Hugging Face's logo

Datasets:
Cccccz
/
temp

Dataset card Files
xet
 Community
temp / FlowCache /README.md
Cccccz's picture
Cccccz
Add files using upload-large-folder tool
153dd42 verified
preview code
|
Raw
History Blame Contribute Delete
6.8 kB

Flow Caching for Autoregressive Video Generation

ICLR 2026

Paper | arXiv |

The first caching framework specifically designed for autoregressive video generation

Achieving 2.38× speedup on MAGI-1 and 6.7× on SkyReels-V2 with negligible quality degradation

License: MIT Python 3.8+ PyTorch

📋 Table of Contents

📰 News

  • 📄 2026.02.12: Paper available on arXiv!
  • 🚀 2026.02.10: Code release for MAGI-1 and SkyReels-V2!
  • 🎉 2026.01.26: Paper accepted by ICLR 2026!

🌟 Overview

FlowCache is a caching framework designed specifically for autoregressive video generation models. Unlike traditional caching methods that treat all frames uniformly, FlowCache introduces a chunkwise caching strategy where each video chunk maintains independent caching policies, complemented by importance-based KV cache compression that maintains fixed memory bounds while preserving generation quality.

Overview

🔬 Method

Key Findings

Our key insight: Different video chunks exhibit heterogeneous denoising states at identical timesteps, necessitating independent caching policies for optimal performance.

Framework Overview

FlowCache introduces three key innovations for training-free acceleration of autoregressive video generation:

  • Chunkwise Denoising Heterogeneity: We identify and formalize that denoising progress varies significantly across video chunks—even at the same timestep—necessitating per-chunk caching decisions.

  • Chunkwise Adaptive Caching: A novel design where each chunk independently decides whether to reuse or recompute based on its own similarity trajectory.

  • KV Cache Compression Tailored for Video: We adapt importance–redundancy scoring to autoregressive video generation KV cache compression by introducing an efficient, equivalence-preserving similarity computation, thereby enhancing cache diversity without sacrificing efficiency.

These contributions collectively make FlowCache the first theoretically grounded, training-free caching framework for efficient autoregressive video generation.

For more details, please refer to the original paper.

📊 Main Results

Quantitative Performance

MAGI-1 (4.5B model)

Method PFLOPs Speedup Latency (s) VBench LPIPS SSIM PSNR
Vanilla 306 1.0× 2873 77.06% - - -
TeaCache-slow 294 1.12× 2579 77.50% 0.6211 0.2801 13.26
TeaCache-fast 225 1.44× 1998 70.11% 0.8160 0.1138 8.94
FlowCache-slow 161 1.86× 1546 78.96% 0.3160 0.6497 22.34
FlowCache-fast 140 2.38× 1209 77.93% 0.4311 0.5140 19.27

SkyReels-V2 (1.3B model)

Method PFLOPs Speedup Latency (s) VBench LPIPS SSIM PSNR
Vanilla 113 1.0× 1540 83.84% - - -
TeaCache-slow 58 1.89× 814 82.67% 0.1472 0.7501 21.96
TeaCache-fast 49 2.2× 686 80.06% 0.3063 0.6121 18.39
FlowCache-slow 36 5.88× 262 83.12% 0.1225 0.7890 23.74
FlowCache-fast 28 6.7× 230 83.05% 0.1467 0.7635 22.95

Visualization

🛠️ Installation

Prerequisites

  • Python 3.8+
  • CUDA 11.8+ (or 12.x)
  • PyTorch 2.0+

MAGI-1 Setup 

cd FlowCache4MAGI-1
pip install -r requirements.txt

SkyReels-V2 Setup 

cd FlowCache4SkyReels-V2
pip install -r requirements.txt

🚀 Quick Start

MAGI-1 

cd FlowCache4MAGI-1

bash scripts/single_run/flowcache_t2v.sh

SkyReels-V2 

cd FlowCache4SkyReels-V2

bash run_flowcache_fast.sh

🎯 Supported Models

Model Type Status
MAGI-1 4.5B-distill ✅
SkyReels-V2 1.3B ✅

📝 Todo List

  • Support more autoregressive video generation models (e.g., self-forcing, causal-forcing, etc.)
  • Integrate other training-free acceleration methods (e.g., quantization, etc.)

📮 Contact

For questions and collaboration inquiries, please contact the co-first authors. The following are all WeChat IDs:

  • Yuexiao Ma: ma18640400169
  • Xuzhe Zheng: zhengxuzhe_
  • Jing Xu: a2665048215

📚 Citation

If you find FlowCache useful for your research, please cite:

@misc{ma2026flowcachingautoregressivevideo,
      title={Flow caching for autoregressive video generation}, 
      author={Yuexiao Ma and Xuzhe Zheng and Jing Xu and Xiwei Xu and Feng Ling and Xiawu Zheng and Huafeng Kuang and Huixia Li and Xing Wang and Xuefeng Xiao and Fei Chao and Rongrong Ji},
      year={2026},
      eprint={2602.10825},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.10825}, 
}

🙏 Acknowledgments

We thank the authors of the following projects for their valuable contributions:

⭐ If you find this project useful, please consider giving it a star! ⭐

For questions and feedback, please open an issue on GitHub.