SpectralCache reuses stable singular subspaces and extrapolates singular values to reach 5.22x acceleration on HunyuanWorld-Voyager-13B while holding a WorldScore of 65.90

The paper reveals that world-model features exhibit highly stable singular subspaces across nearby denoising steps, while their singular values follow predictable evolution patterns. Existing caching methods mainly exploit temporal redundancy at the feature or token level; this work relocates the source of redundancy to the spectral structure of diffusion features. The observation is presented in the abstract as an empirical finding that motivates the subsequent caching design.

It proposes SpectralCache, a training-free spectral caching framework that reuses stable singular subspaces and estimates only low-dimensional singular values through linear extrapolation. Unlike acceleration schemes that rely on training or fine-tuning, the method can be inserted into existing world-model inference pipelines without training. The method description comes from the abstract and is a framework-level design statement.

It further exploits spectral consistency between neighboring full-computation features to skip selected expensive backbone evaluations via singular value scaling. Beyond singular-subspace reuse, it adds a mechanism for skipping part of the backbone computation, forming a second source of acceleration. The mechanism is stated in the abstract as a component of the method, without separate ablation data.

Experiments on representative world models show SpectralCache consistently improves inference efficiency while preserving generation quality; on HunyuanWorld-Voyager-13B it achieves 5.22x acceleration while maintaining a WorldScore of 65.90 for static scenes, substantially outperforming existing training-free caching methods in inference efficiency. It reports concrete acceleration factors and a quality metric on a specific model, and compares efficiency against existing training-free caching methods. The evidence is the acceleration factor and WorldScore value reported in the abstract for a single model, plus the efficiency comparison against existing training-free caching methods.

The work targets inference acceleration for diffusion-based world models, suited to interactive environment generation where Transformer evaluations repeat during denoising, and it is inserted into existing models in a training-free manner. The concrete validation reported in the abstract centers on static scenes with HunyuanWorld-Voyager-13B, reporting 5.22x acceleration and a WorldScore of 65.90, and compares inference efficiency against existing training-free caching methods. Natural next steps include extending spectral caching to more world models and dynamic scenes, and applying the combination of singular-subspace stability and singular-value extrapolation to other diffusion-based generation pipelines.

The visible text is abstract-level content without experiment tables, ablation settings, baseline lists, or statistical details, so the stability of the acceleration factor across models, scenes, and quality metrics cannot be judged, nor can the scope of the singular-subspace stability observation be confirmed. Open questions include how the spectral-consistency backbone skipping affects generation quality as the skip ratio varies, and how the method behaves on dynamic scenes and longer denoising step counts.