WorldDynCache: Risk-Controlled Latent Dynamics Approximation for Diffusion World Model
2026-08-03 • Machine Learning
Machine LearningComputer Vision and Pattern Recognition
AI summaryⓘ
The authors address the problem of slow predictions in diffusion world models caused by repeated transformer use. They introduce WorldDynCache, a system that smartly estimates and controls errors from skipping some computation steps. This system uses two main parts: one predicts error risks over time, and another approximates changes without running the heavy model repeatedly. Their method speeds up predictions significantly on two models while keeping or improving quality compared to other shortcuts.
diffusion world modelstransformerlatent dynamicscaching methodsrisk estimationlatent transitionapproximation errormodel inference speedHunyuanVoyager-13BAether-5B
Authors
Leyang Chen, Junyi Wu, Shaoqiu Zhang, Yulun Zhang
Abstract
Diffusion world models generate high-quality futures, but re- peated transformer evaluations make inference prohibitively slow. Existing caches reuse intermediate features, selectively update tokens, or reuse and extrapolate denoising outputs ac- cording to local drift or short native-space histories. These criteria can miss both approximation-induced latent transition defects that accumulate across skipped steps and phase- or condition-dependent changes in the direction of latent evo- lution. We propose WorldDynCache, a risk-controlled latent dynamics approximation framework with two core compo- nents. First, a lightweight latent-transition risk estimator tracks the accumulated future impact of approximation defects and calibrates its predictions against counterfactual defects ob- served at exact anchors. Second, a condition- and phase- aware lifted latent surrogate approximates latent evolution without extra transformer evaluations. On HunyuanVoyager- 13B and Aether-5B, WorldDynCache achieves 4.92 times and 2.15 times speedups, respectively, while attaining the best gen- eration quality among the compared caching methods across WorldScore, PSNR, SSIM, and LPIPS.