Motion-Aware Caching for Efficient Autoregressive Video Generation

Autoregressive video generation paradigms offer theoretical promise for long video synthesis, yet their practical deployment is hindered by the computational burden of sequential iterative denoising. While cache reuse strategies can accelerate generation by skipping redundant denoising steps, existing methods rely on coarse-grained chunk-level skipping that fails to capture fine-grained pixel dynamics. This oversight is critical: pixels with high motion require more denoising steps to prevent error accumulation, while static pixels tolerate aggressive skipping. We formalize this insight theoretically by linking cache errors to residual instability, and propose $\textbf{MotionCache}$, a motion-aware cache framework that exploits inter-frame differences as a lightweight proxy for pixel-level motion characteristics. MotionCache employs a coarse-to-fine strategy: an initial warm-up phase establishes semantic coherence, followed by motion-weighted cache reuse that dynamically adjusts update frequencies per token. Extensive experiments on state-of-the-art models like SkyReels-V2 and MAGI-1 demonstrate that MotionCache achieves significant speedups of $\textbf{6.28}\times$ and $\textbf{1.64}\times$ respectively, while effectively preserving generation quality (VBench: 1%$\downarrow$ and 0.01%$\downarrow$ respectively). The code is available at https://github.com/ywlq/MotionCache.

MotionCache is a method for accelerating autoregressive video generation. Existing video generation models are computationally expensive because they repeatedly denoise every frame during inference, making long video generation very slow. Prior acceleration methods reuse cached computations, but they typically apply coarse-grained strategies that fail to handle regions with different motion dynamics. Our method introduces a motion-aware caching mechanism that selectively updates regions with large motion while reusing cached results for static areas. We further provide a theoretical analysis showing that caching errors are closely related to residual instability, which can be effectively estimated using frame differences. Based on this insight, MotionCache dynamically allocates computation through a coarse-to-fine inference strategy. Experiments on state-of-the-art autoregressive video generation models demonstrate that MotionCache achieves significant inference speedups while preserving high visual quality and temporal consistency.