WorldCache: Content-Aware Caching for Accelerated Video World Models

TL;DR

WorldCache is a novel content-aware caching framework that enhances inference speed of diffusion-based video models by adaptively reusing features, reducing artifacts, and maintaining high quality without retraining.

Contribution

It introduces motion-adaptive thresholds, saliency-based drift estimation, and phase-aware scheduling for dynamic, perception-constrained feature caching in video diffusion models.

Findings

2.3x inference speedup achieved.

Preserves 99.4% of baseline quality.

Outperforms prior training-free caching methods.

Abstract

Diffusion Transformers (DiTs) power high-fidelity video world models but remain computationally expensive due to sequential denoising and costly spatio-temporal attention. Training-free feature caching accelerates inference by reusing intermediate activations across denoising steps; however, existing methods largely rely on a Zero-Order Hold assumption i.e., reusing cached features as static snapshots when global drift is small. This often leads to ghosting artifacts, blur, and motion inconsistencies in dynamic scenes. We propose \textbf{WorldCache}, a Perception-Constrained Dynamical Caching framework that improves both when and how to reuse features. WorldCache introduces motion-adaptive thresholds, saliency-weighted drift estimation, optimal approximation via blending and warping, and phase-aware threshold scheduling across diffusion steps. Our cohesive approach enables adaptive, motion-consistent feature reuse without retraining. On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves \textbf{2.3$\times$} inference speedup while preserving \textbf{99.4\%} of baseline quality, substantially outperforming prior training-free caching approaches. Our code can be accessed on \href{https://umair1221.github.io/World-Cache/}{World-Cache}.