This topic investigates the application of process reward models (PRMs) to video generation tasks, specifically localized video editing and object trajectory control.
This topic investigates the application of process reward models (PRMs) to video generation tasks, specifically localized video editing and object trajectory control. The PRM pipeline aims to verify intended edits or motions while preventing unintended changes, leveraging an origin-aware failure analysis across perception, planning, and generation stages. We want to study whether generated objects follow user-specified paths while preserving identity. For example, MagicMotion focuses on trajectory-controlled image-to-video generation using masks, bounding boxes, and sparse boxes, and introduces benchmarks for trajectory-control accuracy. Also, localized video editing where the task has a clear edit target and a clear “do not change anything else” constraint.