- Updated: March 10, 2026
- 1 min read
Screen, Match, and Cache: A Training‑Free Causality‑Consistent Reference Frame Framework for Human Animation
Human animation aims to generate temporally coherent and visually consistent videos over long sequences. In this article we present the FrameCache framework, a training‑free three‑stage approach consisting of Screen, Cache, and Match. The Screen stage dynamically selects informative frames using a quality‑aware mechanism with adaptive thresholds. The Cache stage maintains a reference pool with a dynamic replacement‑hit strategy to preserve diversity and relevance. The Match stage extracts behavioral features to perform motion‑consistent reference matching, guiding coherent animation.
Extensive experiments on standard benchmarks demonstrate that FrameCache consistently improves temporal coherence and visual stability while integrating seamlessly with diverse baselines. The framework’s effectiveness depends on baseline temporal reasoning and real‑synthetic consistency, motivating future work on compatibility conditions and adaptive cache mechanisms.
Read the full paper on arXiv and explore related resources on ubos.tech.