Reproduce “WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weilun Feng, Guoxin Fan, Haotong Qin, Mingqiang Wu, Yuqi Li, Xiangqi Li, Zhulin An, Libo Huang and 5 more
The WorldCache framework accelerates diffusion-based world models by employing heterogeneous token caching and adaptive skipping, tailored to multi-modal data. It achieves up to 3.65x speedup on HunyuanVoyager-13B for world generation, maintaining over 98% of baseline quality, and a 2.61x speedup on Aether-5B for 3D reconstruction with high fidelity.
Key insights
- Tokens in multi-modal diffusion world models exhibit significant heterogeneity in their temporal dynamics and predictability, requiring adaptive, token-specific caching strategies.
- The local non-linearity of a token's temporal trajectory, quantified by a curvature score, is an effective indicator for its predictability and guides the application of appropriate caching rules.
- Monitoring the accumulated drift of a small subset of 'chaotic' (hard-to-predict) tokens using a dimensionless, curvature-normalized score provides a robust mechanism for adaptive re-evaluation, preventing divergence while maximizing skipping.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.