SpectralCache: Accelerating Diffusion-Based World Models via Spectral Feature Caching

This research proposes SpectralCache, a training-free spectral caching framework for diffusion-based world models. By exploiting the stable singular subspaces and predictable evolution of singular values across denoising steps, SpectralCache improves inference efficiency while preserving generation quality. Experiments show 5.22x acceleration on HunyuanWorld-Voyager-13B with a WorldScore of 65.90 for static scenes.