feat(wan): Add prior-based diffusion step skip for ~70% fewer inference steps#1325
feat(wan): Add prior-based diffusion step skip for ~70% fewer inference steps#1325Efrat-Taig wants to merge 2 commits intomodelscope:mainfrom
Conversation
…ce steps - Add step_callback, prior_latents, prior_timesteps, prior_sigmas, start_from_step to WanVideoPipeline for resuming from saved latents - Add examples/wanvideo/prior_based_step_skip/ with generate_prior.py, infer_from_prior.py, prior_utils.py, and README - Add --download_example to generate_prior.py for easy onboarding - Add Prior-based step skip section to Wan docs (en + zh) - Supports fixed identity/scene with varying motion (e.g. lip-sync, different actions) Made-with: Cursor
Summary of ChangesHello, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request introduces a significant optimization for Wan video models by enabling a prior-based diffusion step skip. This feature allows users to drastically reduce the number of inference steps required for video generation, particularly when the identity or scene remains constant and only motion varies. By saving and reusing intermediate latent states, the system can skip redundant early diffusion steps, leading to faster generation times without compromising video quality. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a prior-based diffusion step skipping feature to accelerate inference for Wan video models. The changes include modifications to the WanVideoPipeline to support resuming from a saved latent state, along with new example scripts and documentation. My review focuses on ensuring the correctness of the new logic and improving the maintainability of the example code. I've identified a critical issue with an inconsistent condition that could lead to incorrect behavior, and a couple of medium-severity issues related to code duplication and style in the new example files.
| if prior_latents is not None and start_from_step is not None: | ||
| inputs_shared["latents"] = prior_latents.to(dtype=self.torch_dtype, device=self.device) |
There was a problem hiding this comment.
The condition for replacing latents with the prior is missing a check for prior_timesteps. The check at line 262 correctly requires prior_latents, prior_timesteps, and start_from_step. If prior_timesteps is not provided here, the scheduler will use incorrect timesteps with the loaded prior latents, which can lead to incorrect generation results. To ensure consistency and prevent bugs, the condition should be the same as the one at line 262.
| if prior_latents is not None and start_from_step is not None: | |
| inputs_shared["latents"] = prior_latents.to(dtype=self.torch_dtype, device=self.device) | |
| if prior_latents is not None and prior_timesteps is not None and start_from_step is not None: | |
| inputs_shared["latents"] = prior_latents.to(dtype=self.torch_dtype, device=self.device) | |
| if args.model == "I2V-480P": | ||
| pipe = WanVideoPipeline.from_pretrained( | ||
| torch_dtype=torch.bfloat16, | ||
| device="cuda", | ||
| model_configs=[ | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-I2V-14B-480P", | ||
| origin_file_pattern="diffusion_pytorch_model*.safetensors", | ||
| **vram_config, | ||
| ), | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-I2V-14B-480P", | ||
| origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", | ||
| **vram_config, | ||
| ), | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-I2V-14B-480P", | ||
| origin_file_pattern="Wan2.1_VAE.pth", | ||
| **vram_config, | ||
| ), | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-I2V-14B-480P", | ||
| origin_file_pattern="models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth", | ||
| **vram_config, | ||
| ), | ||
| ], | ||
| tokenizer_config=ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-T2V-1.3B", | ||
| origin_file_pattern="google/umt5-xxl/", | ||
| ), | ||
| vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024**3) - 2, | ||
| ) | ||
| else: | ||
| # T2V-1.3B (no image) | ||
| pipe = WanVideoPipeline.from_pretrained( | ||
| torch_dtype=torch.bfloat16, | ||
| device="cuda", | ||
| model_configs=[ | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-T2V-1.3B", | ||
| origin_file_pattern="diffusion_pytorch_model*.safetensors", | ||
| **vram_config, | ||
| ), | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-T2V-1.3B", | ||
| origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", | ||
| **vram_config, | ||
| ), | ||
| ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-T2V-1.3B", | ||
| origin_file_pattern="Wan2.1_VAE.pth", | ||
| **vram_config, | ||
| ), | ||
| ], | ||
| tokenizer_config=ModelConfig( | ||
| model_id="Wan-AI/Wan2.1-T2V-1.3B", | ||
| origin_file_pattern="google/umt5-xxl/", | ||
| ), | ||
| vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024**3) - 2, | ||
| ) |
There was a problem hiding this comment.
The model loading logic for I2V-480P and T2V-1.3B is largely duplicated. This can be refactored to improve maintainability and readability. By extracting the common patterns and parameterizing the differences, you can make the code more concise and easier to extend with new models in the future.
model_file_patterns = {
"I2V-480P": [
"diffusion_pytorch_model*.safetensors",
"models_t5_umt5-xxl-enc-bf16.pth",
"Wan2.1_VAE.pth",
"models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth",
],
"T2V-1.3B": [
"diffusion_pytorch_model*.safetensors",
"models_t5_umt5-xxl-enc-bf16.pth",
"Wan2.1_VAE.pth",
],
}
model_ids = {
"I2V-480P": "Wan-AI/Wan2.1-I2V-14B-480P",
"T2V-1.3B": "Wan-AI/Wan2.1-T2V-1.3B",
}
model_id = model_ids[args.model]
file_patterns = model_file_patterns[args.model]
model_configs = [
ModelConfig(
model_id=model_id,
origin_file_pattern=pattern,
**vram_config,
)
for pattern in file_patterns
]
pipe = WanVideoPipeline.from_pretrained(
torch_dtype=torch.bfloat16,
device="cuda",
model_configs=model_configs,
tokenizer_config=ModelConfig(
model_id="Wan-AI/Wan2.1-T2V-1.3B",
origin_file_pattern="google/umt5-xxl/",
),
vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024**3) - 2,
)| import imageio | ||
| import numpy as np |
There was a problem hiding this comment.
According to the PEP 8 style guide, imports should be at the top of the file. This makes dependencies clear and avoids potential issues with repeated imports. Please move import imageio and import numpy as np to the top of the file.
References
- PEP 8: E402 module level import not at top of file (link)
…ip README Made-with: Cursor
1 participant
Summary
Adds prior-based diffusion step skip for Wan video models — ~70% fewer inference steps with same quality, zero retraining.
Changes
Use case
When identity/scene is fixed and only motion varies (e.g. lip-sync, different actions), early diffusion steps are largely redundant.
Quick start
Made with Cursor