MagCache: Fast Video Generation with Magnitude-Aware Cache (2506.09045v1)

Abstract: 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 and 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 2.1x and 2.68x speedups on Open-Sora and Wan 2.1, respectively, while preserving superior visual fidelity. It significantly outperforms existing methods in LPIPS, SSIM, and PSNR, under comparable computational budgets.

• The paper introduces a novel magnitude-aware caching framework to accelerate video diffusion by intelligently skipping redundant timesteps.
• It employs accurate error modeling and adaptive caching based on a unified magnitude law to ensure minimal loss in visual fidelity.
• Experimental results demonstrate speedups of 2.1× and 2.68× on Open-Sora and Wan 2.1, outperforming current caching methods in quality metrics.

Overview of "MagCache: Fast Video Generation with Magnitude-Aware Cache"

The paper "MagCache: Fast Video Generation with Magnitude-Aware Cache" introduces a novel framework tailored to enhance the efficiency of video diffusion models. These models have gained significant prominence in visual generative tasks but suffer from inherent inefficiencies, predominantly in their inference speed. The proposed MagCache system addresses these limitations by leveraging a magnitude-aware caching strategy, which derives its effectiveness from a newly discovered unified magnitude law applicable across multiple models and prompts.

Approach and Methodology

The authors provide a meticulous analysis of the magnitude ratio behavior of successful residual outputs during the diffusion process. Their empirical findings illustrate that this ratio decreases steadily across the majority of timesteps and only plummets rapidly during the concluding steps. The paper capitalizes on this observation by developing an adaptive caching strategy that intelligently skips redundant timesteps.

MagCache comprises two core mechanisms: accurate error modeling and adaptive caching. The error modeling leverages the magnitude ratio to reliably predict errors introduced by skipping timesteps, ensuring minimal compromise on visual fidelity. The adaptive caching strategy utilizes these predictions to determine whether a timestep can be skipped based on predefined error thresholds and maximum permissible step lengths.

Experimental Evaluation

Rigorous experimental assessments of MagCache demonstrate substantial improvements in inference speed while maintaining or enhancing visual quality. On video diffusion models such as Open-Sora and Wan 2.1, MagCache achieves significant speedups—2.1× on Open-Sora and 2.68× on Wan 2.1—compared to existing methodologies. Furthermore, the models accelerated by MagCache outperform existing caching-based methods in visual quality metrics, including LPIPS, SSIM, and PSNR under similar computational constraints.

Implications

From a practical standpoint, the deployment of MagCache means video generation can be executed in real-time or on resource-constrained platforms without compromising the quality of the produced videos. Theoretically, the identification of a unified magnitude law offers a robust criterion for accelerating inference that could extend beyond video diffusion models to other domains in AI.

Future Directions

The implications of MagCache encourage exploration into broader applicability across various diffusion models and tasks, especially in text-to-image synthesis. Additionally, further refinement of error modeling mechanisms to accommodate a diverse range of prompts or unconventional model architectures may reveal new efficiencies or improvements in visual generation fidelity.

Conclusion

The paper articulates a compelling argument for the adoption of magnitude-aware caching strategies in diffusion models, not just for enhancing speed but also for refining visual outputs. "MagCache: Fast Video Generation with Magnitude-Aware Cache" represents an important step towards optimizing video synthesis and potentially other generative tasks, and sets the stage for ongoing research in adaptive caching and acceleration methods for complex AI systems.