

Ziqi Ma - LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation
Camera-controlled video models are rapidly advancing toward long generation horizons and complex camera control. A key failure mode is 3D inconsistency: as the camera moves, objects lose permanence and scene structures shift. Existing post-training techniques, which assign a single scalar reward to the entire generation, are poorly suited to correcting these inconsistencies over long horizons. Ma et al. introduce LoGo, which blends global and spatially localized rewards for camera-controlled video models. The local reward provides fine-grained credit assignment, which substantially improves 3D consistency, while the global reward preserves camera following and video quality. Across three base models, LoGo shows a clear advantage on DL3DV and TrajectoryBench, a new benchmark for long-horizon, complex-camera-control generation that current evaluations lack. LoGo effectively reduces local object shifts, artifacts, and global scene changes, illustrating the importance of credit assignment in post-training video models.
Bio: Ziqi is a fourth-year PhD student at Caltech. Her research centers around video-based world modeling: learning spatial and physical priors from videos. Her works enable today's generative video models to capture physical properties, such as consistency and accurate dynamics. Ziqi has interned at World Labs and Meta SuperIntelligence Labs. She is advised by Georgia Gkioxari and Yisong Yue.