LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation
The paper’s fix is to stop scoring long generated videos as one undivided output.
LoGo combines a global reward with spatially localized rewards for camera-controlled video models. The local reward is meant to catch object shifts, artifacts, and scene changes as the camera moves, while the global reward keeps camera following and overall video quality intact. The authors report gains across three base models on DL3DV and a new long-horizon benchmark, TrajectoryBench. Source: ArXiv · AI/CL/LG's note.
LoGo combines a global reward with spatially localized rewards for camera-controlled video models. The local reward is meant to catch object shifts, artifacts, and scene changes as the camera moves, while the global reward keeps camera following and overall video quality intact. The authors report gains across three base models on DL3DV and a new long-horizon benchmark, TrajectoryBench. Source: ArXiv · AI/CL/LG's note.
score 5