STRIKE: Learning Visual State Transitions for Physical World Modeling

Physical world modeling requires predicting how interactions change a scene, not merely generating coherent motion. We propose STRIKE, a framework that explicitly learns visual state transitions and uses the predicted states to condition dense video generation. At its core is an image-based transition model trained on paired visual states to predict the next scene configuration from the current image, a transition specification, and elapsed time. A pretrained vision-language model supplies timed transition specifications, and recursive application of the learned transition model produces a sequence of future visual states. A conditional video diffusion model then generates the full rollout from these states and their temporal locations. This separates supervised state-transition learning from conditional dynamics modeling. Experiments on Physics-IQ Verified, PhyGenBench, Pisa-Experiments, and RoboTwin2.0 show improvements over the corresponding video-backbone baselines in benchmark measures of physical realism and manipulation-video fidelity. These results support learned visual state transitions as an effective intermediate representation for physical world modeling.