WorldDynCache: Risk-Controlled Latent Dynamics Approximation for Diffusion World Model

2026-08-03

AI Summary

Diffusion-based world models suffer from low inference efficiency, and existing caching methods struggle to balance speed and generation quality due to either neglecting the dynamic dependencies between conditions and denoising stages or accumulating approximation errors. To address this, this work proposes a risk-controlled latent dynamic approximation framework that achieves accurate modeling of latent dynamics without additional Transformer calls. The approach integrates a lightweight risk estimator with counterfactual anchor points to calibrate approximation errors and introduces a condition- and stage-aware latent surrogate model to capture dynamic evolution precisely. Evaluated on HunyuanVoyager-13B and Aether-5B, the method achieves speedups of 4.92× and 2.15×, respectively, while outperforming prior caching strategies across multiple metrics, including WorldScore, PSNR, SSIM, and LPIPS.

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.

Problem

• diffusion world model
• inference acceleration
• latent dynamics approximation
• caching
• risk control

Innovation

• risk-controlled approximation
• latent dynamics
• diffusion world model
• caching mechanism
• surrogate modeling