📚English | <a href="./README_CN.md">📚中文阅读 </a> | <a href="https://mp.weixin.qq.com/s/o6MMOzbmGBRpB_a_9U8JMw?">机器之心</a>

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

<div class="is-size-5 publication-authors" align="center"> <a href="https://github.com/joelulu" target="_blank">Huanlin Gao</a>1,2*,  <a href="https://scholar.google.com/citations?hl=zh-CN&view_op=list_works&user=gpNOW2UAAAAJ" target="_blank">Ping Chen</a>1,2*,  <a href="https://github.com/stone002" target="_blank">Fuyuan Shi</a>1,2,  <a href="https://github.com/tanchaow" target="_blank">Chao Tan</a>1,2,  <a href="https://scholar.google.com/citations?hl=en&user=L4OXOs0AAAAJ" target="_blank">Zhaoxiang Liu</a>1,2 <a href="https://github.com/FangGet" target="_blank">Fang Zhao</a>1,2†,  <a href="https://scholar.google.com/citations?user=CFUQLCAAAAAJ&hl=en" target="_blank">Kai Wang</a>1,2,  <a href="https://scholar.google.com.hk/citations?user=kCC2oKwAAAAJ&hl=zh-CN&oi=ao" target="_blank">Shiguo Lian</a>1,2† </div> <div class="is-size-5 publication-authors" align="center"> 1Data Science & Artificial Intelligence Research Institute, China Unicom,  2Unicom Data Intelligence, China Unicom </div> <div class="is-size-5 publication-authors" align="center"> (* Equal contribution. † Corresponding author.) </div> <h5 align="center"> <a href="https://unicomai.github.io/LeMiCa/" target="_blank"> <img src="https://img.shields.io/badge/Project-Website-blue.svg" alt="Project Page"> </a>  <a href="https://arxiv.org/abs/2511.00090" target="_blank"> <img src="https://img.shields.io/badge/Paper-PDF-critical.svg?logo=adobeacrobatreader" alt="Paper"> </a>  </a> <a href="./LICENSE" target="_blank"> <img src="https://img.shields.io/badge/License-Apache%202.0-yellow.svg" alt="License"> </a> <a href="https://github.com/UnicomAI/LeMiCa/stargazers" target="_blank"> <img src="https://img.shields.io/github/stars/UnicomAI/LeMiCa.svg?style=social" alt="GitHub Stars"> </a> </h5>

LeMiCa Overview

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

[2025/01/29] 🔥 Our latest work "MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference" is accepted by ICLR 2026! Codes are available at MeanCache! MeanCache achieves 4.12×, 4.56×, and 3.59× acceleration on FLUX.1, Qwen-Image, and HunyuanVideo, while consistently outperforming state-of-the-art caching baselines in generation quality. For more details, please refer to our latest research paper.
[2025/01/20] 🔥 Added support for FLUX.1-dev and FLUX.2-Klein in LeMiCa4FLUX
[2025/12/15] ✨ ComfyUI-LeMiCa has been seamlessly integrated into ComfyUI and is fully compatible with ComfyUI’s native nodes.
[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

ComfyUI-LeMiCa

FLUX.2 [Klein-9B]

| Method | Flux.2(klein-9B) | LeMiCa-slow | LeMiCa-medium | LeMiCa-fast | LeMiCa-ultra | |:-------------------:|:--------------------:|:--------------:|:--------------:|:-------------:|:-------------:| | Latency | 20.04 s | 10.77 s | 8.45 s | 6.54 s | 4.59 s | | T2I | <img width="120" alt="Flux.2(klein-9B)" src="https://github.com/user-attachments/assets/07989284-1856-44c8-8a6b-2b46d9532ff2" /> | <img width="120" alt="LeMiCa-slow" src="https://github.com/user-attachments/assets/66130c70-12e3-479f-9062-55c249128821" /> | <img width="120" alt="LeMiCa-medium" src="https://github.com/user-attachments/assets/445cd63a-a896-4bfa-8705-a1043ed42bef" /> | <img width="120" alt="LeMiCa-fast" src="https://github.com/user-attachments/assets/042f45c6-e9ac-4b60-a6a4-ec7fc603f6e3" /> | <img width="120" alt="LeMiCa-ultra" src="https://github.com/user-attachments/assets/c9356da6-924f-4502-b0e4-a902820f7740" /> |

Qwen-Image-2512

| Method | Qwen-Image-2512 | LeMiCa-slow | LeMiCa-medium | LeMiCa-fast | |:-------:|:-------:|:-----------:|:-------------:|:-----------:| | Latency | 32.8 s | 18.83 s | 14.35 s | 10.41 s | | T2I | <img width="160" alt="Qwen-Image-2512" src="https://github.com/user-attachments/assets/ae47f0c2-e687-4fc8-bf78-8f9d968feda5" /> | <img width="160" alt="LeMiCa-slow" src="https://github.com/user-attachments/assets/9cc9c647-6650-4294-b8fd-6263c74e20b5" /> | <img width="160" alt="LeMiCa-medium" src="https://github.com/user-attachments/assets/0ff53c58-e1be-4e4f-9574-c6e60cf4e342" /> | <img width="160" alt="LeMiCa-fast" src="https://github.com/user-attachments/assets/220e85ea-17e7-4d86-b9cc-6fdb62049d62" /> |

HunyuanVideo1.5

T2V 720P (Up to a 2.86× speedup）

https://github.com/user-attachments/assets/ebed2e0f-87f4-408e-98e3-93bd29bbc99f

I2V 720P (Up to a 3.88× speedup）

https://github.com/user-attachments/assets/d1a83d45-579f-4174-9477-ba0b9aebb322

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 | <img width="120" alt="Flux2 CPU-offload" src="https://github.com/user-attachments/assets/76fda91e-8819-4914-87e4-8a832135da0f" /> | <img width="120" alt="Flux2" src="https://github.com/user-attachments/assets/a3f320e3-9d36-4618-9953-f714646e6bf7" /> | <img width="120" alt="LeMiCa-slow" src="https://github.com/user-attachments/assets/b28fdd2b-8178-4ba7-bf23-3da66f555593" /> | <img width="120" alt="LeMiCa-medium" src="https://github.com/user-attachments/assets/72b4361d-8afe-4c94-9654-77697e3c1444" /> | <img width="120" alt="LeMiCa-fast" src="https://github.com/user-attachments/assets/56ea6af3-e1a5-4134-890b-24f5666081e9" /> |

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 | <img width="120" alt="Z-Image" src="https://github.com/user-attachments/assets/e7aa76a9-2ffd-4cfc-8c9d-2240f357850b" /> | <img width="120" alt="LeMiCa-slow" src="https://github.com/user-attachments/assets/e7ff50b9-44bb-48ff-86f9-14dacc1b5144" /> | <img width="120" alt="LeMiCa-medium" src="https://github.com/user-attachments/assets/786ad801-ac92-4467-86a6-661b5e7dca53" /> | <img width="120" alt="LeMiCa-fast" src="https://github.com/user-attachments/assets/722d79b1-69fb-4683-914f-e92533394393" /> |

Wan2.1

https://github.com/user-attachments/assets/3d99b959-7253-47ec-af0a-da13a66e6d49

Open-Sora

<details> <summary>Click to expand Open-Sora example</summary>

https://github.com/user-attachments/assets/ba205856-2d77-494a-aaa9-09189ba2915c

</details>

Qwen-Image

<details> <summary>Click to expand Qwen-Image example</s

LeMiCa

Install / Use

README