CachedSearch: Training-Free Caching for Video Diffusion Test-Time Search

A novel approach known as CachedSearch significantly lowers the computational demands of test-time searches in video diffusion models through the use of caching that does not require training. While test-time searches enable smaller models to rival larger ones, they usually demand 2-10 times more computational resources since all candidates undergo complete denoising, even if most are ultimately rejected. The implementation of training-free caching accelerates each rollout by 2-3 times while maintaining nearly lossless quality. Research conducted on Wan2.1-T2V-1.3B, utilizing an adaptive caching wrapper, reveals that seed-matched cached rollouts and full rollouts exhibit a median Spearman rank correlation of 0.905 and 72% top-1 agreement on the VBench suite. VBench-2.0 confirms these results on a more challenging suite. Re-evaluating the cached winner at full compute preserves 90-94% of the gains from full searches, with errors mainly occurring among closely ranked candidates, thus limiting corruption. CachedSearch effectively evaluates all candidates through robust caching techniques.