update

* add ChronoEdit

* add ref to  original function & remove wan2.2 logics

* Update src/diffusers/pipelines/chronoedit/pipeline_chronoedit.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update src/diffusers/pipelines/chronoedit/pipeline_chronoedit.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* add ChronoeEdit test

* add docs

* add docs

* make fix-copies

* fix chronoedit test

---------

Co-authored-by: wjay <wjay@nvidia.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

.github/workflows/push_tests.yml

@@ -76,6 +76,7 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
       - name: Environment
         run: |
           python utils/print_env.py
@@ -127,6 +128,7 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
     - name: Environment
       run: |
@@ -178,6 +180,7 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
+        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
     - name: Environment
       run: |
         python utils/print_env.py

docs/source/en/_toctree.yml

@@ -329,6 +329,8 @@
         title: BriaTransformer2DModel
       - local: api/models/chroma_transformer
         title: ChromaTransformer2DModel
+      - local: api/models/chronoedit_transformer_3d
+        title: ChronoEditTransformer3DModel
       - local: api/models/cogvideox_transformer3d
         title: CogVideoXTransformer3DModel
       - local: api/models/cogview3plus_transformer2d
@@ -373,6 +375,8 @@
         title: QwenImageTransformer2DModel
       - local: api/models/sana_transformer2d
         title: SanaTransformer2DModel
+      - local: api/models/sana_video_transformer3d
+        title: SanaVideoTransformer3DModel
       - local: api/models/sd3_transformer2d
         title: SD3Transformer2DModel
       - local: api/models/skyreels_v2_transformer_3d
@@ -563,6 +567,8 @@
         title: Sana
       - local: api/pipelines/sana_sprint
         title: Sana Sprint
+      - local: api/pipelines/sana_video
+        title: Sana Video
       - local: api/pipelines/self_attention_guidance
         title: Self-Attention Guidance
       - local: api/pipelines/semantic_stable_diffusion
@@ -624,6 +630,8 @@
     - sections:
       - local: api/pipelines/allegro
         title: Allegro
+      - local: api/pipelines/chronoedit
+        title: ChronoEdit
       - local: api/pipelines/cogvideox
         title: CogVideoX
       - local: api/pipelines/consisid

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

@@ -0,0 +1,32 @@
+<!-- Copyright 2025 The ChronoEdit Team and HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# ChronoEditTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data from [ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation](https://huggingface.co/papers/2510.04290) from NVIDIA and University of Toronto, by Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M. Alvarez, Jun Gao, Sanja Fidler, Zian Wang, Huan Ling.
+
+> **TL;DR:** ChronoEdit reframes image editing as a video generation task, using input and edited images as start/end frames to leverage pretrained video models with temporal consistency. A temporal reasoning stage introduces reasoning tokens to ensure physically plausible edits and visualize the editing trajectory.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import ChronoEditTransformer3DModel
+
+transformer = ChronoEditTransformer3DModel.from_pretrained("nvidia/ChronoEdit-14B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## ChronoEditTransformer3DModel
+
+[[autodoc]] ChronoEditTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

@@ -0,0 +1,36 @@
+<!-- Copyright 2025 The SANA-Video Authors and HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# SanaVideoTransformer3DModel
+
+A Diffusion Transformer model for 3D data (video) from [SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer](https://huggingface.co/papers/2509.24695) from NVIDIA and MIT HAN Lab, by Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie.
+
+The abstract from the paper is:
+
+*We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation.*
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import SanaVideoTransformer3DModel
+import torch
+
+transformer = SanaVideoTransformer3DModel.from_pretrained("Efficient-Large-Model/SANA-Video_2B_480p_diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## SanaVideoTransformer3DModel
+
+[[autodoc]] SanaVideoTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput
+

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

@@ -0,0 +1,156 @@
+<!-- Copyright 2025 The ChronoEdit Team and HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <a href="https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference" target="_blank" rel="noopener">
+      <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+    </a>
+  </div>
+</div>
+
+# ChronoEdit
+
+[ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation](https://huggingface.co/papers/2510.04290) from NVIDIA and University of Toronto, by Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M. Alvarez, Jun Gao, Sanja Fidler, Zian Wang, Huan Ling.
+
+> **TL;DR:** ChronoEdit reframes image editing as a video generation task, using input and edited images as start/end frames to leverage pretrained video models with temporal consistency. A temporal reasoning stage introduces reasoning tokens to ensure physically plausible edits and visualize the editing trajectory.
+
+*Recent advances in large generative models have greatly enhanced both image editing and in-context image generation, yet a critical gap remains in ensuring physical consistency, where edited objects must remain coherent. This capability is especially vital for world simulation related tasks. In this paper, we present ChronoEdit, a framework that reframes image editing as a video generation problem. First, ChronoEdit treats the input and edited images as the first and last frames of a video, allowing it to leverage large pretrained video generative models that capture not only object appearance but also the implicit physics of motion and interaction through learned temporal consistency. Second, ChronoEdit introduces a temporal reasoning stage that explicitly performs editing at inference time. Under this setting, target frame is jointly denoised with reasoning tokens to imagine a plausible editing trajectory that constrains the solution space to physically viable transformations. The reasoning tokens are then dropped after a few steps to avoid the high computational cost of rendering a full video. To validate ChronoEdit, we introduce PBench-Edit, a new benchmark of image-prompt pairs for contexts that require physical consistency, and demonstrate that ChronoEdit surpasses state-of-the-art baselines in both visual fidelity and physical plausibility. Project page for code and models: [this https URL](https://research.nvidia.com/labs/toronto-ai/chronoedit).*
+
+The ChronoEdit pipeline is developed by the ChronoEdit Team. The original code is available on [GitHub](https://github.com/nv-tlabs/ChronoEdit), and pretrained models can be found in the [nvidia/ChronoEdit](https://huggingface.co/collections/nvidia/chronoedit) collection on Hugging Face.
+
+
+### Image Editing
+
+```py
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, ChronoEditTransformer3DModel, ChronoEditPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+from PIL import Image
+
+model_id = "nvidia/ChronoEdit-14B-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = ChronoEditTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+pipe = ChronoEditPipeline.from_pretrained(model_id, image_encoder=image_encoder, transformer=transformer, vae=vae, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/spaces/nvidia/ChronoEdit/resolve/main/examples/3.png"
+)
+max_area = 720 * 1280
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+print("width", width, "height", height)
+image = image.resize((width, height))
+prompt = (
+    "The user wants to transform the image by adding a small, cute mouse sitting inside the floral teacup, enjoying a spa bath. The mouse should appear relaxed and cheerful, with a tiny white bath towel draped over its head like a turban. It should be positioned comfortably in the cup’s liquid, with gentle steam rising around it to blend with the cozy atmosphere. "
+    "The mouse’s pose should be natural—perhaps sitting upright with paws resting lightly on the rim or submerged in the tea. The teacup’s floral design, gold trim, and warm lighting must remain unchanged to preserve the original aesthetic. The steam should softly swirl around the mouse, enhancing the spa-like, whimsical mood."
+)
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=5,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+    enable_temporal_reasoning=False,
+    num_temporal_reasoning_steps=0,
+).frames[0]
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+Optionally, enable **temporal reasoning** for improved physical consistency:
+```py
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=29,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+    enable_temporal_reasoning=True,
+    num_temporal_reasoning_steps=50,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+### Inference with 8-Step Distillation Lora
+
+```py
+import torch
+import numpy as np
+from diffusers import AutoencoderKLWan, ChronoEditTransformer3DModel, ChronoEditPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+from PIL import Image
+
+model_id = "nvidia/ChronoEdit-14B-Diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+transformer = ChronoEditTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
+pipe = ChronoEditPipeline.from_pretrained(model_id, image_encoder=image_encoder, transformer=transformer, vae=vae, torch_dtype=torch.bfloat16)
+lora_path = hf_hub_download(repo_id=model_id, filename="lora/chronoedit_distill_lora.safetensors")
+pipe.load_lora_weights(lora_path)
+pipe.fuse_lora(lora_scale=1.0)
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=2.0)
+pipe.to("cuda")
+
+image = load_image(
+    "https://huggingface.co/spaces/nvidia/ChronoEdit/resolve/main/examples/3.png"
+)
+max_area = 720 * 1280
+aspect_ratio = image.height / image.width
+mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+print("width", width, "height", height)
+image = image.resize((width, height))
+prompt = (
+    "The user wants to transform the image by adding a small, cute mouse sitting inside the floral teacup, enjoying a spa bath. The mouse should appear relaxed and cheerful, with a tiny white bath towel draped over its head like a turban. It should be positioned comfortably in the cup’s liquid, with gentle steam rising around it to blend with the cozy atmosphere. "
+    "The mouse’s pose should be natural—perhaps sitting upright with paws resting lightly on the rim or submerged in the tea. The teacup’s floral design, gold trim, and warm lighting must remain unchanged to preserve the original aesthetic. The steam should softly swirl around the mouse, enhancing the spa-like, whimsical mood."
+)
+
+output = pipe(
+    image=image,
+    prompt=prompt,
+    height=height,
+    width=width,
+    num_frames=5,
+    num_inference_steps=8,
+    guidance_scale=1.0,
+    enable_temporal_reasoning=False,
+    num_temporal_reasoning_steps=0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+Image.fromarray((output[-1] * 255).clip(0, 255).astype("uint8")).save("output.png")
+```
+
+## ChronoEditPipeline
+
+[[autodoc]] ChronoEditPipeline
+  - all
+  - __call__
+
+## ChronoEditPipelineOutput
+
+[[autodoc]] pipelines.chronoedit.pipeline_output.ChronoEditPipelineOutput

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

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

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

@@ -0,0 +1,102 @@
+<!-- Copyright 2025 The SANA-Video Authors and HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# SanaVideoPipeline
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+  <img alt="MPS" src="https://img.shields.io/badge/MPS-000000?style=flat&logo=apple&logoColor=white%22">
+</div>
+
+[SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer](https://huggingface.co/papers/2509.24695) from NVIDIA and MIT HAN Lab, by Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie.
+
+The abstract from the paper is:
+
+*We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation. [this https URL](https://github.com/NVlabs/SANA).*
+
+This pipeline was contributed by SANA Team. The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://hf.co/collections/Efficient-Large-Model/sana-video).
+
+Available models:
+
+| Model | Recommended dtype |
+|:-----:|:-----------------:|
+| [`Efficient-Large-Model/SANA-Video_2B_480p_diffusers`](https://huggingface.co/Efficient-Large-Model/ANA-Video_2B_480p_diffusers) | `torch.bfloat16` |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-video) collection for more information.
+
+Note: The recommended dtype mentioned is for the transformer weights. The text encoder and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype. 
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaVideoPipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaVideoTransformer3DModel, SanaVideoPipeline
+from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
+
+quant_config = BitsAndBytesConfig(load_in_8bit=True)
+text_encoder_8bit = AutoModel.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    subfolder="text_encoder",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = SanaVideoTransformer3DModel.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+pipeline = SanaVideoPipeline.from_pretrained(
+    "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
+    text_encoder=text_encoder_8bit,
+    transformer=transformer_8bit,
+    torch_dtype=torch.float16,
+    device_map="balanced",
+)
+
+model_score = 30
+prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest, golden light glimmers on his hair as sunlight filters through the leaves. He wears a light shirt, wind gently blowing his hair and collar, light dances across his face with his movements. The background is blurred, with dappled light and soft tree shadows in the distance. The camera focuses on his lifted gaze, clear and emotional."
+negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
+motion_prompt = f" motion score: {model_score}."
+prompt = prompt + motion_prompt
+
+output = pipeline(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=832,
+    num_frames=81,
+    guidance_scale=6.0,
+    num_inference_steps=50
+).frames[0]
+export_to_video(output, "sana-video-output.mp4", fps=16)
+```
+
+## SanaVideoPipeline
+
+[[autodoc]] SanaVideoPipeline
+  - all
+  - __call__
+
+
+## SanaVideoPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_sana_video.SanaVideoPipelineOutput

examples/unconditional_image_generation/README.md

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

examples/unconditional_image_generation/train_unconditional.py

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

scripts/convert_sana_video_to_diffusers.py

@@ -0,0 +1,324 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import os
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download, snapshot_download
+from termcolor import colored
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from diffusers import (
+    AutoencoderKLWan,
+    DPMSolverMultistepScheduler,
+    FlowMatchEulerDiscreteScheduler,
+    SanaVideoPipeline,
+    SanaVideoTransformer3DModel,
+    UniPCMultistepScheduler,
+)
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+ckpt_ids = ["Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth"]
+# https://github.com/NVlabs/Sana/blob/main/inference_video_scripts/inference_sana_video.py
+
+
+def main(args):
+    cache_dir_path = os.path.expanduser("~/.cache/huggingface/hub")
+
+    if args.orig_ckpt_path is None or args.orig_ckpt_path in ckpt_ids:
+        ckpt_id = args.orig_ckpt_path or ckpt_ids[0]
+        snapshot_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+        file_path = hf_hub_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            filename=f"{'/'.join(ckpt_id.split('/')[2:])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+    else:
+        file_path = args.orig_ckpt_path
+
+    print(colored(f"Loading checkpoint from {file_path}", "green", attrs=["bold"]))
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embedding.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embedding.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    # scheduler
+    flow_shift = 8.0
+
+    # model config
+    layer_num = 20
+    # Positional embedding interpolation scale.
+    qk_norm = True
+
+    # sample size
+    if args.video_size == 480:
+        sample_size = 30  # Wan-VAE: 8xp2 downsample factor
+        patch_size = (1, 2, 2)
+    elif args.video_size == 720:
+        sample_size = 22  # Wan-VAE: 32xp1 downsample factor
+        patch_size = (1, 1, 1)
+    else:
+        raise ValueError(f"Video size {args.video_size} is not supported.")
+
+    for depth in range(layer_num):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        if qk_norm is not None:
+            # Add Q/K normalization for self-attention (attn1) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.k_norm.weight"
+            )
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.point_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_temp.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.t_conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+        if qk_norm is not None:
+            # Add Q/K normalization for cross-attention (attn2) - needed for Sana-Sprint and Sana-1.5
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_q.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn2.norm_k.weight"] = state_dict.pop(
+                f"blocks.{depth}.cross_attn.k_norm.weight"
+            )
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.bias"
+        )
+
+    # Final block.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
+    converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
+
+    # Transformer
+    with CTX():
+        transformer_kwargs = {
+            "in_channels": 16,
+            "out_channels": 16,
+            "num_attention_heads": 20,
+            "attention_head_dim": 112,
+            "num_layers": 20,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "caption_channels": 2304,
+            "mlp_ratio": 3.0,
+            "attention_bias": False,
+            "sample_size": sample_size,
+            "patch_size": patch_size,
+            "norm_elementwise_affine": False,
+            "norm_eps": 1e-6,
+            "qk_norm": "rms_norm_across_heads",
+            "rope_max_seq_len": 1024,
+        }
+
+        transformer = SanaVideoTransformer3DModel(**transformer_kwargs)
+
+    transformer.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    try:
+        state_dict.pop("y_embedder.y_embedding")
+        state_dict.pop("pos_embed")
+        state_dict.pop("logvar_linear.weight")
+        state_dict.pop("logvar_linear.bias")
+    except KeyError:
+        print("y_embedder.y_embedding or pos_embed not found in the state_dict")
+
+    assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    transformer = transformer.to(weight_dtype)
+
+    if not args.save_full_pipeline:
+        print(
+            colored(
+                f"Only saving transformer model of {args.model_type}. "
+                f"Set --save_full_pipeline to save the whole Pipeline",
+                "green",
+                attrs=["bold"],
+            )
+        )
+        transformer.save_pretrained(
+            os.path.join(args.dump_path, "transformer"), safe_serialization=True, max_shard_size="5GB"
+        )
+    else:
+        print(colored(f"Saving the whole Pipeline containing {args.model_type}", "green", attrs=["bold"]))
+        # VAE
+        vae = AutoencoderKLWan.from_pretrained(
+            "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32
+        )
+
+        # Text Encoder
+        text_encoder_model_path = "Efficient-Large-Model/gemma-2-2b-it"
+        tokenizer = AutoTokenizer.from_pretrained(text_encoder_model_path)
+        tokenizer.padding_side = "right"
+        text_encoder = AutoModelForCausalLM.from_pretrained(
+            text_encoder_model_path, torch_dtype=torch.bfloat16
+        ).get_decoder()
+
+        # Choose the appropriate pipeline and scheduler based on model type
+        # Original Sana scheduler
+        if args.scheduler_type == "flow-dpm_solver":
+            scheduler = DPMSolverMultistepScheduler(
+                flow_shift=flow_shift,
+                use_flow_sigmas=True,
+                prediction_type="flow_prediction",
+            )
+        elif args.scheduler_type == "flow-euler":
+            scheduler = FlowMatchEulerDiscreteScheduler(shift=flow_shift)
+        elif args.scheduler_type == "uni-pc":
+            scheduler = UniPCMultistepScheduler(
+                prediction_type="flow_prediction",
+                use_flow_sigmas=True,
+                num_train_timesteps=1000,
+                flow_shift=flow_shift,
+            )
+        else:
+            raise ValueError(f"Scheduler type {args.scheduler_type} is not supported")
+
+        pipe = SanaVideoPipeline(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+        )
+
+        pipe.save_pretrained(args.dump_path, safe_serialization=True, max_shard_size="5GB")
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--video_size",
+        default=480,
+        type=int,
+        choices=[480, 720],
+        required=False,
+        help="Video size of pretrained model, 480 or 720.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="SanaVideo",
+        type=str,
+        choices=[
+            "SanaVideo",
+        ],
+    )
+    parser.add_argument(
+        "--scheduler_type",
+        default="flow-dpm_solver",
+        type=str,
+        choices=["flow-dpm_solver", "flow-euler", "uni-pc"],
+        help="Scheduler type to use.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipeline elements in one.")
+    parser.add_argument("--dtype", default="fp32", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+
+    main(args)

src/diffusers/__init__.py

@@ -202,6 +202,7 @@ else:
             "BriaTransformer2DModel",
             "CacheMixin",
             "ChromaTransformer2DModel",
+            "ChronoEditTransformer3DModel",
             "CogVideoXTransformer3DModel",
             "CogView3PlusTransformer2DModel",
             "CogView4Transformer2DModel",
@@ -246,6 +247,7 @@ else:
             "QwenImageTransformer2DModel",
             "SanaControlNetModel",
             "SanaTransformer2DModel",
+            "SanaVideoTransformer3DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
             "SD3Transformer2DModel",
@@ -435,6 +437,7 @@ else:
             "BriaPipeline",
             "ChromaImg2ImgPipeline",
             "ChromaPipeline",
+            "ChronoEditPipeline",
             "CLIPImageProjection",
             "CogVideoXFunControlPipeline",
             "CogVideoXImageToVideoPipeline",
@@ -544,6 +547,7 @@ else:
             "SanaPipeline",
             "SanaSprintImg2ImgPipeline",
             "SanaSprintPipeline",
+            "SanaVideoPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -907,6 +911,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             BriaTransformer2DModel,
             CacheMixin,
             ChromaTransformer2DModel,
+            ChronoEditTransformer3DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,
@@ -951,6 +956,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageTransformer2DModel,
             SanaControlNetModel,
             SanaTransformer2DModel,
+            SanaVideoTransformer3DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
             SD3Transformer2DModel,
@@ -1110,6 +1116,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             BriaPipeline,
             ChromaImg2ImgPipeline,
             ChromaPipeline,
+            ChronoEditPipeline,
             CLIPImageProjection,
             CogVideoXFunControlPipeline,
             CogVideoXImageToVideoPipeline,
@@ -1219,6 +1226,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             SanaPipeline,
             SanaSprintImg2ImgPipeline,
             SanaSprintPipeline,
+            SanaVideoPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,

src/diffusers/hooks/context_parallel.py

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

src/diffusers/image_processor.py

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

src/diffusers/models/__init__.py

@@ -86,6 +86,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_bria"] = ["BriaTransformer2DModel"]
     _import_structure["transformers.transformer_bria_fibo"] = ["BriaFiboTransformer2DModel"]
     _import_structure["transformers.transformer_chroma"] = ["ChromaTransformer2DModel"]
+    _import_structure["transformers.transformer_chronoedit"] = ["ChronoEditTransformer3DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_cogview4"] = ["CogView4Transformer2DModel"]
     _import_structure["transformers.transformer_cosmos"] = ["CosmosTransformer3DModel"]
@@ -102,6 +103,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_omnigen"] = ["OmniGenTransformer2DModel"]
     _import_structure["transformers.transformer_prx"] = ["PRXTransformer2DModel"]
     _import_structure["transformers.transformer_qwenimage"] = ["QwenImageTransformer2DModel"]
+    _import_structure["transformers.transformer_sana_video"] = ["SanaVideoTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
@@ -178,6 +180,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             BriaFiboTransformer2DModel,
             BriaTransformer2DModel,
             ChromaTransformer2DModel,
+            ChronoEditTransformer3DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,
             CogView4Transformer2DModel,
@@ -204,6 +207,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PRXTransformer2DModel,
             QwenImageTransformer2DModel,
             SanaTransformer2DModel,
+            SanaVideoTransformer3DModel,
             SD3Transformer2DModel,
             SkyReelsV2Transformer3DModel,
             StableAudioDiTModel,

src/diffusers/models/transformers/__init__.py

@@ -20,6 +20,7 @@ if is_torch_available():
     from .transformer_bria import BriaTransformer2DModel
     from .transformer_bria_fibo import BriaFiboTransformer2DModel
     from .transformer_chroma import ChromaTransformer2DModel
+    from .transformer_chronoedit import ChronoEditTransformer3DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_cogview4 import CogView4Transformer2DModel
     from .transformer_cosmos import CosmosTransformer3DModel
@@ -36,6 +37,7 @@ if is_torch_available():
     from .transformer_omnigen import OmniGenTransformer2DModel
     from .transformer_prx import PRXTransformer2DModel
     from .transformer_qwenimage import QwenImageTransformer2DModel
+    from .transformer_sana_video import SanaVideoTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_chronoedit.py

@@ -0,0 +1,735 @@
+# Copyright 2025 The ChronoEdit Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import maybe_allow_in_graph
+from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
+from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
+from ..attention_dispatch import dispatch_attention_fn
+from ..cache_utils import CacheMixin
+from ..embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import FP32LayerNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.models.transformers.transformer_wan._get_qkv_projections
+def _get_qkv_projections(attn: "WanAttention", hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor):
+    # encoder_hidden_states is only passed for cross-attention
+    if encoder_hidden_states is None:
+        encoder_hidden_states = hidden_states
+
+    if attn.fused_projections:
+        if attn.cross_attention_dim_head is None:
+            # In self-attention layers, we can fuse the entire QKV projection into a single linear
+            query, key, value = attn.to_qkv(hidden_states).chunk(3, dim=-1)
+        else:
+            # In cross-attention layers, we can only fuse the KV projections into a single linear
+            query = attn.to_q(hidden_states)
+            key, value = attn.to_kv(encoder_hidden_states).chunk(2, dim=-1)
+    else:
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+    return query, key, value
+
+
+# Copied from diffusers.models.transformers.transformer_wan._get_added_kv_projections
+def _get_added_kv_projections(attn: "WanAttention", encoder_hidden_states_img: torch.Tensor):
+    if attn.fused_projections:
+        key_img, value_img = attn.to_added_kv(encoder_hidden_states_img).chunk(2, dim=-1)
+    else:
+        key_img = attn.add_k_proj(encoder_hidden_states_img)
+        value_img = attn.add_v_proj(encoder_hidden_states_img)
+    return key_img, value_img
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanAttnProcessor
+class WanAttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "WanAttnProcessor requires PyTorch 2.0. To use it, please upgrade PyTorch to version 2.0 or higher."
+            )
+
+    def __call__(
+        self,
+        attn: "WanAttention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            # 512 is the context length of the text encoder, hardcoded for now
+            image_context_length = encoder_hidden_states.shape[1] - 512
+            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
+            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]
+
+        query, key, value = _get_qkv_projections(attn, hidden_states, encoder_hidden_states)
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(
+                hidden_states: torch.Tensor,
+                freqs_cos: torch.Tensor,
+                freqs_sin: torch.Tensor,
+            ):
+                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                cos = freqs_cos[..., 0::2]
+                sin = freqs_sin[..., 1::2]
+                out = torch.empty_like(hidden_states)
+                out[..., 0::2] = x1 * cos - x2 * sin
+                out[..., 1::2] = x1 * sin + x2 * cos
+                return out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, *rotary_emb)
+            key = apply_rotary_emb(key, *rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img, value_img = _get_added_kv_projections(attn, encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1))
+            value_img = value_img.unflatten(2, (attn.heads, -1))
+
+            hidden_states_img = dispatch_attention_fn(
+                query,
+                key_img,
+                value_img,
+                attn_mask=None,
+                dropout_p=0.0,
+                is_causal=False,
+                backend=self._attention_backend,
+                parallel_config=self._parallel_config,
+            )
+            hidden_states_img = hidden_states_img.flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanAttnProcessor2_0
+class WanAttnProcessor2_0:
+    def __new__(cls, *args, **kwargs):
+        deprecation_message = (
+            "The WanAttnProcessor2_0 class is deprecated and will be removed in a future version. "
+            "Please use WanAttnProcessor instead. "
+        )
+        deprecate("WanAttnProcessor2_0", "1.0.0", deprecation_message, standard_warn=False)
+        return WanAttnProcessor(*args, **kwargs)
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanAttention
+class WanAttention(torch.nn.Module, AttentionModuleMixin):
+    _default_processor_cls = WanAttnProcessor
+    _available_processors = [WanAttnProcessor]
+
+    def __init__(
+        self,
+        dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        eps: float = 1e-5,
+        dropout: float = 0.0,
+        added_kv_proj_dim: Optional[int] = None,
+        cross_attention_dim_head: Optional[int] = None,
+        processor=None,
+        is_cross_attention=None,
+    ):
+        super().__init__()
+
+        self.inner_dim = dim_head * heads
+        self.heads = heads
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.cross_attention_dim_head = cross_attention_dim_head
+        self.kv_inner_dim = self.inner_dim if cross_attention_dim_head is None else cross_attention_dim_head * heads
+
+        self.to_q = torch.nn.Linear(dim, self.inner_dim, bias=True)
+        self.to_k = torch.nn.Linear(dim, self.kv_inner_dim, bias=True)
+        self.to_v = torch.nn.Linear(dim, self.kv_inner_dim, bias=True)
+        self.to_out = torch.nn.ModuleList(
+            [
+                torch.nn.Linear(self.inner_dim, dim, bias=True),
+                torch.nn.Dropout(dropout),
+            ]
+        )
+        self.norm_q = torch.nn.RMSNorm(dim_head * heads, eps=eps, elementwise_affine=True)
+        self.norm_k = torch.nn.RMSNorm(dim_head * heads, eps=eps, elementwise_affine=True)
+
+        self.add_k_proj = self.add_v_proj = None
+        if added_kv_proj_dim is not None:
+            self.add_k_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=True)
+            self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=True)
+            self.norm_added_k = torch.nn.RMSNorm(dim_head * heads, eps=eps)
+
+        self.is_cross_attention = cross_attention_dim_head is not None
+
+        self.set_processor(processor)
+
+    def fuse_projections(self):
+        if getattr(self, "fused_projections", False):
+            return
+
+        if self.cross_attention_dim_head is None:
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_qkv = nn.Linear(in_features, out_features, bias=True)
+            self.to_qkv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+        else:
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_kv = nn.Linear(in_features, out_features, bias=True)
+            self.to_kv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+
+        if self.added_kv_proj_dim is not None:
+            concatenated_weights = torch.cat([self.add_k_proj.weight.data, self.add_v_proj.weight.data])
+            concatenated_bias = torch.cat([self.add_k_proj.bias.data, self.add_v_proj.bias.data])
+            out_features, in_features = concatenated_weights.shape
+            with torch.device("meta"):
+                self.to_added_kv = nn.Linear(in_features, out_features, bias=True)
+            self.to_added_kv.load_state_dict(
+                {"weight": concatenated_weights, "bias": concatenated_bias}, strict=True, assign=True
+            )
+
+        self.fused_projections = True
+
+    @torch.no_grad()
+    def unfuse_projections(self):
+        if not getattr(self, "fused_projections", False):
+            return
+
+        if hasattr(self, "to_qkv"):
+            delattr(self, "to_qkv")
+        if hasattr(self, "to_kv"):
+            delattr(self, "to_kv")
+        if hasattr(self, "to_added_kv"):
+            delattr(self, "to_added_kv")
+
+        self.fused_projections = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, rotary_emb, **kwargs)
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanImageEmbedding
+class WanImageEmbedding(torch.nn.Module):
+    def __init__(self, in_features: int, out_features: int, pos_embed_seq_len=None):
+        super().__init__()
+
+        self.norm1 = FP32LayerNorm(in_features)
+        self.ff = FeedForward(in_features, out_features, mult=1, activation_fn="gelu")
+        self.norm2 = FP32LayerNorm(out_features)
+        if pos_embed_seq_len is not None:
+            self.pos_embed = nn.Parameter(torch.zeros(1, pos_embed_seq_len, in_features))
+        else:
+            self.pos_embed = None
+
+    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        if self.pos_embed is not None:
+            batch_size, seq_len, embed_dim = encoder_hidden_states_image.shape
+            encoder_hidden_states_image = encoder_hidden_states_image.view(-1, 2 * seq_len, embed_dim)
+            encoder_hidden_states_image = encoder_hidden_states_image + self.pos_embed
+
+        hidden_states = self.norm1(encoder_hidden_states_image)
+        hidden_states = self.ff(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+        return hidden_states
+
+
+# Copied from diffusers.models.transformers.transformer_wan.WanTimeTextImageEmbedding
+class WanTimeTextImageEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        time_freq_dim: int,
+        time_proj_dim: int,
+        text_embed_dim: int,
+        image_embed_dim: Optional[int] = None,
+        pos_embed_seq_len: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.timesteps_proj = Timesteps(num_channels=time_freq_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.time_embedder = TimestepEmbedding(in_channels=time_freq_dim, time_embed_dim=dim)
+        self.act_fn = nn.SiLU()
+        self.time_proj = nn.Linear(dim, time_proj_dim)
+        self.text_embedder = PixArtAlphaTextProjection(text_embed_dim, dim, act_fn="gelu_tanh")
+
+        self.image_embedder = None
+        if image_embed_dim is not None:
+            self.image_embedder = WanImageEmbedding(image_embed_dim, dim, pos_embed_seq_len=pos_embed_seq_len)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        timestep_seq_len: Optional[int] = None,
+    ):
+        timestep = self.timesteps_proj(timestep)
+        if timestep_seq_len is not None:
+            timestep = timestep.unflatten(0, (-1, timestep_seq_len))
+
+        time_embedder_dtype = next(iter(self.time_embedder.parameters())).dtype
+        if timestep.dtype != time_embedder_dtype and time_embedder_dtype != torch.int8:
+            timestep = timestep.to(time_embedder_dtype)
+        temb = self.time_embedder(timestep).type_as(encoder_hidden_states)
+        timestep_proj = self.time_proj(self.act_fn(temb))
+
+        encoder_hidden_states = self.text_embedder(encoder_hidden_states)
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)
+
+        return temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image
+
+
+class ChronoEditRotaryPosEmbed(nn.Module):
+    def __init__(
+        self,
+        attention_head_dim: int,
+        patch_size: Tuple[int, int, int],
+        max_seq_len: int,
+        theta: float = 10000.0,
+        temporal_skip_len: int = 8,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+        self.temporal_skip_len = temporal_skip_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+        freqs_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+
+        freqs_cos = []
+        freqs_sin = []
+
+        for dim in [t_dim, h_dim, w_dim]:
+            freq_cos, freq_sin = get_1d_rotary_pos_embed(
+                dim,
+                max_seq_len,
+                theta,
+                use_real=True,
+                repeat_interleave_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            freqs_cos.append(freq_cos)
+            freqs_sin.append(freq_sin)
+
+        self.register_buffer("freqs_cos", torch.cat(freqs_cos, dim=1), persistent=False)
+        self.register_buffer("freqs_sin", torch.cat(freqs_sin, dim=1), persistent=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        split_sizes = [
+            self.attention_head_dim - 2 * (self.attention_head_dim // 3),
+            self.attention_head_dim // 3,
+            self.attention_head_dim // 3,
+        ]
+
+        freqs_cos = self.freqs_cos.split(split_sizes, dim=1)
+        freqs_sin = self.freqs_sin.split(split_sizes, dim=1)
+
+        if num_frames == 2:
+            freqs_cos_f = freqs_cos[0][: self.temporal_skip_len][[0, -1]].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        else:
+            freqs_cos_f = freqs_cos[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_h = freqs_cos[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_w = freqs_cos[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        if num_frames == 2:
+            freqs_sin_f = freqs_sin[0][: self.temporal_skip_len][[0, -1]].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        else:
+            freqs_sin_f = freqs_sin[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_h = freqs_sin[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_w = freqs_sin[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_cos = torch.cat([freqs_cos_f, freqs_cos_h, freqs_cos_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+        freqs_sin = torch.cat([freqs_sin_f, freqs_sin_h, freqs_sin_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+
+        return freqs_cos, freqs_sin
+
+
+@maybe_allow_in_graph
+# Copied from diffusers.models.transformers.transformer_wan.WanTransformerBlock
+class WanTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        ffn_dim: int,
+        num_heads: int,
+        qk_norm: str = "rms_norm_across_heads",
+        cross_attn_norm: bool = False,
+        eps: float = 1e-6,
+        added_kv_proj_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # 1. Self-attention
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+        self.attn1 = WanAttention(
+            dim=dim,
+            heads=num_heads,
+            dim_head=dim // num_heads,
+            eps=eps,
+            cross_attention_dim_head=None,
+            processor=WanAttnProcessor(),
+        )
+
+        # 2. Cross-attention
+        self.attn2 = WanAttention(
+            dim=dim,
+            heads=num_heads,
+            dim_head=dim // num_heads,
+            eps=eps,
+            added_kv_proj_dim=added_kv_proj_dim,
+            cross_attention_dim_head=dim // num_heads,
+            processor=WanAttnProcessor(),
+        )
+        self.norm2 = FP32LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+
+        # 3. Feed-forward
+        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
+        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        rotary_emb: torch.Tensor,
+    ) -> torch.Tensor:
+        if temb.ndim == 4:
+            # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table.unsqueeze(0) + temb.float()
+            ).chunk(6, dim=2)
+            # batch_size, seq_len, 1, inner_dim
+            shift_msa = shift_msa.squeeze(2)
+            scale_msa = scale_msa.squeeze(2)
+            gate_msa = gate_msa.squeeze(2)
+            c_shift_msa = c_shift_msa.squeeze(2)
+            c_scale_msa = c_scale_msa.squeeze(2)
+            c_gate_msa = c_gate_msa.squeeze(2)
+        else:
+            # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table + temb.float()
+            ).chunk(6, dim=1)
+
+        # 1. Self-attention
+        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)
+        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)
+
+        # 2. Cross-attention
+        norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
+        attn_output = self.attn2(norm_hidden_states, encoder_hidden_states, None, None)
+        hidden_states = hidden_states + attn_output
+
+        # 3. Feed-forward
+        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
+            hidden_states
+        )
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
+
+        return hidden_states
+
+
+# modified from diffusers.models.transformers.transformer_wan.WanTransformer3DModel
+class ChronoEditTransformer3DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin, AttentionMixin
+):
+    r"""
+    A Transformer model for video-like data used in the ChronoEdit model.
+
+    Args:
+        patch_size (`Tuple[int]`, defaults to `(1, 2, 2)`):
+            3D patch dimensions for video embedding (t_patch, h_patch, w_patch).
+        num_attention_heads (`int`, defaults to `40`):
+            Fixed length for text embeddings.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_dim (`int`, defaults to `512`):
+            Input dimension for text embeddings.
+        freq_dim (`int`, defaults to `256`):
+            Dimension for sinusoidal time embeddings.
+        ffn_dim (`int`, defaults to `13824`):
+            Intermediate dimension in feed-forward network.
+        num_layers (`int`, defaults to `40`):
+            The number of layers of transformer blocks to use.
+        window_size (`Tuple[int]`, defaults to `(-1, -1)`):
+            Window size for local attention (-1 indicates global attention).
+        cross_attn_norm (`bool`, defaults to `True`):
+            Enable cross-attention normalization.
+        qk_norm (`bool`, defaults to `True`):
+            Enable query/key normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        add_img_emb (`bool`, defaults to `False`):
+            Whether to use img_emb.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
+    _no_split_modules = ["WanTransformerBlock"]
+    _keep_in_fp32_modules = ["time_embedder", "scale_shift_table", "norm1", "norm2", "norm3"]
+    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
+    _repeated_blocks = ["WanTransformerBlock"]
+    _cp_plan = {
+        "rope": {
+            0: ContextParallelInput(split_dim=1, expected_dims=4, split_output=True),
+            1: ContextParallelInput(split_dim=1, expected_dims=4, split_output=True),
+        },
+        "blocks.0": {
+            "hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
+        },
+        "blocks.*": {
+            "encoder_hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
+        },
+        "proj_out": ContextParallelOutput(gather_dim=1, expected_dims=3),
+    }
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: Tuple[int] = (1, 2, 2),
+        num_attention_heads: int = 40,
+        attention_head_dim: int = 128,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        text_dim: int = 4096,
+        freq_dim: int = 256,
+        ffn_dim: int = 13824,
+        num_layers: int = 40,
+        cross_attn_norm: bool = True,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        eps: float = 1e-6,
+        image_dim: Optional[int] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        rope_max_seq_len: int = 1024,
+        pos_embed_seq_len: Optional[int] = None,
+        rope_temporal_skip_len: int = 8,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Patch & position embedding
+        self.rope = ChronoEditRotaryPosEmbed(
+            attention_head_dim, patch_size, rope_max_seq_len, temporal_skip_len=rope_temporal_skip_len
+        )
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Condition embeddings
+        # image_embedding_dim=1280 for I2V model
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=freq_dim,
+            time_proj_dim=inner_dim * 6,
+            text_embed_dim=text_dim,
+            image_embed_dim=image_dim,
+            pos_embed_seq_len=pos_embed_seq_len,
+        )
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList(
+            [
+                WanTransformerBlock(
+                    inner_dim, ffn_dim, num_attention_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
+        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        # timestep shape: batch_size, or batch_size, seq_len (wan 2.2 ti2v)
+        if timestep.ndim == 2:
+            ts_seq_len = timestep.shape[1]
+            timestep = timestep.flatten()  # batch_size * seq_len
+        else:
+            ts_seq_len = None
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image, timestep_seq_len=ts_seq_len
+        )
+        if ts_seq_len is not None:
+            # batch_size, seq_len, 6, inner_dim
+            timestep_proj = timestep_proj.unflatten(2, (6, -1))
+        else:
+            # batch_size, 6, inner_dim
+            timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        if temb.ndim == 3:
+            # batch_size, seq_len, inner_dim (wan 2.2 ti2v)
+            shift, scale = (self.scale_shift_table.unsqueeze(0).to(temb.device) + temb.unsqueeze(2)).chunk(2, dim=2)
+            shift = shift.squeeze(2)
+            scale = scale.squeeze(2)
+        else:
+            # batch_size, inner_dim
+            shift, scale = (self.scale_shift_table.to(temb.device) + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_hunyuan_video.py

@@ -24,6 +24,7 @@ from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import PeftAdapterMixin
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ..attention import FeedForward
+from ..attention_dispatch import dispatch_attention_fn
 from ..attention_processor import Attention, AttentionProcessor
 from ..cache_utils import CacheMixin
 from ..embeddings import (
@@ -42,6 +43,9 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 class HunyuanVideoAttnProcessor2_0:
+    _attention_backend = None
+    _parallel_config = None
+
     def __init__(self):
         if not hasattr(F, "scaled_dot_product_attention"):
             raise ImportError(
@@ -64,9 +68,9 @@ class HunyuanVideoAttnProcessor2_0:
         key = attn.to_k(hidden_states)
         value = attn.to_v(hidden_states)
 
-        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)
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
 
         # 2. QK normalization
         if attn.norm_q is not None:
@@ -81,21 +85,29 @@ class HunyuanVideoAttnProcessor2_0:
             if attn.add_q_proj is None and encoder_hidden_states is not None:
                 query = torch.cat(
                     [
-                        apply_rotary_emb(query[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
-                        query[:, :, -encoder_hidden_states.shape[1] :],
+                        apply_rotary_emb(
+                            query[:, : -encoder_hidden_states.shape[1]],
+                            image_rotary_emb,
+                            sequence_dim=1,
+                        ),
+                        query[:, -encoder_hidden_states.shape[1] :],
                     ],
-                    dim=2,
+                    dim=1,
                 )
                 key = torch.cat(
                     [
-                        apply_rotary_emb(key[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
-                        key[:, :, -encoder_hidden_states.shape[1] :],
+                        apply_rotary_emb(
+                            key[:, : -encoder_hidden_states.shape[1]],
+                            image_rotary_emb,
+                            sequence_dim=1,
+                        ),
+                        key[:, -encoder_hidden_states.shape[1] :],
                     ],
-                    dim=2,
+                    dim=1,
                 )
             else:
-                query = apply_rotary_emb(query, image_rotary_emb)
-                key = apply_rotary_emb(key, image_rotary_emb)
+                query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+                key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
 
         # 4. Encoder condition QKV projection and normalization
         if attn.add_q_proj is not None and encoder_hidden_states is not None:
@@ -103,24 +115,31 @@ class HunyuanVideoAttnProcessor2_0:
             encoder_key = attn.add_k_proj(encoder_hidden_states)
             encoder_value = attn.add_v_proj(encoder_hidden_states)
 
-            encoder_query = encoder_query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
-            encoder_key = encoder_key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
-            encoder_value = encoder_value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            encoder_query = encoder_query.unflatten(2, (attn.heads, -1))
+            encoder_key = encoder_key.unflatten(2, (attn.heads, -1))
+            encoder_value = encoder_value.unflatten(2, (attn.heads, -1))
 
             if attn.norm_added_q is not None:
                 encoder_query = attn.norm_added_q(encoder_query)
             if attn.norm_added_k is not None:
                 encoder_key = attn.norm_added_k(encoder_key)
 
-            query = torch.cat([query, encoder_query], dim=2)
-            key = torch.cat([key, encoder_key], dim=2)
-            value = torch.cat([value, encoder_value], dim=2)
+            query = torch.cat([query, encoder_query], dim=1)
+            key = torch.cat([key, encoder_key], dim=1)
+            value = torch.cat([value, encoder_value], dim=1)
 
         # 5. Attention
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
         )
-        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.flatten(2, 3)
         hidden_states = hidden_states.to(query.dtype)
 
         # 6. Output projection

src/diffusers/models/transformers/transformer_sana_video.py

@@ -0,0 +1,703 @@
+# Copyright 2025 The HuggingFace Team and SANA-Video 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.functional as F
+from torch import nn
+
+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 AttentionMixin
+from ..attention_dispatch import dispatch_attention_fn
+from ..attention_processor import Attention
+from ..embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class GLUMBTempConv(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: Optional[str] = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+        self.conv_temp = nn.Conv2d(
+            out_channels, out_channels, kernel_size=(3, 1), stride=1, padding=(1, 0), bias=False
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.residual_connection:
+            residual = hidden_states
+        batch_size, num_frames, height, width, num_channels = hidden_states.shape
+        hidden_states = hidden_states.view(batch_size * num_frames, height, width, num_channels).permute(0, 3, 1, 2)
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        # Temporal aggregation
+        hidden_states_temporal = hidden_states.view(batch_size, num_frames, num_channels, height * width).permute(
+            0, 2, 1, 3
+        )
+        hidden_states = hidden_states_temporal + self.conv_temp(hidden_states_temporal)
+        hidden_states = hidden_states.permute(0, 2, 3, 1).view(batch_size, num_frames, height, width, num_channels)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class SanaLinearAttnProcessor3_0:
+    r"""
+    Processor for implementing scaled dot-product linear attention.
+    """
+
+    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:
+        original_dtype = hidden_states.dtype
+
+        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))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+        # B,N,H,C
+
+        query = F.relu(query)
+        key = F.relu(key)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(
+                hidden_states: torch.Tensor,
+                freqs_cos: torch.Tensor,
+                freqs_sin: torch.Tensor,
+            ):
+                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                cos = freqs_cos[..., 0::2]
+                sin = freqs_sin[..., 1::2]
+                out = torch.empty_like(hidden_states)
+                out[..., 0::2] = x1 * cos - x2 * sin
+                out[..., 1::2] = x1 * sin + x2 * cos
+                return out.type_as(hidden_states)
+
+            query_rotate = apply_rotary_emb(query, *rotary_emb)
+            key_rotate = apply_rotary_emb(key, *rotary_emb)
+
+        # B,H,C,N
+        query = query.permute(0, 2, 3, 1)
+        key = key.permute(0, 2, 3, 1)
+        query_rotate = query_rotate.permute(0, 2, 3, 1)
+        key_rotate = key_rotate.permute(0, 2, 3, 1)
+        value = value.permute(0, 2, 3, 1)
+
+        query_rotate, key_rotate, value = query_rotate.float(), key_rotate.float(), value.float()
+
+        z = 1 / (key.sum(dim=-1, keepdim=True).transpose(-2, -1) @ query + 1e-15)
+
+        scores = torch.matmul(value, key_rotate.transpose(-1, -2))
+        hidden_states = torch.matmul(scores, query_rotate)
+
+        hidden_states = hidden_states * z
+        # B,H,C,N
+        hidden_states = hidden_states.flatten(1, 2).transpose(1, 2)
+        hidden_states = hidden_states.to(original_dtype)
+
+        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.WanRotaryPosEmbed
+class WanRotaryPosEmbed(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_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+
+        freqs_cos = []
+        freqs_sin = []
+
+        for dim in [t_dim, h_dim, w_dim]:
+            freq_cos, freq_sin = get_1d_rotary_pos_embed(
+                dim,
+                max_seq_len,
+                theta,
+                use_real=True,
+                repeat_interleave_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            freqs_cos.append(freq_cos)
+            freqs_sin.append(freq_sin)
+
+        self.register_buffer("freqs_cos", torch.cat(freqs_cos, dim=1), persistent=False)
+        self.register_buffer("freqs_sin", torch.cat(freqs_sin, dim=1), persistent=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        split_sizes = [
+            self.attention_head_dim - 2 * (self.attention_head_dim // 3),
+            self.attention_head_dim // 3,
+            self.attention_head_dim // 3,
+        ]
+
+        freqs_cos = self.freqs_cos.split(split_sizes, dim=1)
+        freqs_sin = self.freqs_sin.split(split_sizes, dim=1)
+
+        freqs_cos_f = freqs_cos[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_h = freqs_cos[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_w = freqs_cos[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_sin_f = freqs_sin[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_h = freqs_sin[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_w = freqs_sin[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_cos = torch.cat([freqs_cos_f, freqs_cos_h, freqs_cos_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+        freqs_sin = torch.cat([freqs_sin_f, freqs_sin_h, freqs_sin_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+
+        return freqs_cos, freqs_sin
+
+
+# Copied from diffusers.models.transformers.sana_transformer.SanaModulatedNorm
+class SanaModulatedNorm(nn.Module):
+    def __init__(self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim, elementwise_affine=elementwise_affine, eps=eps)
+
+    def forward(
+        self, hidden_states: torch.Tensor, temb: torch.Tensor, scale_shift_table: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.norm(hidden_states)
+        shift, scale = (scale_shift_table[None] + temb[:, None].to(scale_shift_table.device)).chunk(2, dim=1)
+        hidden_states = hidden_states * (1 + scale) + shift
+        return hidden_states
+
+
+class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+
+    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
+
+        guidance_proj = self.guidance_condition_proj(guidance)
+        guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
+        conditioning = timesteps_emb + guidance_emb
+
+        return self.linear(self.silu(conditioning)), conditioning
+
+
+class SanaAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("SanaAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim)
+        key = key.view(batch_size, -1, attn.heads, head_dim)
+        value = value.view(batch_size, -1, attn.heads, head_dim)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class SanaVideoTransformerBlock(nn.Module):
+    r"""
+    Transformer block introduced in [Sana-Video](https://huggingface.co/papers/2509.24695).
+    """
+
+    def __init__(
+        self,
+        dim: int = 2240,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        dropout: float = 0.0,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        attention_bias: bool = True,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        attention_out_bias: bool = True,
+        mlp_ratio: float = 3.0,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        # 1. Self Attention
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            kv_heads=num_attention_heads if qk_norm is not None else None,
+            qk_norm=qk_norm,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            processor=SanaLinearAttnProcessor3_0(),
+        )
+
+        # 2. Cross Attention
+        if cross_attention_dim is not None:
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            self.attn2 = Attention(
+                query_dim=dim,
+                qk_norm=qk_norm,
+                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_cross_attention_heads,
+                dim_head=cross_attention_head_dim,
+                dropout=dropout,
+                bias=True,
+                out_bias=attention_out_bias,
+                processor=SanaAttnProcessor2_0(),
+            )
+
+        # 3. Feed-forward
+        self.ff = GLUMBTempConv(dim, dim, mlp_ratio, norm_type=None, residual_connection=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        frames: int = None,
+        height: int = None,
+        width: int = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size = hidden_states.shape[0]
+
+        # 1. Modulation
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+        ).chunk(6, dim=1)
+
+        # 2. Self Attention
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.to(hidden_states.dtype)
+
+        attn_output = self.attn1(norm_hidden_states, rotary_emb=rotary_emb)
+        hidden_states = hidden_states + gate_msa * attn_output
+
+        # 3. Cross Attention
+        if self.attn2 is not None:
+            attn_output = self.attn2(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        norm_hidden_states = norm_hidden_states.unflatten(1, (frames, height, width))
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = ff_output.flatten(1, 3)
+        hidden_states = hidden_states + gate_mlp * ff_output
+
+        return hidden_states
+
+
+class SanaVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, AttentionMixin):
+    r"""
+    A 3D Transformer model introduced in [Sana-Video](https://huggingface.co/papers/2509.24695) family of models.
+
+    Args:
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `16`):
+            The number of channels in the output.
+        num_attention_heads (`int`, defaults to `20`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `112`):
+            The number of channels in each head.
+        num_layers (`int`, defaults to `20`):
+            The number of layers of Transformer blocks to use.
+        num_cross_attention_heads (`int`, *optional*, defaults to `20`):
+            The number of heads to use for cross-attention.
+        cross_attention_head_dim (`int`, *optional*, defaults to `112`):
+            The number of channels in each head for cross-attention.
+        cross_attention_dim (`int`, *optional*, defaults to `2240`):
+            The number of channels in the cross-attention output.
+        caption_channels (`int`, defaults to `2304`):
+            The number of channels in the caption embeddings.
+        mlp_ratio (`float`, defaults to `2.5`):
+            The expansion ratio to use in the GLUMBConv layer.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability.
+        attention_bias (`bool`, defaults to `False`):
+            Whether to use bias in the attention layer.
+        sample_size (`int`, defaults to `32`):
+            The base size of the input latent.
+        patch_size (`int`, defaults to `1`):
+            The size of the patches to use in the patch embedding layer.
+        norm_elementwise_affine (`bool`, defaults to `False`):
+            Whether to use elementwise affinity in the normalization layer.
+        norm_eps (`float`, defaults to `1e-6`):
+            The epsilon value for the normalization layer.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for the query and key.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaVideoTransformerBlock", "SanaModulatedNorm"]
+    _skip_layerwise_casting_patterns = ["patch_embedding", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: Optional[int] = 16,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        num_layers: int = 20,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 30,
+        patch_size: Tuple[int, int, int] = (1, 2, 2),
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+        guidance_embeds: bool = False,
+        guidance_embeds_scale: float = 0.1,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(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. Additional condition embeddings
+        if guidance_embeds:
+            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
+        else:
+            self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaVideoTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output blocks
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.norm_out = SanaModulatedNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(inner_dim, math.prod(patch_size) * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples: Optional[Tuple[torch.Tensor]] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        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."
+                )
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        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)
+
+        if guidance is not None:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, guidance=guidance, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            timestep, embedded_timestep = self.time_embed(
+                timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for index_block, block in enumerate(self.transformer_blocks):
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                )
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
+
+        else:
+            for index_block, block in enumerate(self.transformer_blocks):
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                )
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
+
+        # 3. Normalization
+        hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)
+
+        hidden_states = self.proj_out(hidden_states)
+
+        # 5. Unpatchify
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_wan.py

@@ -555,6 +555,9 @@ class WanTransformer3DModel(
             "encoder_hidden_states": ContextParallelInput(split_dim=1, expected_dims=3, split_output=False),
         },
         "proj_out": ContextParallelOutput(gather_dim=1, expected_dims=3),
+        "": {
+            "timestep": ContextParallelInput(split_dim=1, expected_dims=2, split_output=False),
+        },
     }
 
     @register_to_config

src/diffusers/pipelines/__init__.py

@@ -308,6 +308,7 @@ else:
         "SanaSprintPipeline",
         "SanaControlNetPipeline",
         "SanaSprintImg2ImgPipeline",
+        "SanaVideoPipeline",
     ]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
@@ -403,6 +404,7 @@ else:
         "QwenImageControlNetInpaintPipeline",
         "QwenImageControlNetPipeline",
     ]
+    _import_structure["chronoedit"] = ["ChronoEditPipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -565,6 +567,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .bria import BriaPipeline
         from .bria_fibo import BriaFiboPipeline
         from .chroma import ChromaImg2ImgPipeline, ChromaPipeline
+        from .chronoedit import ChronoEditPipeline
         from .cogvideo import (
             CogVideoXFunControlPipeline,
             CogVideoXImageToVideoPipeline,
@@ -735,7 +738,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageInpaintPipeline,
             QwenImagePipeline,
         )
-        from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline
+        from .sana import (
+            SanaControlNetPipeline,
+            SanaPipeline,
+            SanaSprintImg2ImgPipeline,
+            SanaSprintPipeline,
+            SanaVideoPipeline,
+        )
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

src/diffusers/pipelines/chronoedit/__init__.py

@@ -0,0 +1,47 @@
+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_chronoedit"] = ["ChronoEditPipeline"]
+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_chronoedit import ChronoEditPipeline
+
+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)

src/diffusers/pipelines/chronoedit/pipeline_chronoedit.py

@@ -0,0 +1,752 @@
+# Copyright 2025 The ChronoEdit Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL
+import regex as re
+import torch
+from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...loaders import WanLoraLoaderMixin
+from ...models import AutoencoderKLWan, ChronoEditTransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+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 ChronoEditPipelineOutput
+
+
+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:
+        ```python
+        >>> import torch
+        >>> import numpy as np
+        >>> from diffusers import AutoencoderKLWan, ChronoEditTransformer3DModel, ChronoEditPipeline
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from transformers import CLIPVisionModel
+
+        >>> # Available models: nvidia/ChronoEdit-14B-Diffusers
+        >>> model_id = "nvidia/ChronoEdit-14B-Diffusers"
+        >>> image_encoder = CLIPVisionModel.from_pretrained(
+        ...     model_id, subfolder="image_encoder", torch_dtype=torch.float32
+        ... )
+        >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+        >>> transformer = ChronoEditTransformer3DModel.from_pretrained(
+        ...     model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe = ChronoEditPipeline.from_pretrained(
+        ...     model_id, vae=vae, image_encoder=image_encoder, transformer=transformer, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> image = load_image("https://huggingface.co/spaces/nvidia/ChronoEdit/resolve/main/examples/3.png")
+        >>> max_area = 720 * 1280
+        >>> aspect_ratio = image.height / image.width
+        >>> mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+        >>> height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+        >>> width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+        >>> image = image.resize((width, height))
+        >>> prompt = (
+        ...     "The user wants to transform the image by adding a small, cute mouse sitting inside the floral teacup, enjoying a spa bath. The mouse should appear relaxed and cheerful, with a tiny white bath towel draped over its head like a turban. It should be positioned comfortably in the cup’s liquid, with gentle steam rising around it to blend with the cozy atmosphere. "
+        ...     "The mouse’s pose should be natural—perhaps sitting upright with paws resting lightly on the rim or submerged in the tea. The teacup’s floral design, gold trim, and warm lighting must remain unchanged to preserve the original aesthetic. The steam should softly swirl around the mouse, enhancing the spa-like, whimsical mood."
+        ... )
+
+        >>> output = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     height=height,
+        ...     width=width,
+        ...     num_frames=5,
+        ...     guidance_scale=5.0,
+        ...     enable_temporal_reasoning=False,
+        ...     num_temporal_reasoning_steps=0,
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=16)
+        ```
+"""
+
+
+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 ChronoEditPipeline(DiffusionPipeline, WanLoraLoaderMixin):
+    r"""
+    Pipeline for image-to-video generation using Wan.
+
+    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 ([`CLIPVisionModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModel), specifically
+            the
+            [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large)
+            variant.
+        transformer ([`WanTransformer3DModel`]):
+            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: CLIPVisionModel,
+        image_processor: CLIPImageProcessor,
+        transformer: ChronoEditTransformer3DModel,
+        vae: AutoencoderKLWan,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+    ):
+        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 = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial 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.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
+
+    # modified 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)}")
+
+    # modified 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,
+    ) -> 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)  # [batch_size, channels, 1, height, width]
+        video_condition = torch.cat(
+            [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], 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)
+        mask_lat_size[:, :, list(range(1, num_frames))] = 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 = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        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,
+        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,
+        enable_temporal_reasoning: bool = False,
+        num_temporal_reasoning_steps: int = 0,
+    ):
+        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 `480`):
+                The height of the generated video.
+            width (`int`, defaults to `832`):
+                The width of the generated video.
+            num_frames (`int`, defaults to `81`):
+                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 [`ChronoEditPipelineOutput`] 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`, defaults to `512`):
+                The maximum sequence length of the text encoder. If the prompt is longer than this, it will be
+                truncated. If the prompt is shorter, it will be padded to this length.
+            enable_temporal_reasoning (`bool`, *optional*, defaults to `False`):
+                Whether to enable temporal reasoning.
+            num_temporal_reasoning_steps (`int`, *optional*, defaults to `0`):
+                The number of steps to enable temporal reasoning.
+
+        Examples:
+
+        Returns:
+            [`~ChronoEditPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`ChronoEditPipelineOutput`] 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,
+        )
+
+        num_frames = 5 if not enable_temporal_reasoning else num_frames
+
+        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:
+            image_embeds = self.encode_image(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)
+        latents, condition = self.prepare_latents(
+            image,
+            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
+
+                if enable_temporal_reasoning and i == num_temporal_reasoning_steps:
+                    latents = latents[:, :, [0, -1]]
+                    condition = condition[:, :, [0, -1]]
+
+                    for j in range(len(self.scheduler.model_outputs)):
+                        if self.scheduler.model_outputs[j] is not None:
+                            if latents.shape[-3] != self.scheduler.model_outputs[j].shape[-3]:
+                                self.scheduler.model_outputs[j] = self.scheduler.model_outputs[j][:, :, [0, -1]]
+                    if self.scheduler.last_sample is not None:
+                        self.scheduler.last_sample = self.scheduler.last_sample[:, :, [0, -1]]
+
+                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
+            if enable_temporal_reasoning and latents.shape[2] > 2:
+                video_edit = self.vae.decode(latents[:, :, [0, -1]], return_dict=False)[0]
+                video_reason = self.vae.decode(latents[:, :, :-1], return_dict=False)[0]
+                video = torch.cat([video_reason, video_edit[:, :, 1:]], dim=2)
+            else:
+                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 ChronoEditPipelineOutput(frames=video)

src/diffusers/pipelines/chronoedit/pipeline_output.py

@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+import torch
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class ChronoEditPipelineOutput(BaseOutput):
+    r"""
+    Output class for ChronoEdit 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

src/diffusers/pipelines/sana/__init__.py

@@ -26,6 +26,7 @@ else:
     _import_structure["pipeline_sana_controlnet"] = ["SanaControlNetPipeline"]
     _import_structure["pipeline_sana_sprint"] = ["SanaSprintPipeline"]
     _import_structure["pipeline_sana_sprint_img2img"] = ["SanaSprintImg2ImgPipeline"]
+    _import_structure["pipeline_sana_video"] = ["SanaVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -39,6 +40,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_sana_controlnet import SanaControlNetPipeline
         from .pipeline_sana_sprint import SanaSprintPipeline
         from .pipeline_sana_sprint_img2img import SanaSprintImg2ImgPipeline
+        from .pipeline_sana_video import SanaVideoPipeline
 else:
     import sys
 

src/diffusers/pipelines/sana/pipeline_output.py

@@ -3,6 +3,7 @@ from typing import List, Union
 
 import numpy as np
 import PIL.Image
+import torch
 
 from ...utils import BaseOutput
 
@@ -19,3 +20,18 @@ class SanaPipelineOutput(BaseOutput):
     """
 
     images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+@dataclass
+class SanaVideoPipelineOutput(BaseOutput):
+    r"""
+    Output class for Sana-Video 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

src/diffusers/pipelines/sana/pipeline_sana.py

@@ -1,4 +1,4 @@
-# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved.
+# Copyright 2025 SANA Authors 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.

src/diffusers/pipelines/sana/pipeline_sana_sprint.py

@@ -1,4 +1,4 @@
-# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved.
+# Copyright 2025 SANA-Sprint Authors 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.

src/diffusers/utils/dummy_pt_objects.py

@@ -648,6 +648,21 @@ class ChromaTransformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class ChronoEditTransformer3DModel(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 CogVideoXTransformer3DModel(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -1308,6 +1323,21 @@ class SanaTransformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class SanaVideoTransformer3DModel(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 SD3ControlNetModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -542,6 +542,21 @@ class ChromaPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class ChronoEditPipeline(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 CLIPImageProjection(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -2177,6 +2192,21 @@ class SanaSprintPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class SanaVideoPipeline(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 SemanticStableDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

src/diffusers/video_processor.py

@@ -13,11 +13,12 @@
 # limitations under the License.
 
 import warnings
-from typing import List, Optional, Union
+from typing import List, Optional, Tuple, Union
 
 import numpy as np
 import PIL
 import torch
+import torch.nn.functional as F
 
 from .image_processor import VaeImageProcessor, is_valid_image, is_valid_image_imagelist
 
@@ -111,3 +112,65 @@ class VideoProcessor(VaeImageProcessor):
             raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
 
         return outputs
+
+    @staticmethod
+    def classify_height_width_bin(height: int, width: int, ratios: dict) -> Tuple[int, int]:
+        r"""
+        Returns the binned height and width based on the aspect ratio.
+
+        Args:
+            height (`int`): The height of the image.
+            width (`int`): The width of the image.
+            ratios (`dict`): A dictionary where keys are aspect ratios and values are tuples of (height, width).
+
+        Returns:
+            `Tuple[int, int]`: The closest binned height and width.
+        """
+        ar = float(height / width)
+        closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar))
+        default_hw = ratios[closest_ratio]
+        return int(default_hw[0]), int(default_hw[1])
+
+    @staticmethod
+    def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor:
+        r"""
+        Resizes and crops a tensor of videos to the specified dimensions.
+
+        Args:
+            samples (`torch.Tensor`):
+                A tensor of shape (N, C, T, H, W) where N is the batch size, C is the number of channels, T is the
+                number of frames, H is the height, and W is the width.
+            new_width (`int`): The desired width of the output videos.
+            new_height (`int`): The desired height of the output videos.
+
+        Returns:
+            `torch.Tensor`: A tensor containing the resized and cropped videos.
+        """
+        orig_height, orig_width = samples.shape[3], samples.shape[4]
+
+        # Check if resizing is needed
+        if orig_height != new_height or orig_width != new_width:
+            ratio = max(new_height / orig_height, new_width / orig_width)
+            resized_width = int(orig_width * ratio)
+            resized_height = int(orig_height * ratio)
+
+            # Reshape to (N*T, C, H, W) for interpolation
+            n, c, t, h, w = samples.shape
+            samples = samples.permute(0, 2, 1, 3, 4).reshape(n * t, c, h, w)
+
+            # Resize
+            samples = F.interpolate(
+                samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+            )
+
+            # Center Crop
+            start_x = (resized_width - new_width) // 2
+            end_x = start_x + new_width
+            start_y = (resized_height - new_height) // 2
+            end_y = start_y + new_height
+            samples = samples[:, :, start_y:end_y, start_x:end_x]
+
+            # Reshape back to (N, C, T, H, W)
+            samples = samples.reshape(n, t, c, new_height, new_width).permute(0, 2, 1, 3, 4)
+
+        return samples

tests/models/transformers/test_models_transformer_sana_video.py

@@ -0,0 +1,97 @@
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import torch
+
+from diffusers import SanaVideoTransformer3DModel
+
+from ...testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+
+
+enable_full_determinism()
+
+
+class SanaVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
+    model_class = SanaVideoTransformer3DModel
+    main_input_name = "hidden_states"
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 1
+        num_channels = 16
+        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 (16, 2, 16, 16)
+
+    @property
+    def output_shape(self):
+        return (16, 2, 16, 16)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "in_channels": 16,
+            "out_channels": 16,
+            "num_attention_heads": 2,
+            "attention_head_dim": 12,
+            "num_layers": 2,
+            "num_cross_attention_heads": 2,
+            "cross_attention_head_dim": 12,
+            "cross_attention_dim": 24,
+            "caption_channels": 16,
+            "mlp_ratio": 2.5,
+            "dropout": 0.0,
+            "attention_bias": False,
+            "sample_size": 8,
+            "patch_size": (1, 2, 2),
+            "norm_elementwise_affine": False,
+            "norm_eps": 1e-6,
+            "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 = {"SanaVideoTransformer3DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+
+class SanaVideoTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
+    model_class = SanaVideoTransformer3DModel
+
+    def prepare_init_args_and_inputs_for_common(self):
+        return SanaVideoTransformer3DTests().prepare_init_args_and_inputs_for_common()

tests/modular_pipelines/flux/test_modular_pipeline_flux.py

@@ -15,7 +15,6 @@
 
 import random
 import tempfile
-import unittest
 
 import numpy as np
 import PIL
@@ -34,21 +33,16 @@ from ...testing_utils import floats_tensor, torch_device
 from ..test_modular_pipelines_common import ModularPipelineTesterMixin
 
 
-class FluxModularTests:
+class TestFluxModularPipelineFast(ModularPipelineTesterMixin):
     pipeline_class = FluxModularPipeline
     pipeline_blocks_class = FluxAutoBlocks
     repo = "hf-internal-testing/tiny-flux-modular"
 
-    def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
-        pipeline = self.pipeline_blocks_class().init_pipeline(self.repo, components_manager=components_manager)
-        pipeline.load_components(torch_dtype=torch_dtype)
-        return pipeline
+    params = frozenset(["prompt", "height", "width", "guidance_scale"])
+    batch_params = frozenset(["prompt"])
 
-    def get_dummy_inputs(self, device, seed=0):
-        if str(device).startswith("mps"):
-            generator = torch.manual_seed(seed)
-        else:
-            generator = torch.Generator(device=device).manual_seed(seed)
+    def get_dummy_inputs(self, seed=0):
+        generator = self.get_generator(seed)
         inputs = {
             "prompt": "A painting of a squirrel eating a burger",
             "generator": generator,
@@ -57,36 +51,47 @@ class FluxModularTests:
             "height": 8,
             "width": 8,
             "max_sequence_length": 48,
-            "output_type": "np",
+            "output_type": "pt",
         }
         return inputs
 
 
-class FluxModularPipelineFastTests(FluxModularTests, ModularPipelineTesterMixin, unittest.TestCase):
-    params = frozenset(["prompt", "height", "width", "guidance_scale"])
-    batch_params = frozenset(["prompt"])
+class TestFluxImg2ImgModularPipelineFast(ModularPipelineTesterMixin):
+    pipeline_class = FluxModularPipeline
+    pipeline_blocks_class = FluxAutoBlocks
+    repo = "hf-internal-testing/tiny-flux-modular"
 
-
-class FluxImg2ImgModularPipelineFastTests(FluxModularTests, ModularPipelineTesterMixin, unittest.TestCase):
     params = frozenset(["prompt", "height", "width", "guidance_scale", "image"])
     batch_params = frozenset(["prompt", "image"])
 
     def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
         pipeline = super().get_pipeline(components_manager, torch_dtype)
+
         # Override `vae_scale_factor` here as currently, `image_processor` is initialized with
         # fixed constants instead of
         # https://github.com/huggingface/diffusers/blob/d54622c2679d700b425ad61abce9b80fc36212c0/src/diffusers/pipelines/flux/pipeline_flux_img2img.py#L230C9-L232C10
         pipeline.image_processor = VaeImageProcessor(vae_scale_factor=2)
         return pipeline
 
-    def get_dummy_inputs(self, device, seed=0):
-        inputs = super().get_dummy_inputs(device, seed)
-        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
-        image = image / 2 + 0.5
-        inputs["image"] = image
-        inputs["strength"] = 0.8
-        inputs["height"] = 8
-        inputs["width"] = 8
+    def get_dummy_inputs(self, seed=0):
+        generator = self.get_generator(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 4,
+            "guidance_scale": 5.0,
+            "height": 8,
+            "width": 8,
+            "max_sequence_length": 48,
+            "output_type": "pt",
+        }
+        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(torch_device)
+        image = image.cpu().permute(0, 2, 3, 1)[0]
+        init_image = PIL.Image.fromarray(np.uint8(image)).convert("RGB")
+
+        inputs["image"] = init_image
+        inputs["strength"] = 0.5
+
         return inputs
 
     def test_save_from_pretrained(self):
@@ -96,6 +101,7 @@ class FluxImg2ImgModularPipelineFastTests(FluxModularTests, ModularPipelineTeste
 
         with tempfile.TemporaryDirectory() as tmpdirname:
             base_pipe.save_pretrained(tmpdirname)
+
             pipe = ModularPipeline.from_pretrained(tmpdirname).to(torch_device)
             pipe.load_components(torch_dtype=torch.float32)
             pipe.to(torch_device)
@@ -105,26 +111,62 @@ class FluxImg2ImgModularPipelineFastTests(FluxModularTests, ModularPipelineTeste
 
         image_slices = []
         for pipe in pipes:
-            inputs = self.get_dummy_inputs(torch_device)
+            inputs = self.get_dummy_inputs()
             image = pipe(**inputs, output="images")
 
             image_slices.append(image[0, -3:, -3:, -1].flatten())
 
-        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert torch.abs(image_slices[0] - image_slices[1]).max() < 1e-3
 
 
-class FluxKontextModularPipelineFastTests(FluxImg2ImgModularPipelineFastTests):
+class TestFluxKontextModularPipelineFast(ModularPipelineTesterMixin):
     pipeline_class = FluxKontextModularPipeline
     pipeline_blocks_class = FluxKontextAutoBlocks
     repo = "hf-internal-testing/tiny-flux-kontext-pipe"
 
-    def get_dummy_inputs(self, device, seed=0):
-        inputs = super().get_dummy_inputs(device, seed)
+    params = frozenset(["prompt", "height", "width", "guidance_scale", "image"])
+    batch_params = frozenset(["prompt", "image"])
+
+    def get_dummy_inputs(self, seed=0):
+        generator = self.get_generator(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 8,
+            "width": 8,
+            "max_sequence_length": 48,
+            "output_type": "pt",
+        }
         image = PIL.Image.new("RGB", (32, 32), 0)
-        _ = inputs.pop("strength")
+
         inputs["image"] = image
-        inputs["height"] = 8
-        inputs["width"] = 8
-        inputs["max_area"] = 8 * 8
+        inputs["max_area"] = inputs["height"] * inputs["width"]
         inputs["_auto_resize"] = False
+
         return inputs
+
+    def test_save_from_pretrained(self):
+        pipes = []
+        base_pipe = self.get_pipeline().to(torch_device)
+        pipes.append(base_pipe)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            base_pipe.save_pretrained(tmpdirname)
+
+            pipe = ModularPipeline.from_pretrained(tmpdirname).to(torch_device)
+            pipe.load_components(torch_dtype=torch.float32)
+            pipe.to(torch_device)
+            pipe.image_processor = VaeImageProcessor(vae_scale_factor=2)
+
+        pipes.append(pipe)
+
+        image_slices = []
+        for pipe in pipes:
+            inputs = self.get_dummy_inputs()
+            image = pipe(**inputs, output="images")
+
+            image_slices.append(image[0, -3:, -3:, -1].flatten())
+
+        assert torch.abs(image_slices[0] - image_slices[1]).max() < 1e-3

tests/modular_pipelines/stable_diffusion_xl/test_modular_pipeline_stable_diffusion_xl.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 
 import random
-import unittest
 from typing import Any, Dict
 
 import numpy as np
@@ -32,63 +31,26 @@ from ..test_modular_pipelines_common import ModularPipelineTesterMixin
 enable_full_determinism()
 
 
-class SDXLModularTests:
+class SDXLModularTesterMixin:
     """
     This mixin defines method to create pipeline, base input and base test across all SDXL modular tests.
     """
 
-    pipeline_class = StableDiffusionXLModularPipeline
-    pipeline_blocks_class = StableDiffusionXLAutoBlocks
-    repo = "hf-internal-testing/tiny-sdxl-modular"
-    params = frozenset(
-        [
-            "prompt",
-            "height",
-            "width",
-            "negative_prompt",
-            "cross_attention_kwargs",
-            "image",
-            "mask_image",
-        ]
-    )
-    batch_params = frozenset(["prompt", "negative_prompt", "image", "mask_image"])
-
-    def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
-        pipeline = self.pipeline_blocks_class().init_pipeline(self.repo, components_manager=components_manager)
-        pipeline.load_components(torch_dtype=torch_dtype)
-        return pipeline
-
-    def get_dummy_inputs(self, device, seed=0):
-        if str(device).startswith("mps"):
-            generator = torch.manual_seed(seed)
-        else:
-            generator = torch.Generator(device=device).manual_seed(seed)
-        inputs = {
-            "prompt": "A painting of a squirrel eating a burger",
-            "generator": generator,
-            "num_inference_steps": 2,
-            "output_type": "np",
-        }
-        return inputs
-
     def _test_stable_diffusion_xl_euler(self, expected_image_shape, expected_slice, expected_max_diff=1e-2):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         sd_pipe = self.get_pipeline()
-        sd_pipe = sd_pipe.to(device)
+        sd_pipe = sd_pipe.to(torch_device)
         sd_pipe.set_progress_bar_config(disable=None)
 
-        inputs = self.get_dummy_inputs(device)
+        inputs = self.get_dummy_inputs()
         image = sd_pipe(**inputs, output="images")
         image_slice = image[0, -3:, -3:, -1]
 
         assert image.shape == expected_image_shape
-
-        assert np.abs(image_slice.flatten() - expected_slice).max() < expected_max_diff, (
-            "Image Slice does not match expected slice"
-        )
+        max_diff = torch.abs(image_slice.flatten() - expected_slice).max()
+        assert max_diff < expected_max_diff, f"Image slice does not match expected slice. Max Difference: {max_diff}"
 
 
-class SDXLModularIPAdapterTests:
+class SDXLModularIPAdapterTesterMixin:
     """
     This mixin is designed to test IP Adapter.
     """
@@ -127,7 +89,7 @@ class SDXLModularIPAdapterTests:
         if "image" in parameters and "strength" in parameters:
             inputs["num_inference_steps"] = 4
 
-        inputs["output_type"] = "np"
+        inputs["output_type"] = "pt"
         return inputs
 
     def test_ip_adapter(self, expected_max_diff: float = 1e-4, expected_pipe_slice=None):
@@ -152,7 +114,7 @@ class SDXLModularIPAdapterTests:
         cross_attention_dim = pipe.unet.config.get("cross_attention_dim")
 
         # forward pass without ip adapter
-        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs())
         if expected_pipe_slice is None:
             output_without_adapter = pipe(**inputs, output="images")
         else:
@@ -163,7 +125,7 @@ class SDXLModularIPAdapterTests:
         pipe.unet._load_ip_adapter_weights(adapter_state_dict)
 
         # forward pass with single ip adapter, but scale=0 which should have no effect
-        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs())
         inputs["ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         inputs["negative_ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         pipe.set_ip_adapter_scale(0.0)
@@ -172,7 +134,7 @@ class SDXLModularIPAdapterTests:
             output_without_adapter_scale = output_without_adapter_scale[0, -3:, -3:, -1].flatten()
 
         # forward pass with single ip adapter, but with scale of adapter weights
-        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs())
         inputs["ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         inputs["negative_ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         pipe.set_ip_adapter_scale(42.0)
@@ -180,8 +142,8 @@ class SDXLModularIPAdapterTests:
         if expected_pipe_slice is not None:
             output_with_adapter_scale = output_with_adapter_scale[0, -3:, -3:, -1].flatten()
 
-        max_diff_without_adapter_scale = np.abs(output_without_adapter_scale - output_without_adapter).max()
-        max_diff_with_adapter_scale = np.abs(output_with_adapter_scale - output_without_adapter).max()
+        max_diff_without_adapter_scale = torch.abs(output_without_adapter_scale - output_without_adapter).max()
+        max_diff_with_adapter_scale = torch.abs(output_with_adapter_scale - output_without_adapter).max()
 
         assert max_diff_without_adapter_scale < expected_max_diff, (
             "Output without ip-adapter must be same as normal inference"
@@ -194,7 +156,7 @@ class SDXLModularIPAdapterTests:
         pipe.unet._load_ip_adapter_weights([adapter_state_dict_1, adapter_state_dict_2])
 
         # forward pass with multi ip adapter, but scale=0 which should have no effect
-        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs())
         inputs["ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)] * 2
         inputs["negative_ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)] * 2
         pipe.set_ip_adapter_scale([0.0, 0.0])
@@ -203,7 +165,7 @@ class SDXLModularIPAdapterTests:
             output_without_multi_adapter_scale = output_without_multi_adapter_scale[0, -3:, -3:, -1].flatten()
 
         # forward pass with multi ip adapter, but with scale of adapter weights
-        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs())
         inputs["ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)] * 2
         inputs["negative_ip_adapter_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)] * 2
         pipe.set_ip_adapter_scale([42.0, 42.0])
@@ -211,10 +173,10 @@ class SDXLModularIPAdapterTests:
         if expected_pipe_slice is not None:
             output_with_multi_adapter_scale = output_with_multi_adapter_scale[0, -3:, -3:, -1].flatten()
 
-        max_diff_without_multi_adapter_scale = np.abs(
+        max_diff_without_multi_adapter_scale = torch.abs(
             output_without_multi_adapter_scale - output_without_adapter
         ).max()
-        max_diff_with_multi_adapter_scale = np.abs(output_with_multi_adapter_scale - output_without_adapter).max()
+        max_diff_with_multi_adapter_scale = torch.abs(output_with_multi_adapter_scale - output_without_adapter).max()
         assert max_diff_without_multi_adapter_scale < expected_max_diff, (
             "Output without multi-ip-adapter must be same as normal inference"
         )
@@ -223,7 +185,7 @@ class SDXLModularIPAdapterTests:
         )
 
 
-class SDXLModularControlNetTests:
+class SDXLModularControlNetTesterMixin:
     """
     This mixin is designed to test ControlNet.
     """
@@ -262,24 +224,26 @@ class SDXLModularControlNetTests:
         pipe.set_progress_bar_config(disable=None)
 
         # forward pass without controlnet
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         output_without_controlnet = pipe(**inputs, output="images")
         output_without_controlnet = output_without_controlnet[0, -3:, -3:, -1].flatten()
 
         # forward pass with single controlnet, but scale=0 which should have no effect
-        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs())
         inputs["controlnet_conditioning_scale"] = 0.0
         output_without_controlnet_scale = pipe(**inputs, output="images")
         output_without_controlnet_scale = output_without_controlnet_scale[0, -3:, -3:, -1].flatten()
 
         # forward pass with single controlnet, but with scale of adapter weights
-        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs())
         inputs["controlnet_conditioning_scale"] = 42.0
         output_with_controlnet_scale = pipe(**inputs, output="images")
         output_with_controlnet_scale = output_with_controlnet_scale[0, -3:, -3:, -1].flatten()
 
-        max_diff_without_controlnet_scale = np.abs(output_without_controlnet_scale - output_without_controlnet).max()
-        max_diff_with_controlnet_scale = np.abs(output_with_controlnet_scale - output_without_controlnet).max()
+        max_diff_without_controlnet_scale = torch.abs(
+            output_without_controlnet_scale - output_without_controlnet
+        ).max()
+        max_diff_with_controlnet_scale = torch.abs(output_with_controlnet_scale - output_without_controlnet).max()
 
         assert max_diff_without_controlnet_scale < expected_max_diff, (
             "Output without controlnet must be same as normal inference"
@@ -295,21 +259,21 @@ class SDXLModularControlNetTests:
         guider = ClassifierFreeGuidance(guidance_scale=1.0)
         pipe.update_components(guider=guider)
 
-        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs())
         out_no_cfg = pipe(**inputs, output="images")
 
         # forward pass with CFG applied
         guider = ClassifierFreeGuidance(guidance_scale=7.5)
         pipe.update_components(guider=guider)
-        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs(torch_device))
+        inputs = self._modify_inputs_for_controlnet_test(self.get_dummy_inputs())
         out_cfg = pipe(**inputs, output="images")
 
         assert out_cfg.shape == out_no_cfg.shape
-        max_diff = np.abs(out_cfg - out_no_cfg).max()
+        max_diff = torch.abs(out_cfg - out_no_cfg).max()
         assert max_diff > 1e-2, "Output with CFG must be different from normal inference"
 
 
-class SDXLModularGuiderTests:
+class SDXLModularGuiderTesterMixin:
     def test_guider_cfg(self):
         pipe = self.get_pipeline()
         pipe = pipe.to(torch_device)
@@ -319,13 +283,13 @@ class SDXLModularGuiderTests:
         guider = ClassifierFreeGuidance(guidance_scale=1.0)
         pipe.update_components(guider=guider)
 
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         out_no_cfg = pipe(**inputs, output="images")
 
         # forward pass with CFG applied
         guider = ClassifierFreeGuidance(guidance_scale=7.5)
         pipe.update_components(guider=guider)
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         out_cfg = pipe(**inputs, output="images")
 
         assert out_cfg.shape == out_no_cfg.shape
@@ -333,30 +297,57 @@ class SDXLModularGuiderTests:
         assert max_diff > 1e-2, "Output with CFG must be different from normal inference"
 
 
-class SDXLModularPipelineFastTests(
-    SDXLModularTests,
-    SDXLModularIPAdapterTests,
-    SDXLModularControlNetTests,
-    SDXLModularGuiderTests,
+class TestSDXLModularPipelineFast(
+    SDXLModularTesterMixin,
+    SDXLModularIPAdapterTesterMixin,
+    SDXLModularControlNetTesterMixin,
+    SDXLModularGuiderTesterMixin,
     ModularPipelineTesterMixin,
-    unittest.TestCase,
 ):
     """Test cases for Stable Diffusion XL modular pipeline fast tests."""
 
+    pipeline_class = StableDiffusionXLModularPipeline
+    pipeline_blocks_class = StableDiffusionXLAutoBlocks
+    repo = "hf-internal-testing/tiny-sdxl-modular"
+    params = frozenset(
+        [
+            "prompt",
+            "height",
+            "width",
+            "negative_prompt",
+            "cross_attention_kwargs",
+        ]
+    )
+    batch_params = frozenset(["prompt", "negative_prompt"])
+    expected_image_output_shape = (1, 3, 64, 64)
+
+    def get_dummy_inputs(self, seed=0):
+        generator = self.get_generator(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "output_type": "pt",
+        }
+        return inputs
+
     def test_stable_diffusion_xl_euler(self):
         self._test_stable_diffusion_xl_euler(
-            expected_image_shape=(1, 64, 64, 3),
-            expected_slice=[
-                0.5966781,
-                0.62939394,
-                0.48465094,
-                0.51573336,
-                0.57593524,
-                0.47035995,
-                0.53410417,
-                0.51436996,
-                0.47313565,
-            ],
+            expected_image_shape=self.expected_image_output_shape,
+            expected_slice=torch.tensor(
+                [
+                    0.5966781,
+                    0.62939394,
+                    0.48465094,
+                    0.51573336,
+                    0.57593524,
+                    0.47035995,
+                    0.53410417,
+                    0.51436996,
+                    0.47313565,
+                ],
+                device=torch_device,
+            ),
             expected_max_diff=1e-2,
         )
 
@@ -364,39 +355,65 @@ class SDXLModularPipelineFastTests(
         super().test_inference_batch_single_identical(expected_max_diff=3e-3)
 
 
-class SDXLImg2ImgModularPipelineFastTests(
-    SDXLModularTests,
-    SDXLModularIPAdapterTests,
-    SDXLModularControlNetTests,
-    SDXLModularGuiderTests,
+class TestSDXLImg2ImgModularPipelineFast(
+    SDXLModularTesterMixin,
+    SDXLModularIPAdapterTesterMixin,
+    SDXLModularControlNetTesterMixin,
+    SDXLModularGuiderTesterMixin,
     ModularPipelineTesterMixin,
-    unittest.TestCase,
 ):
     """Test cases for Stable Diffusion XL image-to-image modular pipeline fast tests."""
 
-    def get_dummy_inputs(self, device, seed=0):
-        inputs = super().get_dummy_inputs(device, seed)
-        image = floats_tensor((1, 3, 64, 64), rng=random.Random(seed)).to(device)
-        image = image / 2 + 0.5
-        inputs["image"] = image
-        inputs["strength"] = 0.8
+    pipeline_class = StableDiffusionXLModularPipeline
+    pipeline_blocks_class = StableDiffusionXLAutoBlocks
+    repo = "hf-internal-testing/tiny-sdxl-modular"
+    params = frozenset(
+        [
+            "prompt",
+            "height",
+            "width",
+            "negative_prompt",
+            "cross_attention_kwargs",
+            "image",
+        ]
+    )
+    batch_params = frozenset(["prompt", "negative_prompt", "image"])
+    expected_image_output_shape = (1, 3, 64, 64)
+
+    def get_dummy_inputs(self, seed=0):
+        generator = self.get_generator(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 4,
+            "output_type": "pt",
+        }
+        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(torch_device)
+        image = image.cpu().permute(0, 2, 3, 1)[0]
+        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+
+        inputs["image"] = init_image
+        inputs["strength"] = 0.5
 
         return inputs
 
     def test_stable_diffusion_xl_euler(self):
         self._test_stable_diffusion_xl_euler(
-            expected_image_shape=(1, 64, 64, 3),
-            expected_slice=[
-                0.56943184,
-                0.4702148,
-                0.48048905,
-                0.6235963,
-                0.551138,
-                0.49629188,
-                0.60031277,
-                0.5688907,
-                0.43996853,
-            ],
+            expected_image_shape=self.expected_image_output_shape,
+            expected_slice=torch.tensor(
+                [
+                    0.56943184,
+                    0.4702148,
+                    0.48048905,
+                    0.6235963,
+                    0.551138,
+                    0.49629188,
+                    0.60031277,
+                    0.5688907,
+                    0.43996853,
+                ],
+                device=torch_device,
+            ),
             expected_max_diff=1e-2,
         )
 
@@ -405,20 +422,43 @@ class SDXLImg2ImgModularPipelineFastTests(
 
 
 class SDXLInpaintingModularPipelineFastTests(
-    SDXLModularTests,
-    SDXLModularIPAdapterTests,
-    SDXLModularControlNetTests,
-    SDXLModularGuiderTests,
+    SDXLModularTesterMixin,
+    SDXLModularIPAdapterTesterMixin,
+    SDXLModularControlNetTesterMixin,
+    SDXLModularGuiderTesterMixin,
     ModularPipelineTesterMixin,
-    unittest.TestCase,
 ):
     """Test cases for Stable Diffusion XL inpainting modular pipeline fast tests."""
 
+    pipeline_class = StableDiffusionXLModularPipeline
+    pipeline_blocks_class = StableDiffusionXLAutoBlocks
+    repo = "hf-internal-testing/tiny-sdxl-modular"
+    params = frozenset(
+        [
+            "prompt",
+            "height",
+            "width",
+            "negative_prompt",
+            "cross_attention_kwargs",
+            "image",
+            "mask_image",
+        ]
+    )
+    batch_params = frozenset(["prompt", "negative_prompt", "image", "mask_image"])
+    expected_image_output_shape = (1, 3, 64, 64)
+
     def get_dummy_inputs(self, device, seed=0):
-        inputs = super().get_dummy_inputs(device, seed)
+        generator = self.get_generator(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 4,
+            "output_type": "pt",
+        }
         image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
         image = image.cpu().permute(0, 2, 3, 1)[0]
         init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+
         # create mask
         image[8:, 8:, :] = 255
         mask_image = Image.fromarray(np.uint8(image)).convert("L").resize((64, 64))
@@ -431,18 +471,21 @@ class SDXLInpaintingModularPipelineFastTests(
 
     def test_stable_diffusion_xl_euler(self):
         self._test_stable_diffusion_xl_euler(
-            expected_image_shape=(1, 64, 64, 3),
-            expected_slice=[
-                0.40872607,
-                0.38842705,
-                0.34893104,
-                0.47837183,
-                0.43792963,
-                0.5332134,
-                0.3716843,
-                0.47274873,
-                0.45000193,
-            ],
+            expected_image_shape=self.expected_image_output_shape,
+            expected_slice=torch.tensor(
+                [
+                    0.40872607,
+                    0.38842705,
+                    0.34893104,
+                    0.47837183,
+                    0.43792963,
+                    0.5332134,
+                    0.3716843,
+                    0.47274873,
+                    0.45000193,
+                ],
+                device=torch_device,
+            ),
             expected_max_diff=1e-2,
         )
 

tests/modular_pipelines/test_modular_pipelines_common.py

@@ -1,9 +1,7 @@
 import gc
 import tempfile
-import unittest
 from typing import Callable, Union
 
-import numpy as np
 import torch
 
 import diffusers
@@ -19,17 +17,9 @@ from ..testing_utils import (
 )
 
 
-def to_np(tensor):
-    if isinstance(tensor, torch.Tensor):
-        tensor = tensor.detach().cpu().numpy()
-
-    return tensor
-
-
 @require_torch
 class ModularPipelineTesterMixin:
     """
-    This mixin is designed to be used with unittest.TestCase classes.
     It provides a set of common tests for each modular pipeline,
     including:
     - test_pipeline_call_signature: check if the pipeline's __call__ method has all required parameters
@@ -57,9 +47,8 @@ class ModularPipelineTesterMixin:
         ]
     )
 
-    def get_generator(self, seed):
-        device = torch_device if torch_device != "mps" else "cpu"
-        generator = torch.Generator(device).manual_seed(seed)
+    def get_generator(self, seed=0):
+        generator = torch.Generator("cpu").manual_seed(seed)
         return generator
 
     @property
@@ -82,13 +71,7 @@ class ModularPipelineTesterMixin:
             "See existing pipeline tests for reference."
         )
 
-    def get_pipeline(self):
-        raise NotImplementedError(
-            "You need to implement `get_pipeline(self)` in the child test class. "
-            "See existing pipeline tests for reference."
-        )
-
-    def get_dummy_inputs(self, device, seed=0):
+    def get_dummy_inputs(self, seed=0):
         raise NotImplementedError(
             "You need to implement `get_dummy_inputs(self, device, seed)` in the child test class. "
             "See existing pipeline tests for reference."
@@ -123,20 +106,23 @@ class ModularPipelineTesterMixin:
             "See existing pipeline tests for reference."
         )
 
-    def setUp(self):
+    def setup_method(self):
         # clean up the VRAM before each test
-        super().setUp()
         torch.compiler.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
-    def tearDown(self):
+    def teardown_method(self):
         # clean up the VRAM after each test in case of CUDA runtime errors
-        super().tearDown()
         torch.compiler.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
+    def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
+        pipeline = self.pipeline_blocks_class().init_pipeline(self.repo, components_manager=components_manager)
+        pipeline.load_components(torch_dtype=torch_dtype)
+        return pipeline
+
     def test_pipeline_call_signature(self):
         pipe = self.get_pipeline()
         input_parameters = pipe.blocks.input_names
@@ -156,7 +142,7 @@ class ModularPipelineTesterMixin:
         pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
 
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         inputs["generator"] = self.get_generator(0)
 
         logger = logging.get_logger(pipe.__module__)
@@ -196,7 +182,7 @@ class ModularPipelineTesterMixin:
         pipe = self.get_pipeline()
         pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
 
         # Reset generator in case it is has been used in self.get_dummy_inputs
         inputs["generator"] = self.get_generator(0)
@@ -226,10 +212,9 @@ class ModularPipelineTesterMixin:
 
         assert output_batch.shape[0] == batch_size
 
-        max_diff = np.abs(to_np(output_batch[0]) - to_np(output[0])).max()
+        max_diff = torch.abs(output_batch[0] - output[0]).max()
         assert max_diff < expected_max_diff, "Batch inference results different from single inference results"
 
-    @unittest.skipIf(torch_device not in ["cuda", "xpu"], reason="float16 requires CUDA or XPU")
     @require_accelerator
     def test_float16_inference(self, expected_max_diff=5e-2):
         pipe = self.get_pipeline()
@@ -240,13 +225,13 @@ class ModularPipelineTesterMixin:
         pipe_fp16.to(torch_device, torch.float16)
         pipe_fp16.set_progress_bar_config(disable=None)
 
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         # Reset generator in case it is used inside dummy inputs
         if "generator" in inputs:
             inputs["generator"] = self.get_generator(0)
         output = pipe(**inputs, output="images")
 
-        fp16_inputs = self.get_dummy_inputs(torch_device)
+        fp16_inputs = self.get_dummy_inputs()
         # Reset generator in case it is used inside dummy inputs
         if "generator" in fp16_inputs:
             fp16_inputs["generator"] = self.get_generator(0)
@@ -283,8 +268,8 @@ class ModularPipelineTesterMixin:
         pipe.set_progress_bar_config(disable=None)
         pipe.to("cpu")
 
-        output = pipe(**self.get_dummy_inputs("cpu"), output="images")
-        assert np.isnan(to_np(output)).sum() == 0, "CPU Inference returns NaN"
+        output = pipe(**self.get_dummy_inputs(), output="images")
+        assert torch.isnan(output).sum() == 0, "CPU Inference returns NaN"
 
     @require_accelerator
     def test_inference_is_not_nan(self):
@@ -292,8 +277,8 @@ class ModularPipelineTesterMixin:
         pipe.set_progress_bar_config(disable=None)
         pipe.to(torch_device)
 
-        output = pipe(**self.get_dummy_inputs(torch_device), output="images")
-        assert np.isnan(to_np(output)).sum() == 0, "Accelerator Inference returns NaN"
+        output = pipe(**self.get_dummy_inputs(), output="images")
+        assert torch.isnan(output).sum() == 0, "Accelerator Inference returns NaN"
 
     def test_num_images_per_prompt(self):
         pipe = self.get_pipeline()
@@ -309,7 +294,7 @@ class ModularPipelineTesterMixin:
 
         for batch_size in batch_sizes:
             for num_images_per_prompt in num_images_per_prompts:
-                inputs = self.get_dummy_inputs(torch_device)
+                inputs = self.get_dummy_inputs()
 
                 for key in inputs.keys():
                     if key in self.batch_params:
@@ -329,12 +314,12 @@ class ModularPipelineTesterMixin:
 
         image_slices = []
         for pipe in [base_pipe, offload_pipe]:
-            inputs = self.get_dummy_inputs(torch_device)
+            inputs = self.get_dummy_inputs()
             image = pipe(**inputs, output="images")
 
             image_slices.append(image[0, -3:, -3:, -1].flatten())
 
-        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert torch.abs(image_slices[0] - image_slices[1]).max() < 1e-3
 
     def test_save_from_pretrained(self):
         pipes = []
@@ -351,9 +336,9 @@ class ModularPipelineTesterMixin:
 
         image_slices = []
         for pipe in pipes:
-            inputs = self.get_dummy_inputs(torch_device)
+            inputs = self.get_dummy_inputs()
             image = pipe(**inputs, output="images")
 
             image_slices.append(image[0, -3:, -3:, -1].flatten())
 
-        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
+        assert torch.abs(image_slices[0] - image_slices[1]).max() < 1e-3

tests/pipelines/chronoedit/test_chronoedit.py

@@ -0,0 +1,176 @@
+# Copyright 2025 The HuggingFace Team.
+#
+# 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 PIL import Image
+from transformers import (
+    AutoTokenizer,
+    CLIPImageProcessor,
+    CLIPVisionConfig,
+    CLIPVisionModelWithProjection,
+    T5EncoderModel,
+)
+
+from diffusers import (
+    AutoencoderKLWan,
+    ChronoEditPipeline,
+    ChronoEditTransformer3DModel,
+    FlowMatchEulerDiscreteScheduler,
+)
+
+from ...testing_utils import enable_full_determinism
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class ChronoEditPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = ChronoEditPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs", "height", "width"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+    supports_dduf = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        vae = AutoencoderKLWan(
+            base_dim=3,
+            z_dim=16,
+            dim_mult=[1, 1, 1, 1],
+            num_res_blocks=1,
+            temperal_downsample=[False, True, True],
+        )
+
+        torch.manual_seed(0)
+        # TODO: impl FlowDPMSolverMultistepScheduler
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        transformer = ChronoEditTransformer3DModel(
+            patch_size=(1, 2, 2),
+            num_attention_heads=2,
+            attention_head_dim=12,
+            in_channels=36,
+            out_channels=16,
+            text_dim=32,
+            freq_dim=256,
+            ffn_dim=32,
+            num_layers=2,
+            cross_attn_norm=True,
+            qk_norm="rms_norm_across_heads",
+            rope_max_seq_len=32,
+            image_dim=4,
+        )
+
+        torch.manual_seed(0)
+        image_encoder_config = CLIPVisionConfig(
+            hidden_size=4,
+            projection_dim=4,
+            num_hidden_layers=2,
+            num_attention_heads=2,
+            image_size=32,
+            intermediate_size=16,
+            patch_size=1,
+        )
+        image_encoder = CLIPVisionModelWithProjection(image_encoder_config)
+
+        torch.manual_seed(0)
+        image_processor = CLIPImageProcessor(crop_size=32, size=32)
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "image_encoder": image_encoder,
+            "image_processor": image_processor,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        image_height = 16
+        image_width = 16
+        image = Image.new("RGB", (image_width, image_height))
+        inputs = {
+            "image": image,
+            "prompt": "dance monkey",
+            "negative_prompt": "negative",  # TODO
+            "height": image_height,
+            "width": image_width,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "num_frames": 5,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+        self.assertEqual(generated_video.shape, (5, 3, 16, 16))
+
+        # fmt: off
+        expected_slice = torch.tensor([0.4525, 0.4520, 0.4485, 0.4534, 0.4523, 0.4522, 0.4529, 0.4528, 0.5022, 0.5064, 0.5011, 0.5061, 0.5028, 0.4979, 0.5117, 0.5192])
+        # 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))
+
+    @unittest.skip("Test not supported")
+    def test_attention_slicing_forward_pass(self):
+        pass
+
+    @unittest.skip("TODO: revisit failing as it requires a very high threshold to pass")
+    def test_inference_batch_single_identical(self):
+        pass
+
+    @unittest.skip(
+        "ChronoEditPipeline has to run in mixed precision. Save/Load the entire pipeline in FP16 will result in errors"
+    )
+    def test_save_load_float16(self):
+        pass

tests/pipelines/sana/test_sana_video.py

@@ -0,0 +1,225 @@
+# Copyright 2025 The HuggingFace Team.
+#
+# 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 gc
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+from transformers import Gemma2Config, Gemma2Model, GemmaTokenizer
+
+from diffusers import AutoencoderKLWan, DPMSolverMultistepScheduler, SanaVideoPipeline, SanaVideoTransformer3DModel
+
+from ...testing_utils import (
+    backend_empty_cache,
+    enable_full_determinism,
+    require_torch_accelerator,
+    slow,
+    torch_device,
+)
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+class SanaVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = SanaVideoPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+    supports_dduf = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        vae = AutoencoderKLWan(
+            base_dim=3,
+            z_dim=16,
+            dim_mult=[1, 1, 1, 1],
+            num_res_blocks=1,
+            temperal_downsample=[False, True, True],
+        )
+
+        torch.manual_seed(0)
+        scheduler = DPMSolverMultistepScheduler()
+
+        torch.manual_seed(0)
+        text_encoder_config = Gemma2Config(
+            head_dim=16,
+            hidden_size=8,
+            initializer_range=0.02,
+            intermediate_size=64,
+            max_position_embeddings=8192,
+            model_type="gemma2",
+            num_attention_heads=2,
+            num_hidden_layers=1,
+            num_key_value_heads=2,
+            vocab_size=8,
+            attn_implementation="eager",
+        )
+        text_encoder = Gemma2Model(text_encoder_config)
+        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
+
+        torch.manual_seed(0)
+        transformer = SanaVideoTransformer3DModel(
+            in_channels=16,
+            out_channels=16,
+            num_attention_heads=2,
+            attention_head_dim=12,
+            num_layers=2,
+            num_cross_attention_heads=2,
+            cross_attention_head_dim=12,
+            cross_attention_dim=24,
+            caption_channels=8,
+            mlp_ratio=2.5,
+            dropout=0.0,
+            attention_bias=False,
+            sample_size=8,
+            patch_size=(1, 2, 2),
+            norm_elementwise_affine=False,
+            norm_eps=1e-6,
+            qk_norm="rms_norm_across_heads",
+            rope_max_seq_len=32,
+        )
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "height": 32,
+            "width": 32,
+            "frames": 9,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+            "complex_human_instruction": [],
+            "use_resolution_binning": False,
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
+
+    @unittest.skip("Test not supported")
+    def test_attention_slicing_forward_pass(self):
+        pass
+
+    def test_save_load_local(self, expected_max_difference=5e-4):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        torch.manual_seed(0)
+        output = pipe(**inputs)[0]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir, safe_serialization=False)
+            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
+            for component in pipe_loaded.components.values():
+                if hasattr(component, "set_default_attn_processor"):
+                    component.set_default_attn_processor()
+            pipe_loaded.to(torch_device)
+            pipe_loaded.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        torch.manual_seed(0)
+        output_loaded = pipe_loaded(**inputs)[0]
+
+        max_diff = np.abs(output.detach().cpu().numpy() - output_loaded.detach().cpu().numpy()).max()
+        self.assertLess(max_diff, expected_max_difference)
+
+    # TODO(aryan): Create a dummy gemma model with smol vocab size
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_consistent(self):
+        pass
+
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_single_identical(self):
+        pass
+
+    def test_float16_inference(self):
+        # Requires higher tolerance as model seems very sensitive to dtype
+        super().test_float16_inference(expected_max_diff=0.08)
+
+    def test_save_load_float16(self):
+        # Requires higher tolerance as model seems very sensitive to dtype
+        super().test_save_load_float16(expected_max_diff=0.2)
+
+
+@slow
+@require_torch_accelerator
+class SanaVideoPipelineIntegrationTests(unittest.TestCase):
+    prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @unittest.skip("TODO: test needs to be implemented")
+    def test_sana_video_480p(self):
+        pass