LoGo: local-global rewards fix 3D inconsistency in camera-controlled video models

georgiagkioxari · x · 2026-11-20

Caltech PhD student Ziqi Ma will present LoGo at Cohere Labs' open science community (Nov 24, 8am PST), tackling 3D inconsistency in camera-controlled video models. Existing post-training assigns a single scalar reward per generation, poorly suited to long horizons. LoGo blends a spatially localized reward (fine-grained credit assignment, substantially improving 3D consistency) with a global reward (preserving camera following and quality). Across three base models it shows clear gains on DL3DV and a new long-horizon benchmark, TrajectoryBench, reducing object shifts, artifacts, and scene drift. Ma's research focuses on video-based world models capturing physical priors; she has interned at World Labs and Meta SuperIntelligence Labs.

Original post →

More from Multimodal

Multimodal channel →