Video world models are a promising interface for interactive simulation, generating long visual rollouts from actions and camera motion. However, making these rollouts responsive enough for real-time interaction is challenging, especially because coherent generation depends on retaining visual information over time.
The paper introduces Velox, a method that uses spatial memory to make world models more efficient. By explicitly storing spatial context, Velox aims to reduce the computational burden of maintaining coherence during generation, potentially enabling faster responses without sacrificing visual consistency.
The abstract highlights the core tension between responsiveness and coherence, but does not provide quantitative results or implementation details in the available text. The work appears to target a practical bottleneck in deploying video world models for interactive applications, though further reading of the full paper would be needed to assess its claims.