Title: OptiWorld: Achieving Optimal Control in Video World Generation Through Physical Constraints

Abstract: While video generation models are evolving into scalable world models, they predominantly produce plausible motion rather than actively controlling or optimizing the fundamental dynamics. Consequently, objects within generated videos often adhere to trajectories that are physically inconsistent, inefficient, unsafe, or lacking in smoothness. To address this, we introduce OptiWorld, a framework that integrates classical optimal control into video generation during the inference phase. The OptiWorld process begins by extracting a compact world state relevant to the specific task, followed by the planning of an optimal trajectory that respects physical constraints. Finally, the video is rendered based on this planned trajectory. We define planning as a geometric problem situated on a continuous manifold, thereby unifying 3D geometry with task-specific physical constraints into a single planning geometry. By incorporating this optimal-control layer, OptiWorld produces videos with superior dynamics, showing significant promise for applications such as goal-conditioned image-to-video generation, video dynamics editing, and counterfactual generation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC