Abstract
Test-time search allows small video diffusion models to compete with larger ones, but at a cost of 2-10 times more. All candidates are fully denoised, although most are discarded. Training-free caching accelerates each rollout by 2-3 times while maintaining near-lossless quality. Composition is safe only if lossy caching preserves verifier rankings. We present the first study on whether caching corrupts candidate ranking in video test-time search.
Using the Wan2.1-T2V-1.3B model with an adaptive caching wrapper, we achieve about 2x speedup per candidate, with ImageReward scores reflecting the match between cached and full rollouts. The median per-prompt Spearman rank correlation is 0.905, with 72% top-1 agreement on the VBench suite. VBench-2.0 replicates this result on a harder suite. Recomputing the cached winner retains 90-94% of the full-search gain. Errors cluster among near-tied candidates, making corruption self-limiting. This finding leads to CachedSearch.
CachedSearch aggressively explores all candidates with caching, then re-generates only the winner at full compute. At N=8, it captures 94.7% of best-of-N's gain at 63% of the cost. Capture rises with width. At matched budget, it searches twice as wide for 38% more gain. This result holds across six models and four families: Wan, LTX, CogVideoX, and Hunyuan, with Wan2.1-14B matching the fidelity of the 1.3B model. Mid-trajectory pruning enhances exploration savings to 3.11x at 88.6% capture. Ports to other model families require recalibrating a single parameter, indicating that fidelity tracks architecture rather than parameter count.
CachedSearch is training-free, verifier-agnostic, and orthogonal to the search algorithm, making it a plug-in multiplier for test-time scaling.
Blogger's Review: The innovation of CachedSearch lies in its training-free nature and efficient caching strategy, significantly enhancing the performance of video generation models at test time. This approach saves computational resources while maintaining high-quality outputs, showcasing its potential in optimizing model architectures. As model scales increase, CachedSearch is expected to play a crucial role in more applications.