LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation
LoGo combines global and spatially localized rewards to fix 3D inconsistency in camera-controlled long-horizon video generation across three base models.
LoGo is a post-training method for camera-controlled video models that blends a global reward preserving camera following and video quality with a local reward giving fine-grained credit assignment for 3D consistency. Evaluated across three base models on DL3DV and a new TrajectoryBench benchmark for long-horizon complex camera control, it reduces local object shifts, artifacts, and global scene changes. The work highlights the importance of credit assignment in post-training video generation models.
- Local rewards improve 3D consistency; global rewards preserve camera following and quality
- Evaluated on three base models using DL3DV and new TrajectoryBench benchmark
- Reduces object shifts, artifacts, and global scene changes in long-horizon generation
- Organizations
- TrajectoryBench
From the original article
1 min read at arxiv.orgCamera-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.
By Ziqi Ma · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.