Naive Caching Causes Motion Drift

Averages hide local motion. Blindly copying old activations destroys temporal coherence before global loss triggers a refresh.

WorldCache

Content-Aware Caching for Accelerated World Models

A training-free framework that predicts skipped computation rather than copying it over blindly.

Motion-Aware Caching Logic

WorldCache treats caching like a localized prediction. It controls the pace with causal tracking while interpolating the next state.

Driven by Four Key Ideologies

Causal Feature Caching

Dynamically scales caching tolerance based on early layer motion velocity.

Saliency Weighted Drift

Penalizes caching errors in perceptually critical high-frequency regions.

Optimal Feature Approx.

Interpolates skipped cache states using trajectory matching.

Adaptive Scheduling

Exponentially relaxes caching constraints in later denoising stages.

Early

Late

Over baseline architectures while strictly maintaining visual fidelity, motion dynamics, and prompt adherence across Cosmos, WAN2.1, and DreamDojo.

Evaluation & Generalization

Flawless Temporal Coherence