World Labs' LoGo uses local-global rewards to fix 3D inconsistency in long-horizon camera-controlled video
gowthami_s · x · 2026-10-06
Researchers from World Labs and Caltech (including Li Fei-Fei and Ben Mildenhall) present LoGo, a post-training method that substantially improves 3D consistency in long-horizon, camera-controlled video generation.
Problem: as camera-controlled models extend generation horizons, objects lose permanence — scene layouts shift on revisit, objects appear/disappear or deform, and artifacts emerge. Existing post-training assigns a single scalar reward to the whole clip, poorly suited to correcting errors over long horizons.
Method: LoGo blends a spatially localized reward, giving fine-grained credit assignment that markedly improves 3D consistency, with a global reward that preserves camera following and video quality.
Results: across three base models, LoGo clearly outperforms prior methods on DL3DV and TrajectoryBench, a new benchmark for long-horizon complex camera trajectories that current evaluations lack. Paper, code, and benchmark are released.
Related event: World Labs Releases LoGo to Improve 3D Consistency in Long Video Generation(2 posts)→
More from Multimodal
- Google's KeyRec Achieves Best Long-Video VLM Results With Just 10% of Visual Token Budget — google · 2026-10-06
- One Week of Prompting Opus 5.5 with Blender, ElevenLabs and REAPER Yields a TV-Quality Ad — zeeg · 2026-10-06
- Seedance 2.5 demo: a melancholic 30-second forest of walking trees — LudovicCreator · 2026-10-06
- Midjourney style code 3896456162 blends 1970s Kodak film with disco aesthetics in v8.2 — michaelrabone · 2026-10-06
- Four years ago Google's Phenaki generated multi-minute videos from changing prompts — dumierhan · 2026-10-06
- A Lyrics Video for a Song That Never Existed — Fit-Case1093 · 2026-10-06