MotionCache: Efficient Autoregressive Video Generation via Motion-Aware Caching

A new method called MotionCache addresses the computational burden of autoregressive video generation by introducing motion-aware cache reuse. Existing cache strategies skip denoising steps at a coarse chunk level, failing to account for pixel motion: high-motion pixels require more denoising to avoid errors, while static pixels tolerate aggressive skipping. MotionCache formalizes this by linking cache errors to residual instability and uses inter-frame differences as a lightweight proxy for pixel-level motion. It employs a coarse-to-fine strategy with an initial warm-up phase for semantic coherence, followed by motion-weighted caching. The approach is detailed in arXiv paper 2605.01725.