Motion-Aware Caching for Efficient Autoregressive Video Generation

Official implementation of MotionCache

Paper | Code

MotionCache is a training-free, motion-aware caching framework for efficient autoregressive video generation.

---

Table of Contents

• Overview
• Method
• Main Results
• Quick Start
• Citation
• Acknowledgments

---

Overview

Autoregressive video generation can synthesize long videos by generating chunks sequentially, but each chunk still requires expensive iterative denoising. Existing cache-reuse methods usually make coarse chunk-level reuse decisions, which can miss fine-grained motion differences inside a video: highly dynamic tokens need more updates, while static tokens can be reused more aggressively.

MotionCache addresses this with motion-aware token-level caching. It uses inter-frame differences as a lightweight proxy for motion importance, then applies motion-weighted reuse so different tokens can receive different update frequencies during generation.

---

Method

MotionCache follows a coarse-to-fine caching strategy:

• Warm-up for semantic coherence: early denoising steps are computed normally to establish stable video content.
• Motion-aware token weighting: inter-frame differences estimate which spatial tokens are more motion-sensitive.
• Token-wise cache reuse: static or low-motion regions reuse cached activations more aggressively, while high-motion regions are recomputed more often to reduce error accumulation.
• Model-agnostic integration: the method is implemented on both MAGI-1 and SkyReels-V2 without retraining.

For full derivations and ablations, please refer to the paper.

---

Main Results

MotionCache improves the speed-quality tradeoff on both SkyReels-V2 and MAGI-1. The table below reports the main VBench results from the paper.

SkyReels-V2

Method	PFLOPs	Speedup	Latency (s)	VBench	PSNR	SSIM	LPIPS
Vanilla	113	1.00x	1540	83.84%	-	-	-
TeaCache-slow	58	1.89x	814	82.67%	21.96	0.7501	0.1472
TeaCache-fast	49	2.20x	686	80.06%	18.39	0.6121	0.3063
FlowCache-slow	31	6.26x	246	82.70%	21.83	0.8733	0.1417
FlowCache-fast	27	7.19x	214	82.38%	21.17	0.8697	0.1634
MotionCache-slow	30	6.28x	245	82.84%	23.46	0.9093	0.0875
MotionCache-fast	26	7.26x	212	82.75%	21.78	0.8723	0.1478

MAGI-1

Method	PFLOPs	Speedup	Latency (s)	VBench	PSNR	SSIM	LPIPS
Vanilla	139	1.00x	1520	77.26%	-	-	-
TeaCache-slow	129	1.14x	1339	76.64%	14.74	0.4132	0.6189
TeaCache-fast	101	1.41x	1075	68.81%	11.98	0.2632	0.7670
FlowCache-slow	104	1.39x	1094	77.08%	18.16	0.6486	0.3451
FlowCache-fast	78	1.94x	782	73.42%	14.92	0.3998	0.6088
MotionCache-slow	100	1.64x	925	77.25%	19.71	0.7231	0.2510
MotionCache-fast	64	2.07x	733	74.59%	17.70	0.5600	0.4861

---

Quick Start

Environment preparation follows the corresponding base projects and the FlowCache setup style. After dependencies and model checkpoints are ready, run MotionCache with the provided scripts.

MAGI-1 VBench

cd MotionCache4MAGI-1
bash scripts/motioncache.sh

The script uses MotionCache4MAGI-1/addconfig/config.yaml as the MotionCache configuration file.

SkyReels-V2 VBench

cd MotionCache4SkyReels-V2
bash run_vbench.sh

Please update model paths, VBench prompt paths, GPU IDs, and output directories in the scripts before running.

---

Citation

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

@misc{xu2026motionawarecachingefficientautoregressive,
title={Motion-Aware Caching for Efficient Autoregressive Video Generation},
author={Jing Xu and Yuexiao Ma and Xuzhe Zheng and Xing Wang and Shiwei Liu and Chenqian Yan and Xiawu Zheng and Rongrong Ji and Fei Chao and Songwei Liu},
year={2026},
eprint={2605.01725},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.01725}
}

---

Acknowledgments

This repository builds on the following projects:

• MAGI-1
• SkyReels-V2
• TeaCache
• FlowCache