Screen, Match, and Cache: A Training-Free Causality-Consistent Reference Frame Framework for Human Animation

TL;DR

This work tackles the challenge of long-horizon, temporally coherent human animation by introducing FrameCache, a training-free framework with Screen, Cache, and Match stages that build and utilize a dynamic reference frame cache. It emphasizes quality-aware screening to curate high-fidelity references, redundancy-aware caching to preserve diverse and informative content, and motion-consistent matching to align references with the target pose sequence, all without additional training. Empirical results show FrameCache improves temporal coherence and visual stability, particularly with baselines that have weaker temporal modeling, while highlighting that benefits vary with baseline architecture and real-synthetic discrepancies. The framework offers practical, plug-and-play improvements for diffusion-based animation pipelines and sets the stage for adaptive, compatibility-driven cache strategies in real-world robotics and digital-twin contexts; code will be released to enable broader adoption.

Abstract

Human animation aims to generate temporally coherent and visually consistent videos over long sequences, yet modeling long-range dependencies while preserving frame quality remains challenging. Inspired by the human ability to leverage past observations for interpreting ongoing actions, we propose FrameCache, a training-free three-stage framework consisting of Screen, Cache, and Match. In the Screen stage, a multi-dimensional, quality-aware mechanism with adaptive thresholds dynamically selects informative frames; the Cache stage maintains a reference pool using a dynamic replacement-hit strategy, preserving both diversity and relevance; and the Match stage extracts behavioral features to perform motion-consistent reference matching for coherent animation guidance. Extensive experiments on standard benchmarks demonstrate that FrameCache consistently improves temporal coherence and visual stability while integrating seamlessly with diverse baselines. Despite these encouraging results, further analysis reveals that its effectiveness depends on baseline temporal reasoning and real-synthetic consistency, motivating future work on compatibility conditions and adaptive cache mechanisms. Code will be made publicly available.

Figures (5)

• Figure 1: Visual comparison of original animations (top rows) and those enhanced by our FrameCache framework (bottom rows) for MagicAnimate and StableAnimator. It demonstrates how FrameCache improves cross-frame visual consistency in details such as clothing and worn items.
• Figure 2: Overview of the proposed FrameCache framework, consisting of three stages: (a) Screen, which filters high-quality frames using CLIP-IQA and MUSIQ; (b) Cache, which maintains a dynamic and diverse reference buffer using a redundancy-aware replacement strategy; and (c) Match, which selects the most motion-consistent frame to guide generation. FrameCache operates in a training-free and causal manner, enhancing visual consistency and temporal coherence in long-term character animation.
• Figure 3: Illustration of the Cache stage in FrameCache. This module selects the most motion-consistent reference frame from the cache, ensuring temporal alignment with the current pose sequence. By choosing structurally stable and semantically relevant frames, it enhances motion continuity and reduces flickering artifacts in long-sequence character animation.
• Figure 4: Results of qualitative comparison, highlighting the regions enclosed in red boxes and the inter-frame inconsistencies.
• Figure 5: results of qualitative comparison of Magicanimate, highlighting the regions enclosed in red boxes and the inter-frame inconsistencies