Figure 1: Visualization of MagCache on Open-Sora and Wan2.1. Our MagCache can achieve better visual quality with a comparable or less latency.
Existing acceleration techniques for video diffusion models often rely on uniform heuristics or time-embedding variants to skip timesteps and reuse cached features. These approaches typically require extensive calibration with curated prompts and risk inconsistent outputs due to prompt-specific overfitting. In this paper, we introduce a novel and robust discovery: a unified magnitude law observed across different models and prompts. Specifically, the magnitude ratio of successive residual outputs decreases monotonically, steadily in most timesteps while rapidly in the last several steps. Leveraging this insight, we introduce a Magnitude-aware Cache (MagCache) that adaptively skips unimportant timesteps using an error modeling mechanism and adaptive caching strategy. Unlike existing methods requiring dozens of curated samples for calibration, MagCache only requires a single sample for calibration. Experimental results show that MagCache achieves 2.68× and 2.82x speedups on Wan 2.1 and HunyuanVideo, respectively, while preserving superior visual fidelity. It significantly outperforms existing methods in LPIPS, SSIM, and PSNR, under comparable computational budgets.
Existing methods for accelerating video diffusion models via timestep skipping generally fall into two categories: uniform heuristic strategies and adaptive approaches based on prompt-specific calibration. Uniform strategies lack accuracy because they treat all timesteps equally, ignoring the dynamic nature of residual changes during the denoising process. In contrast, adaptive methods like TeaCache attempt to model residual differences using polynomial fitting, but they require extensive calibration on dozens of curated prompts—introducing risks of overfitting and limiting generalization. In this work, we uncover a simple yet robust magnitude decay law that governs the similarity of residual outputs across timesteps: the magnitude ratio between adjacent residuals decreases steadily in early steps and more sharply in later ones. Additionally, both the standard deviation of this magnitude ratio and the token-wise cosine distance remain close to zero throughout most early steps. This suggests that residual differences between adjacent steps are closely tied to the magnitude ratio. These patterns are consistent across various prompts and model variants, making magnitude a reliable and robust indicator of residual difference. Leveraging these insights, we introduce Magnitude-aware Cache (MagCache), a simple yet effective approach, for accelerating video generation.
MagCache uses a single random prompt to calibrate the importance of timesteps. Based on the resulting magnitude curve, it dynamically skips unimportant steps through accurate error modeling and an adaptive caching strategy. Compared to existing methods, our approach significantly accelerates video generation and achieves superior visual quality, while eliminating the need for costly prompt engineering and calibration.
@misc{ma2025magcachefastvideogeneration,
title={MagCache: Fast Video Generation with Magnitude-Aware Cache},
author={Zehong Ma and Longhui Wei and Feng Wang and Shiliang Zhang and Qi Tian},
year={2025},
eprint={2506.09045},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2506.09045},
}