Rover: Interaction-Preserving Video Reshooting

Extending video reshooting from general scenes to object interactions requires preserving contact, motion, and object state across viewpoints. However, existing synchronized multi-view datasets offer limited coverage of interaction dynamics under diverse camera trajectories. We present Rover for interaction-preserving video reshooting. Our approach distills interaction and dynamics priors from a pretrained video generator through ObjAlive, a collection of 13.5K filtered videos produced via image-to-video generation. The collection covers contact-rich manipulation, fluid motion, deformation, topology changes, articulation, and appearance changes. We combine it with synchronized multi-view and real single-view videos to fine-tune a video diffusion model conditioned on source videos and target-view warps. Blinded four-way human evaluation favors Rover, which receives 71.6% and 63.6% of interaction-fidelity votes on held-out human–object interaction and robotic manipulation videos, respectively. These results indicate generalization to robotic manipulation without robot-specific fine-tuning. A controlled, step-matched ablation demonstrates the benefit of ObjAlive for preserving contact and non-rigid dynamics.