[2603.22286] WorldCache: Content-Aware Caching for Accelerated Video World Models

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.22286 (cs) [Submitted on 23 Mar 2026] Title:WorldCache: Content-Aware Caching for Accelerated Video World Models Authors:Umair Nawaz, Ahmed Heakl, Ufaq Khan, Abdelrahman Shaker, Salman Khan, Fahad Shahbaz Khan View a PDF of the paper titled WorldCache: Content-Aware Caching for Accelerated Video World Models, by Umair Nawaz and 5 other authors View PDF HTML (experimental) 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$\times$} infere...