Shadowgraph
About
Sign in

Zehong-Ma/MagCache

The official code for NeurIPS 2025 "MagCache: Fast Video Generation with Magnitude-Aware Cache"

280

stars

49

commits

Python

primary language

Nov 17, 2025

updated

zehong-ma.github.io/MagCache/

README

MagCache: Fast Video Generation with Magnitude-Aware Cache

Zehong Ma1,2†,  Longhui Wei2*‡,  Feng Wang2,  Shiliang Zhang1*,  Qi Tian2
1State Key Laboratory of Multimedia Information Processing,
School of Computer Science, Peking University,  2Huawei Inc.
(† Work was done during internship at Huawei Inc., ‡ Project Leader. * Corresponding author.)

arXiv Home Page License github


Figure 1: Visualization of MagCache on Open-Sora and Wan2.1. MagCache can achieve better visual quality with a comparable or less latency. More demos are shown in the subfolders of each model.

🫖 Introduction

We introduce Magnitude-aware Cache (MagCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps based on the robust magnitude observations, thereby accelerating the inference. MagCache works well for Video Diffusion Models, Image Diffusion models. For more details and results, please visit our project page.


Figure 2: Relationships between output residuals across diffusion timesteps. Magnitude ratio serves as both an accurate and stable criterion for measuring the difference between residuals.

🔥 Latest News

  • If you like our project, please give us a star ⭐ on GitHub for the latest update.
  • [2025/11/17] 🔥Accepted by NeurIPS 2025. Welcome to follow and improve MagCache.
  • [2025/8/20] 🔥 Support Qwen-Image and Qwen-Image-Edit with 1.5x speedup officially.
  • [2025/8/7] 🔥 Support Wan2.2 with 1.5-2x speedup officially.
  • [2025/7/14] 🔥 Support OmniGen2 with 1.9x speedup officially.
  • [2025/6/30] 🔥 Support Flux-Kontext with 2x speedup officially.
  • [2025/6/19] 🔥 Support FramePack officially.
  • [2025/6/17] 🔥 Support VACE 1.3B&14B officially.
  • [2025/6/17] 🔥 MagCache is supported by ComfyUI-WanVideoWrapper. Thanks @kijai.
  • [2025/6/16] 🔥 Support Chroma officially. Please try it in ComfyUI-MagCache.
  • [2025/6/12] 🔥 Support ComfyUI-MagCache officially. Please have a try!
  • [2025/6/10] 🔥 Support Wan2.1, HunyuanVideo, FLUX, Open-Sora.
  • [2025/6/10] 🎉 Release the code of MagCache.
  • [2025/6/10] 🎉 Release the paper of MagCache.

🧩 Community Contributions

If you develop/use MagCache in your projects and you would like more people to see it, please inform us.(zehongma@stu.pku.edu.cn)

ComfyUI

🎉 Supported Models

Text to Video

Image to Video

Video Editing

Text to Image

Image Editing

Demos

  • HunyuanVideo T2V 720P

Prompt: The video shows two astronauts in bulky suits walking slowly on the moon’s surface.... The video shows two astronauts in bulky suits walking slowly on the moon’s surface, against a vast starry universe. Their steps are heavy and slow, kicking up dust in the low-gravity environment. The scene is silent, mysterious, and evokes the courage and dreams of space exploration.
  • Wan2.1 14B I2V 720P

Prompt: A woman in black lace stands confidently in a dim Art Deco interior with polished marble floors.... A woman in black lace stands confidently in a dim Art Deco interior with polished marble floors. Stark chiaroscuro lighting highlights her sharp features as she tilts her head, crimson lips parting in a knowing smile. Her smoldering gaze meets the viewer while she turns gracefully, lace casting shifting shadows on the walls. A medium shot with a subtle dolly zoom, framed by velvet drapes, adds depth. The mysterious, refined atmosphere blends modern elegance with vintage Hollywood glamour, rendered in 8K hyper-realistic detail, metallic gold accents glowing in the soft light.
  • Wan2.1 14B T2V 720P

Prompt: The video shows two astronauts in bulky suits walking slowly on the moon’s surface.... The video shows two astronauts in bulky suits walking slowly on the moon’s surface, against a vast starry universe. Their steps are heavy and slow, kicking up dust in the low-gravity environment. The scene is silent, mysterious, and evokes the courage and dreams of space exploration.

Prompt: A stylish woman walks down a Tokyo street filled with warm glowing neon.... A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.
  • Wan2.1 1.3B T2V 480P

Prompt: Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage. Prompt: Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage.

Prompt: A tranquil tableau of an ornate Victorian streetlamp.... A tranquil tableau of an ornate Victorian streetlamp standing on a cobblestone street corner, illuminating the empty night
  • FramePack

Prompt: The girl dances gracefully, with clear movements, full of charm. The girl dances gracefully, with clear movements, full of charm.
  • FLUX-dev T2I


Prompt: A photo of a black bicycle. Prompt: A photo of a black bicycle.
  • FLUX-Kontext Image

🤖 Instructions for Supporting Other Models

  • Welcome for PRs to support other models.
  • Please implement the magcache_calibration function by referring the code in MagCache4Wan2.1
  • Utilize a random prompt or input to generate the magnitude ratio.
  • Implement the magcache_forward function by adding a few lines into the original forward function.
  • Reuse the default hyper-parameters or set your own hyper-parameters to achieve a trade-off between latency and quality.

💐 Acknowledgement

This repository is built based on VideoSys, Diffusers, Open-Sora, FLUX, Chroma, Wan2.1, HunyuanVideo, FramePack, and TeaCache. Thanks for their contributions!

🔒 License

📖 Citation

If you find MagCache is useful in your research or applications, please consider giving us a star ⭐ and citing it by the following BibTeX entry.

@inproceedings{
  ma2025magcache,
  title={MagCache: Fast Video Generation with Magnitude-Aware Cache},
  author={Zehong Ma and Longhui Wei and Feng Wang and Shiliang Zhang and Qi Tian},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=KZn7TDOL4J}
}

Contributors

Zehong-Ma

49 commits

Similar repos

Zehong-Ma/ComfyUI-MagCache

The official code that integrates MagCache (Fast Video Generation with Magnitude-Aware Cache) with…

276

same org

ali-vilab/TeaCache

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

1,377

wangwenxing-mt/FastVideo

0

Shenyi-Z/TaylorSeer

[ICCV2025] From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers

415

alphaXiv/mage-7255fc67

0

PKU-YuanGroup/MagicTime

[TPAMI 2025🔥] MagicTime: Time-lapse Video Generation Models as Metamorphic Simulators

1,337

horseee/DeepCache

[CVPR 2024] DeepCache: Accelerating Diffusion Models for Free

970

FofGofx/WorldCache

[ICML26] Official Repo for WorldCache: Accelerating World Models for Free via Heterogeneous Token…

43

Languages

Python

99.4%

Zehong-Ma/MagCache

The official code for NeurIPS 2025 "MagCache: Fast Video Generation with Magnitude-Aware Cache"

280

stars

49

commits

Python

primary language

Nov 17, 2025

updated

zehong-ma.github.io/MagCache/

README

MagCache: Fast Video Generation with Magnitude-Aware Cache

Zehong Ma1,2†,  Longhui Wei2*‡,  Feng Wang2,  Shiliang Zhang1*,  Qi Tian2
1State Key Laboratory of Multimedia Information Processing,
School of Computer Science, Peking University,  2Huawei Inc.
(† Work was done during internship at Huawei Inc., ‡ Project Leader. * Corresponding author.)

arXiv Home Page License github


Figure 1: Visualization of MagCache on Open-Sora and Wan2.1. MagCache can achieve better visual quality with a comparable or less latency. More demos are shown in the subfolders of each model.

🫖 Introduction

We introduce Magnitude-aware Cache (MagCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps based on the robust magnitude observations, thereby accelerating the inference. MagCache works well for Video Diffusion Models, Image Diffusion models. For more details and results, please visit our project page.


Figure 2: Relationships between output residuals across diffusion timesteps. Magnitude ratio serves as both an accurate and stable criterion for measuring the difference between residuals.

🔥 Latest News

  • If you like our project, please give us a star ⭐ on GitHub for the latest update.
  • [2025/11/17] 🔥Accepted by NeurIPS 2025. Welcome to follow and improve MagCache.
  • [2025/8/20] 🔥 Support Qwen-Image and Qwen-Image-Edit with 1.5x speedup officially.
  • [2025/8/7] 🔥 Support Wan2.2 with 1.5-2x speedup officially.
  • [2025/7/14] 🔥 Support OmniGen2 with 1.9x speedup officially.
  • [2025/6/30] 🔥 Support Flux-Kontext with 2x speedup officially.
  • [2025/6/19] 🔥 Support FramePack officially.
  • [2025/6/17] 🔥 Support VACE 1.3B&14B officially.
  • [2025/6/17] 🔥 MagCache is supported by ComfyUI-WanVideoWrapper. Thanks @kijai.
  • [2025/6/16] 🔥 Support Chroma officially. Please try it in ComfyUI-MagCache.
  • [2025/6/12] 🔥 Support ComfyUI-MagCache officially. Please have a try!
  • [2025/6/10] 🔥 Support Wan2.1, HunyuanVideo, FLUX, Open-Sora.
  • [2025/6/10] 🎉 Release the code of MagCache.
  • [2025/6/10] 🎉 Release the paper of MagCache.

🧩 Community Contributions

If you develop/use MagCache in your projects and you would like more people to see it, please inform us.(zehongma@stu.pku.edu.cn)

ComfyUI

🎉 Supported Models

Text to Video

Image to Video

Video Editing

Text to Image

Image Editing

Demos

  • HunyuanVideo T2V 720P

Prompt: The video shows two astronauts in bulky suits walking slowly on the moon’s surface.... The video shows two astronauts in bulky suits walking slowly on the moon’s surface, against a vast starry universe. Their steps are heavy and slow, kicking up dust in the low-gravity environment. The scene is silent, mysterious, and evokes the courage and dreams of space exploration.
  • Wan2.1 14B I2V 720P

Prompt: A woman in black lace stands confidently in a dim Art Deco interior with polished marble floors.... A woman in black lace stands confidently in a dim Art Deco interior with polished marble floors. Stark chiaroscuro lighting highlights her sharp features as she tilts her head, crimson lips parting in a knowing smile. Her smoldering gaze meets the viewer while she turns gracefully, lace casting shifting shadows on the walls. A medium shot with a subtle dolly zoom, framed by velvet drapes, adds depth. The mysterious, refined atmosphere blends modern elegance with vintage Hollywood glamour, rendered in 8K hyper-realistic detail, metallic gold accents glowing in the soft light.
  • Wan2.1 14B T2V 720P

Prompt: The video shows two astronauts in bulky suits walking slowly on the moon’s surface.... The video shows two astronauts in bulky suits walking slowly on the moon’s surface, against a vast starry universe. Their steps are heavy and slow, kicking up dust in the low-gravity environment. The scene is silent, mysterious, and evokes the courage and dreams of space exploration.

Prompt: A stylish woman walks down a Tokyo street filled with warm glowing neon.... A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.
  • Wan2.1 1.3B T2V 480P

Prompt: Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage. Prompt: Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage.

Prompt: A tranquil tableau of an ornate Victorian streetlamp.... A tranquil tableau of an ornate Victorian streetlamp standing on a cobblestone street corner, illuminating the empty night
  • FramePack

Prompt: The girl dances gracefully, with clear movements, full of charm. The girl dances gracefully, with clear movements, full of charm.
  • FLUX-dev T2I


Prompt: A photo of a black bicycle. Prompt: A photo of a black bicycle.
  • FLUX-Kontext Image

🤖 Instructions for Supporting Other Models

  • Welcome for PRs to support other models.
  • Please implement the magcache_calibration function by referring the code in MagCache4Wan2.1
  • Utilize a random prompt or input to generate the magnitude ratio.
  • Implement the magcache_forward function by adding a few lines into the original forward function.
  • Reuse the default hyper-parameters or set your own hyper-parameters to achieve a trade-off between latency and quality.

💐 Acknowledgement

This repository is built based on VideoSys, Diffusers, Open-Sora, FLUX, Chroma, Wan2.1, HunyuanVideo, FramePack, and TeaCache. Thanks for their contributions!

🔒 License

📖 Citation

If you find MagCache is useful in your research or applications, please consider giving us a star ⭐ and citing it by the following BibTeX entry.

@inproceedings{
  ma2025magcache,
  title={MagCache: Fast Video Generation with Magnitude-Aware Cache},
  author={Zehong Ma and Longhui Wei and Feng Wang and Shiliang Zhang and Qi Tian},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=KZn7TDOL4J}
}

Similar repos

Zehong-Ma/ComfyUI-MagCache

The official code that integrates MagCache (Fast Video Generation with Magnitude-Aware Cache) with…

276

same org

ali-vilab/TeaCache

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

1,377

wangwenxing-mt/FastVideo

0

Shenyi-Z/TaylorSeer

[ICCV2025] From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers

415

alphaXiv/mage-7255fc67

0

PKU-YuanGroup/MagicTime

[TPAMI 2025🔥] MagicTime: Time-lapse Video Generation Models as Metamorphic Simulators

1,337

horseee/DeepCache

[CVPR 2024] DeepCache: Accelerating Diffusion Models for Free

970

FofGofx/WorldCache

[ICML26] Official Repo for WorldCache: Accelerating World Models for Free via Heterogeneous Token…

43

Contributors

Zehong-Ma

49 commits

Languages

Python

99.4%