Cast speeds up interactive world models while keeping scenes consistent
CAST: Reconstruction-Coupled Acceleration of Interactive World Models
Computer Vision and Pattern Recognition
Summary
Interactive world models have to update quickly when you control them, but it’s hard to keep the scenes looking right when speeding up the process. The authors found that changes caused by interaction relate to where errors happen, and some errors change in predictable ways. They created CAST, a system that picks important reference points and uses smart methods to fix missing details, allowing faster updates without losing quality. Tests show CAST runs over twice as fast as normal methods while keeping scenes looking good and responding well to controls.
What this means in practice
- •For game developers: Speed up rendering of interactive game worlds while maintaining visual quality and responsiveness.
- •For virtual reality engineers: Improve real-time updates in VR environments to keep scenes visually consistent during user interaction.
Authors
Leyang Chen, Junyi Wu, Fanqing Kong, Shaoqiu Zhang, Yulun Zhang
Abstract
Interactive world models must respond quickly to controls while preserving scene consistency. Existing acceleration methods can miss heterogeneous control responses and spatial transport when recovering skipped features. We observe that interaction-induced feature changes correlate with approximation error, while low-frequency interpolation errors are phase-sensitive and show more predictable phase progression. These findings motivate CAST, a reconstruction-coupled inference framework. CAST selects anchors by interaction sensitivity and cross-layer coverage, reconstructs skipped residuals with frequency- and confidence-aware Phase-Aware Reconstruction (PAR), and coordinates historical KV routing according to downstream reconstruction responsibility. On Matrix-Game 3.0 and HY-World 1.5, CAST achieves 2.15x and 3.48x speedups, respectively, while maintaining visual quality close to Native (Figure 1). It also attains the highest VBench scores among compared methods and leads non-native baselines on seven and six of thirteen WorldMark dimensions, demonstrating a balance of generation speed, visual quality, and interactive responsiveness under real-time control. Code is available at https://github.com/lokiniuniu/CAST.