CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion

2026-07-25

AI Summary

This work addresses the high computational cost of test-time search in video diffusion models, which typically requires full denoising of all candidates. The authors propose CachedSearch, a training-free caching acceleration framework that aggressively caches intermediate results to rapidly evaluate candidates and only regenerates the winning candidate at full precision. They theoretically establish for the first time that such caching does not compromise ranking reliability, introduce a self-limiting error mechanism, and design a modular architecture that is plug-and-play, verifier-agnostic, and orthogonal to the underlying search algorithm. Integrated with an adaptive caching wrapper and mid-trajectory pruning, CachedSearch is compatible with diverse video diffusion models. Experiments show it achieves 94.7% of the search gain at 63% computational cost with eight candidates; under equal budgets, it doubles search width and improves gains by 38%, consistently demonstrating effectiveness across six models spanning four families (1.3B–14B parameters).

Abstract

Test-time search lets small video diffusion models rival larger ones, but costs 2-10x more. All candidates are fully denoised, although most are discarded. Training-free caching makes each rollout 2-3x faster at near-lossless quality. Composition is safe only if lossy caching preserves verifier rankings. We present the first study of whether caching corrupts candidate ranking in video test-time search. On Wan2.1-T2V-1.3B with an adaptive caching wrapper (~2x per-candidate speedup), ImageReward scores seed-matched cached and full rollouts. 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 at full compute retains 90-94% of the full-search gain. Errors cluster among near-tied candidates, making corruption self-limiting. This finding leads to CachedSearch. It explores every candidate with aggressive 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. The result holds from 1.3B-14B across six models and four families: Wan, LTX, CogVideoX, and Hunyuan. Wan2.1-14B matches the 1.3B model's fidelity. Mid-trajectory pruning multiplies the exploration saving to 3.11x at 88.6% capture. Ports to other model families require recalibrating a single parameter, showing 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.

Problem

• test-time search
• video diffusion
• candidate ranking
• computational cost
• caching

Innovation

• CachedSearch
• test-time search
• video diffusion
• training-free caching
• candidate ranking preservation