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

📋 Table of Contents

• News
• Overview
• Method
• Main Results
• Installation
• Quick Start
• Supported Models
• Todo
• Citation
• Acknowledgments

---

📰 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.

---

🔬 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.)

---

📚 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:

• MAGI-1
• SkyReels-V2
• TeaCache
• R-KV

---

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

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