = self.image_encoder(**image, output_hidden_states=True)
+        return image_embeds.hidden_states[-2]
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.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_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos 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
+        """
+        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_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        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_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        negative_prompt,
+        image,
+        height,
+        width,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if image is not None and image_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        if image is None and image_embeds is None:
+            raise ValueError(
+                "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined."
+            )
+        if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image):
+            raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}")
+        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 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`: {negative_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)}")
+        elif negative_prompt is not None and (
+            not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list)
+        ):
+            raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+    # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.prepare_latents
+    def prepare_latents(
+        self,
+        image: PipelineImageInput,
+        batch_size: int,
+        num_channels_latents: int = 16,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        last_image: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height = height // self.vae_scale_factor_spatial
+        latent_width = width // self.vae_scale_factor_spatial
+
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        image = image.unsqueeze(2)
+        if last_image is None:
+            video_condition = torch.cat(
+                [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], dim=2
+            )
+        else:
+            last_image = last_image.unsqueeze(2)
+            video_condition = torch.cat(
+                [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 2, height, width), last_image],
+                dim=2,
+            )
+        video_condition = video_condition.to(device=device, dtype=self.vae.dtype)
+
+        latents_mean = (
+            torch.tensor(self.vae.config.latents_mean)
+            .view(1, self.vae.config.z_dim, 1, 1, 1)
+            .to(latents.device, latents.dtype)
+        )
+        latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+            latents.device, latents.dtype
+        )
+
+        if isinstance(generator, list):
+            latent_condition = [
+                retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator
+            ]
+            latent_condition = torch.cat(latent_condition)
+        else:
+            latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax")
+            latent_condition = latent_condition.repeat(batch_size, 1, 1, 1, 1)
+
+        latent_condition = latent_condition.to(dtype)
+        latent_condition = (latent_condition - latents_mean) * latents_std
+
+        mask_lat_size = torch.ones(batch_size, 1, num_frames, latent_height, latent_width)
+
+        if last_image is None:
+            mask_lat_size[:, :, list(range(1, num_frames))] = 0
+        else:
+            mask_lat_size[:, :, list(range(1, num_frames - 1))] = 0
+        first_frame_mask = mask_lat_size[:, :, 0:1]
+        first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal)
+        mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:, :]], dim=2)
+        mask_lat_size = mask_lat_size.view(batch_size, -1, self.vae_scale_factor_temporal, latent_height, latent_width)
+        mask_lat_size = mask_lat_size.transpose(1, 2)
+        mask_lat_size = mask_lat_size.to(latent_condition.device)
+
+        return latents, torch.concat([mask_lat_size, latent_condition], dim=1)
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        height: int = 544,
+        width: int = 960,
+        num_frames: int = 97,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        image_embeds: Optional[torch.Tensor] = None,
+        last_image: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "np",
+        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 = 512,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            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`, defaults to `544`):
+                The height of the generated video.
+            width (`int`, defaults to `960`):
+                The width of the generated video.
+            num_frames (`int`, defaults to `97`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, 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_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `negative_prompt` input argument.
+            image_embeds (`torch.Tensor`, *optional*):
+                Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided,
+                image embeddings are generated from the `image` input argument.
+            output_type (`str`, *optional*, defaults to `"np"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] 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`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. 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`, *optional*, defaults to `512`):
+                The maximum sequence length of the prompt.
+
+        Examples:
+
+        Returns:
+            [`~SkyReelsV2PipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            image,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        if num_frames % self.vae_scale_factor_temporal != 1:
+            logger.warning(
+                f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number."
+            )
+            num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+        num_frames = max(num_frames, 1)
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define 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]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+
+        # Encode image embedding
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        if image_embeds is None:
+            if last_image is None:
+                image_embeds = self.encode_image(image, device)
+            else:
+                image_embeds = self.encode_image([image, last_image], device)
+        image_embeds = image_embeds.repeat(batch_size, 1, 1)
+        image_embeds = image_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.vae.config.z_dim
+        image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32)
+        if last_image is not None:
+            last_image = self.video_processor.preprocess(last_image, height=height, width=width).to(
+                device, dtype=torch.float32
+            )
+        latents, condition = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+            last_image,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        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, condition], dim=1).to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_hidden_states_image=image_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        encoder_hidden_states_image=image_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                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, t, 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)
+
+                # call the callback, if provided
+                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)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return SkyReelsV2PipelineOutput(frames=video)
@@ -383,7 +383,8 @@ class OnnxStableDiffusionPipeline(DiffusionPipeline):
        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma

        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
@@ -483,7 +483,7 @@ class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
        self.scheduler.set_timesteps(num_inference_steps)

        # scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        latents = latents * self.scheduler.init_noise_sigma

        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
@@ -481,7 +481,7 @@ class OnnxStableDiffusionUpscalePipeline(DiffusionPipeline):
        timesteps = self.scheduler.timesteps

        # Scale the initial noise by the standard deviation required by the scheduler
-        latents = latents * np.float64(self.scheduler.init_noise_sigma)
+        latents = latents * self.scheduler.init_noise_sigma

        # 5. Add noise to image
        noise_level = np.array([noise_level]).astype(np.int64)
@@ -153,6 +153,8 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
        flow_shift: Optional[float] = 1.0,
        timestep_spacing: str = "linspace",
        steps_offset: int = 0,
+        use_dynamic_shifting: bool = False,
+        time_shift_type: str = "exponential",
    ):
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
@@ -232,7 +234,9 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
        """
        self._begin_index = begin_index

-    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+    def set_timesteps(
+        self, num_inference_steps: int, device: Union[str, torch.device] = None, mu: Optional[float] = None
+    ):
        """
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).

@@ -242,6 +246,9 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
            device (`str` or `torch.device`, *optional*):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        """
+        if mu is not None:
+            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
+            self.config.flow_shift = np.exp(mu)
        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
        if self.config.timestep_spacing == "linspace":
            timesteps = (
@@ -230,6 +230,8 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
        timestep_spacing: str = "linspace",
        steps_offset: int = 0,
        rescale_betas_zero_snr: bool = False,
+        use_dynamic_shifting: bool = False,
+        time_shift_type: str = "exponential",
    ):
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
@@ -330,6 +332,7 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
+        mu: Optional[float] = None,
        timesteps: Optional[List[int]] = None,
    ):
        """
@@ -345,6 +348,9 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
                based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas`
                must be `None`, and `timestep_spacing` attribute will be ignored.
        """
+        if mu is not None:
+            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
+            self.config.flow_shift = np.exp(mu)
        if num_inference_steps is None and timesteps is None:
            raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.")
        if num_inference_steps is not None and timesteps is not None:
@@ -169,6 +169,8 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
        lambda_min_clipped: float = -float("inf"),
        variance_type: Optional[str] = None,
+        use_dynamic_shifting: bool = False,
+        time_shift_type: str = "exponential",
    ):
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
@@ -301,6 +303,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
+        mu: Optional[float] = None,
        timesteps: Optional[List[int]] = None,
    ):
        """
@@ -316,6 +319,9 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
                timestep spacing strategy of equal spacing between timesteps schedule is used. If `timesteps` is
                passed, `num_inference_steps` must be `None`.
        """
+        if mu is not None:
+            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
+            self.config.flow_shift = np.exp(mu)
        if num_inference_steps is None and timesteps is None:
            raise ValueError("Must pass exactly one of  `num_inference_steps` or `timesteps`.")
        if num_inference_steps is not None and timesteps is not None:
@@ -168,6 +168,8 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
+        use_flow_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use flow sigmas for step sizes in the noise schedule during the sampling process.
        timestep_spacing (`str`, defaults to `"linspace"`):
            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
@@ -212,6 +214,8 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
        steps_offset: int = 0,
        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
        rescale_betas_zero_snr: bool = False,
+        use_dynamic_shifting: bool = False,
+        time_shift_type: str = "exponential",
    ):
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
@@ -298,7 +302,9 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
        """
        self._begin_index = begin_index

-    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+    def set_timesteps(
+        self, num_inference_steps: int, device: Union[str, torch.device] = None, mu: Optional[float] = None
+    ):
        """
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).

@@ -309,6 +315,9 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        """
        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
+        if mu is not None:
+            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
+            self.config.flow_shift = np.exp(mu)
        if self.config.timestep_spacing == "linspace":
            timesteps = (
                np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
@@ -5,12 +5,14 @@ import math
 import random
 import re
 import warnings
+from contextlib import contextmanager
 from typing import Any, Dict, Iterable, List, Optional, Tuple, Union

 import numpy as np
 import torch

 from .models import UNet2DConditionModel
+from .pipelines import DiffusionPipeline
 from .schedulers import SchedulerMixin
 from .utils import (
    convert_state_dict_to_diffusers,
@@ -318,6 +320,39 @@ def free_memory():
        torch.xpu.empty_cache()


+@contextmanager
+def offload_models(
+    *modules: Union[torch.nn.Module, DiffusionPipeline], device: Union[str, torch.device], offload: bool = True
+):
+    """
+    Context manager that, if offload=True, moves each module to `device` on enter, then moves it back to its original
+    device on exit.
+
+    Args:
+        device (`str` or `torch.Device`): Device to move the `modules` to.
+        offload (`bool`): Flag to enable offloading.
+    """
+    if offload:
+        is_model = not any(isinstance(m, DiffusionPipeline) for m in modules)
+        # record where each module was
+        if is_model:
+            original_devices = [next(m.parameters()).device for m in modules]
+        else:
+            assert len(modules) == 1
+            original_devices = modules[0].device
+        # move to target device
+        for m in modules:
+            m.to(device)
+
+    try:
+        yield
+    finally:
+        if offload:
+            # move back to original devices
+            for m, orig_dev in zip(modules, original_devices):
+                m.to(orig_dev)
+
+
 def parse_buckets_string(buckets_str):
    """Parses a string defining buckets into a list of (height, width) tuples."""
    if not buckets_str:
@@ -67,6 +67,9 @@ from .import_utils import (
    is_bitsandbytes_version,
    is_bs4_available,
    is_cosmos_guardrail_available,
+    is_flash_attn_3_available,
+    is_flash_attn_available,
+    is_flash_attn_version,
    is_flax_available,
    is_ftfy_available,
    is_gguf_available,
@@ -90,6 +93,8 @@ from .import_utils import (
    is_peft_version,
    is_pytorch_retinaface_available,
    is_safetensors_available,
+    is_sageattention_available,
+    is_sageattention_version,
    is_scipy_available,
    is_sentencepiece_available,
    is_tensorboard_available,
@@ -108,6 +113,7 @@ from .import_utils import (
    is_unidecode_available,
    is_wandb_available,
    is_xformers_available,
+    is_xformers_version,
    requires_backends,
 )
 from .loading_utils import get_module_from_name, get_submodule_by_name, load_image, load_video
@@ -41,6 +41,8 @@ DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules"
 HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(HF_HOME, "modules"))
 DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"]
 DIFFUSERS_REQUEST_TIMEOUT = 60
+DIFFUSERS_ATTN_BACKEND = os.getenv("DIFFUSERS_ATTN_BACKEND", "native")
+DIFFUSERS_ATTN_CHECKS = os.getenv("DIFFUSERS_ATTN_CHECKS", "0") in ENV_VARS_TRUE_VALUES

 # Below should be `True` if the current version of `peft` and `transformers` are compatible with
 # PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are
@@ -258,6 +258,21 @@ class AsymmetricAutoencoderKL(metaclass=DummyObject):
        requires_backends(cls, ["torch"])


+class AttentionBackendName(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AuraFlowTransformer2DModel(metaclass=DummyObject):
    _backends = ["torch"]

@@ -1098,6 +1113,21 @@ class SD3Transformer2DModel(metaclass=DummyObject):
        requires_backends(cls, ["torch"])


+class SkyReelsV2Transformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SparseControlNetModel(metaclass=DummyObject):
    _backends = ["torch"]

@@ -1353,6 +1383,10 @@ class WanVACETransformer3DModel(metaclass=DummyObject):
        requires_backends(cls, ["torch"])


+def attention_backend(*args, **kwargs):
+    requires_backends(attention_backend, ["torch"])
+
+
 class ComponentsManager(metaclass=DummyObject):
    _backends = ["torch"]

@@ -32,6 +32,36 @@ class StableDiffusionXLModularPipeline(metaclass=DummyObject):
        requires_backends(cls, ["torch", "transformers"])


+class WanAutoBlocks(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class WanModularPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class AllegroPipeline(metaclass=DummyObject):
    _backends = ["torch", "transformers"]

@@ -1817,6 +1847,81 @@ class ShapEPipeline(metaclass=DummyObject):
        requires_backends(cls, ["torch", "transformers"])


+class SkyReelsV2DiffusionForcingImageToVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class SkyReelsV2DiffusionForcingPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class SkyReelsV2DiffusionForcingVideoToVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class SkyReelsV2ImageToVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class SkyReelsV2Pipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableAudioPipeline(metaclass=DummyObject):
    _backends = ["torch", "transformers"]

@@ -220,6 +220,9 @@ _pytorch_retinaface_available, _pytorch_retinaface_version = _is_package_availab
 _better_profanity_available, _better_profanity_version = _is_package_available("better_profanity")
 _nltk_available, _nltk_version = _is_package_available("nltk")
 _cosmos_guardrail_available, _cosmos_guardrail_version = _is_package_available("cosmos_guardrail")
+_sageattention_available, _sageattention_version = _is_package_available("sageattention")
+_flash_attn_available, _flash_attn_version = _is_package_available("flash_attn")
+_flash_attn_3_available, _flash_attn_3_version = _is_package_available("flash_attn_3")


 def is_torch_available():
@@ -378,6 +381,18 @@ def is_hpu_available():
    return all(importlib.util.find_spec(lib) for lib in ("habana_frameworks", "habana_frameworks.torch"))


+def is_sageattention_available():
+    return _sageattention_available
+
+
+def is_flash_attn_available():
+    return _flash_attn_available
+
+
+def is_flash_attn_3_available():
+    return _flash_attn_3_available
+
+
 # docstyle-ignore
 FLAX_IMPORT_ERROR = """
 {0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
@@ -804,6 +819,51 @@ def is_optimum_quanto_version(operation: str, version: str):
    return compare_versions(parse(_optimum_quanto_version), operation, version)


+def is_xformers_version(operation: str, version: str):
+    """
+    Compares the current xformers version to a given reference with an operation.
+
+    Args:
+        operation (`str`):
+            A string representation of an operator, such as `">"` or `"<="`
+        version (`str`):
+            A version string
+    """
+    if not _xformers_available:
+        return False
+    return compare_versions(parse(_xformers_version), operation, version)
+
+
+def is_sageattention_version(operation: str, version: str):
+    """
+    Compares the current sageattention version to a given reference with an operation.
+
+    Args:
+        operation (`str`):
+            A string representation of an operator, such as `">"` or `"<="`
+        version (`str`):
+            A version string
+    """
+    if not _sageattention_available:
+        return False
+    return compare_versions(parse(_sageattention_version), operation, version)
+
+
+def is_flash_attn_version(operation: str, version: str):
+    """
+    Compares the current flash-attention version to a given reference with an operation.
+
+    Args:
+        operation (`str`):
+            A string representation of an operator, such as `">"` or `"<="`
+        version (`str`):
+            A version string
+    """
+    if not _flash_attn_available:
+        return False
+    return compare_versions(parse(_flash_attn_version), operation, version)
+
+
 def get_objects_from_module(module):
    """
    Returns a dict of object names and values in a module, while skipping private/internal objects
@@ -75,7 +75,6 @@ from diffusers.utils.testing_utils import (
    require_torch_2,
    require_torch_accelerator,
    require_torch_accelerator_with_training,
-    require_torch_gpu,
    require_torch_multi_accelerator,
    require_torch_version_greater,
    run_test_in_subprocess,
@@ -1829,8 +1828,8 @@ class ModelTesterMixin:

        assert msg_substring in str(err_ctx.exception)

-    @parameterized.expand([0, "cuda", torch.device("cuda")])
-    @require_torch_gpu
+    @parameterized.expand([0, torch_device, torch.device(torch_device)])
+    @require_torch_accelerator
    def test_passing_non_dict_device_map_works(self, device_map):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).eval()
@@ -1839,8 +1838,8 @@ class ModelTesterMixin:
            loaded_model = self.model_class.from_pretrained(tmpdir, device_map=device_map)
            _ = loaded_model(**inputs_dict)

-    @parameterized.expand([("", "cuda"), ("", torch.device("cuda"))])
-    @require_torch_gpu
+    @parameterized.expand([("", torch_device), ("", torch.device(torch_device))])
+    @require_torch_accelerator
    def test_passing_dict_device_map_works(self, name, device):
        # There are other valid dict-based `device_map` values too. It's best to refer to
        # the docs for those: https://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference#the-devicemap.
@@ -1945,10 +1944,11 @@ class ModelPushToHubTester(unittest.TestCase):
        delete_repo(self.repo_id, token=TOKEN)


-@require_torch_gpu
+@require_torch_accelerator
@require_torch_2
@is_torch_compile
@slow
+@require_torch_version_greater("2.7.1")
 class TorchCompileTesterMixin:
    different_shapes_for_compilation = None

@@ -2013,7 +2013,7 @@ class TorchCompileTesterMixin:
        model.eval()
        # TODO: Can test for other group offloading kwargs later if needed.
        group_offload_kwargs = {
-            "onload_device": "cuda",
+            "onload_device": torch_device,
            "offload_device": "cpu",
            "offload_type": "block_level",
            "num_blocks_per_group": 1,
@@ -2047,6 +2047,7 @@ class TorchCompileTesterMixin:
@require_torch_accelerator
@require_peft_backend
@require_peft_version_greater("0.14.0")
+@require_torch_version_greater("2.7.1")
@is_torch_compile
 class LoraHotSwappingForModelTesterMixin:
    """Test that hotswapping does not result in recompilation on the model directly.
@@ -0,0 +1,84 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import torch
+
+from diffusers import SkyReelsV2Transformer3DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+
+
+enable_full_determinism()
+
+
+class SkyReelsV2Transformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
+    model_class = SkyReelsV2Transformer3DModel
+    main_input_name = "hidden_states"
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 1
+        num_channels = 4
+        num_frames = 2
+        height = 16
+        width = 16
+        text_encoder_embedding_dim = 16
+        sequence_length = 12
+
+        hidden_states = torch.randn((batch_size, num_channels, num_frames, height, width)).to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, text_encoder_embedding_dim)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 1, 16, 16)
+
+    @property
+    def output_shape(self):
+        return (4, 1, 16, 16)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "patch_size": (1, 2, 2),
+            "num_attention_heads": 2,
+            "attention_head_dim": 12,
+            "in_channels": 4,
+            "out_channels": 4,
+            "text_dim": 16,
+            "freq_dim": 256,
+            "ffn_dim": 32,
+            "num_layers": 2,
+            "cross_attn_norm": True,
+            "qk_norm": "rms_norm_across_heads",
+            "rope_max_seq_len": 32,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"SkyReelsV2Transformer3DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
@@ -358,7 +358,7 @@ class UNet2DConditionModelTests(ModelTesterMixin, UNetTesterMixin, unittest.Test
    model_class = UNet2DConditionModel
    main_input_name = "sample"
    # We override the items here because the unet under consideration is small.
-    model_split_percents = [0.5, 0.3, 0.4]
+    model_split_percents = [0.5, 0.34, 0.4]

    @property
    def dummy_input(self):
@@ -7,12 +7,7 @@ from transformers import AutoTokenizer, T5EncoderModel
 from diffusers import AutoencoderKL, ChromaPipeline, ChromaTransformer2DModel, FlowMatchEulerDiscreteScheduler
 from diffusers.utils.testing_utils import torch_device

-from ..test_pipelines_common import (
-    FluxIPAdapterTesterMixin,
-    PipelineTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
-)
+from ..test_pipelines_common import FluxIPAdapterTesterMixin, PipelineTesterMixin, check_qkv_fused_layers_exist


 class ChromaPipelineFastTests(
@@ -126,12 +121,10 @@ class ChromaPipelineFastTests(
        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
        # to the pipeline level.
        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -8,12 +8,7 @@ from transformers import AutoTokenizer, T5EncoderModel
 from diffusers import AutoencoderKL, ChromaImg2ImgPipeline, ChromaTransformer2DModel, FlowMatchEulerDiscreteScheduler
 from diffusers.utils.testing_utils import floats_tensor, torch_device

-from ..test_pipelines_common import (
-    FluxIPAdapterTesterMixin,
-    PipelineTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
-)
+from ..test_pipelines_common import FluxIPAdapterTesterMixin, PipelineTesterMixin, check_qkv_fused_layers_exist


 class ChromaImg2ImgPipelineFastTests(
@@ -129,12 +124,10 @@ class ChromaImg2ImgPipelineFastTests(
        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
        # to the pipeline level.
        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -16,11 +16,7 @@ from diffusers.utils.testing_utils import (
 )
 from diffusers.utils.torch_utils import randn_tensor

-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
-)
+from ..test_pipelines_common import PipelineTesterMixin, check_qkv_fused_layers_exist


 class FluxControlNetImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
@@ -170,12 +166,10 @@ class FluxControlNetImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMi
        original_image_slice = image[0, -3:, -3:, -1]

        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -153,11 +153,15 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
        inputs = self.get_dummy_inputs(device)
        video = pipe(**inputs).frames
        generated_video = video[0]
-
        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.0, 0.9686, 0.8549, 0.8078, 0.0, 0.8431, 1.0, 0.4863, 0.7098, 0.1098, 0.8157, 0.4235, 0.6353, 0.2549, 0.5137, 0.5333])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))

    def test_callback_inputs(self):
        sig = inspect.signature(self.pipeline_class.__call__)
@@ -140,11 +140,15 @@ class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase
        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
        generated_image = image[0]
-
        self.assertEqual(generated_image.shape, (3, 32, 32))
-        expected_video = torch.randn(3, 32, 32)
-        max_diff = np.abs(generated_image - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.451, 0.451, 0.4471, 0.451, 0.451, 0.451, 0.451, 0.451, 0.4784, 0.4784, 0.4784, 0.4784, 0.4784, 0.4902, 0.4588, 0.5333])
+        # fmt: on
+
+        generated_slice = generated_image.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))

    def test_callback_inputs(self):
        sig = inspect.signature(self.pipeline_class.__call__)
@@ -147,11 +147,15 @@ class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCas
        inputs = self.get_dummy_inputs(device)
        video = pipe(**inputs).frames
        generated_video = video[0]
-
        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.451, 0.451, 0.4471, 0.451, 0.451, 0.451, 0.451, 0.451, 0.5098, 0.5137, 0.5176, 0.5098, 0.5255, 0.5412, 0.5098, 0.5059])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))

    def test_components_function(self):
        init_components = self.get_dummy_components()
@@ -159,11 +159,15 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
        inputs = self.get_dummy_inputs(device)
        video = pipe(**inputs).frames
        generated_video = video[0]
-
        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
-        expected_video = torch.randn(9, 3, 32, 32)
-        max_diff = np.abs(generated_video - expected_video).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = torch.tensor([0.0, 0.8275, 0.7529, 0.7294, 0.0, 0.6, 1.0, 0.3804, 0.6667, 0.0863, 0.8784, 0.5922, 0.6627, 0.2784, 0.5725, 0.7765])
+        # fmt: on
+
+        generated_slice = generated_video.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))

    def test_components_function(self):
        init_components = self.get_dummy_components()
@@ -28,8 +28,7 @@ from ..test_pipelines_common import (
    FluxIPAdapterTesterMixin,
    PipelineTesterMixin,
    PyramidAttentionBroadcastTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
+    check_qkv_fused_layers_exist,
 )


@@ -171,12 +170,10 @@ class FluxPipelineFastTests(
        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
        # to the pipeline level.
        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -8,11 +8,7 @@ from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPToken
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxControlPipeline, FluxTransformer2DModel
 from diffusers.utils.testing_utils import torch_device

-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
-)
+from ..test_pipelines_common import PipelineTesterMixin, check_qkv_fused_layers_exist


 class FluxControlPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
@@ -140,12 +136,10 @@ class FluxControlPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
        # to the pipeline level.
        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -15,11 +15,7 @@ from diffusers.utils.testing_utils import (
    torch_device,
 )

-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-    check_qkv_fusion_matches_attn_procs_length,
-    check_qkv_fusion_processors_exist,
-)
+from ..test_pipelines_common import PipelineTesterMixin, check_qkv_fused_layers_exist


 class FluxControlInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
@@ -134,12 +130,10 @@ class FluxControlInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin
        # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
        # to the pipeline level.
        pipe.transformer.fuse_qkv_projections()
-        assert check_qkv_fusion_processors_exist(pipe.transformer), (
-            "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        self.assertTrue(
+            check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]),
+            ("Something wrong with the fused attention layers. Expected all the attention projections to be fused."),
        )
-        assert check_qkv_fusion_matches_attn_procs_length(
-            pipe.transformer, pipe.transformer.original_attn_processors
-        ), "Something wrong with the attention processors concerning the fused QKV projections."

        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
@@ -146,11 +146,15 @@ class HiDreamImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs)[0]
        generated_image = image[0]
-
        self.assertEqual(generated_image.shape, (128, 128, 3))
-        expected_image = torch.randn(128, 128, 3).numpy()
-        max_diff = np.abs(generated_image - expected_image).max()
-        self.assertLessEqual(max_diff, 1e10)
+
+        # fmt: off
+        expected_slice = np.array([0.4507, 0.5256, 0.4205, 0.5791, 0.4848, 0.4831, 0.4443, 0.5107, 0.6586, 0.3163, 0.7318, 0.5933, 0.6252, 0.5512, 0.5357, 0.5983])
+        # fmt: on
+
+        generated_slice = generated_image.flatten()
+        generated_slice = np.concatenate([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(np.allclose(generated_slice, expected_slice, atol=1e-3))

    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-4)
--- a/Show More
+++ b/Show More