📚English | 📚中文阅读   |   机器之心

[NeurIPS 2025 Spotlight] LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation

Huanlin Gao^1,2*,  Ping Chen^1,2*,  Fuyuan Shi^1,2,  Chao Tan^1,2,  Zhaoxiang Liu^1,2
Fang Zhao^1,2†,  Kai Wang^1,2,  Shiguo Lian^1,2†

¹Data Science & Artificial Intelligence Research Institute, China Unicom,  ²Unicom Data Intelligence, China Unicom

(* Equal contribution. † Corresponding author.)

Introduction

LeMiCa is a training-free acceleration framework for diffusion-based video generation (and extendable to image generation). Instead of using local heuristic thresholds, LeMiCa formulates cache scheduling as a global path optimization problem with error-weighted edges and introduces a Lexicographic Minimax strategy to bound the worst-case global error. This global planning improves both inference speed and consistency across frames. For more details and visual results, please visit our project page.

🔥 Latest News

• [2025/12/08] 🔥 Support HunyuanVideo-1.5 for both T2V and I2V.
• [2025/12/02] 🔥 Support Z-Image and FLUX.2.
• [2025/11/14] ⭐ We have open-sourced Awesome-Acceleration-GenAI, collecting the latest generation acceleration techniques. Feel free to check it out !
• [2025/11/13] 🔥 Support Wan2.1 for both T2V and I2V.
• [2025/11/07] 🔥 Support Qwen-Image and Inference Code Released !
• [2025/10/29] 🚀 Code will be released soon !
• [2025/09/18] ✨ Selected as a NeurIPS 2025 Spotlight paper.
• [2025/09/18] ✨ Initial public release of LeMiCa.

Demo

HunyuanVideo1.5

T2V 720P (Up to a 2.86× speedup）

HunyuanVideo1.5_T2V_720P.mp4

I2V 720P (Up to a 3.88× speedup）

HunyuanVideo1.5_I2V_720P.mp4

FLUX.2

Method	Flux.2(cpu-offload)	Flux.2	LeMiCa-slow	LeMiCa-medium	LeMiCa-fast
Latency	101.2 s	32.70 s	13.41 s	10.20 s	6.99 s
T2I

Z-Image

Method	Z-Image	LeMiCa-slow	LeMiCa-medium	LeMiCa-fast
Latency	2.55 s	2.19 s	1.94 s	1.78 s
T2I

Wan2.1

Wan2.1_I2V_14B_832_480.mp4

Open-Sora

Click to expand Open-Sora example

opensora_grid_5x5_with_header_bold.mp4

Qwen-Image

Click to expand Qwen-Image example

Supported Models

LeMiCa currently supports and has been tested on the following diffusion-based models:

Text-to-Video

• Open-Sora
• Latte
• CogVideoX 1.5
• Wan2.1
• HunyuanVideo-1.5

Text-to-Image

• Qwen-Image
• Z-Image
• FLUX.2

ToDo List

• 🗹 Public Project Page
• 🗹 Paper Released
• 🗹 Text-to-Image Forward Inference
• 🗹 Text-to-Video Forward Inference
• ☐ DAG Construction Code
• ☐ Support Acceleration Framework

Community Contributions & Friendly Links

• Qwen-Image and CogVideo featured LeMiCa on their project homepages.

• Cache-DiT A unified and flexible inference engine for DiTs, integrating and applying LeMiCa’s core insights. Welcome to try and explore. Details

Acknowledgement

This repository is built based on or inspired by the following open-source projects: Diffusers, TeaCache, VideoSys. We sincerely thank these communities for their open contributions and inspiration.

License

The majority of this project is released under the Apache 2.0 license as found in the LICENSE file.

📖 Citation

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

@inproceedings{gao2025lemica,
  title     = {LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation},
  author    = {Huanlin Gao and Ping Chen and Fuyuan Shi and Chao Tan and Zhaoxiang Liu and Fang Zhao and Kai Wang and Shiguo Lian},
  journal   = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2025},
  url       = {https://arxiv.org/abs/2511.00090}
}

⭐ Star History