World Labs' LoGo uses local-global rewards to fix 3D inconsistency in long-horizon video generation
georgiagkioxari · x · 2026-10-06
Ziqi Ma, a Caltech intern at World Labs, presents LoGo, a post-training method for camera-controlled video generation targeting 3D inconsistency over long horizons (scene layout shifts on revisit, objects appearing/disappearing, geometry drift, artifacts).
- Key idea: prior post-training assigns one scalar reward per video; LoGo blends a global reward (preserving camera following and quality) with spatially localized rewards for fine-grained credit assignment, substantially improving 3D consistency.
- Validated across three base models, showing clear gains on DL3DV and TrajectoryBench, a new benchmark for long-horizon complex camera trajectories.
- Authors note reward design matters in post-training long-horizon video gen, echoing findings in domains like LLM reasoning.
Paper, code, and benchmark are publicly available.
Related event: World Labs Releases LoGo to Improve 3D Consistency in Long Video Generation(2 posts)→
More from Multimodal
- AI-Generated Music Video: A Ship-Born Human Uses Onboard AI to Discover Ancient Earth Music — most_triumphant_yeah · 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
- Suno launches Albums: bundle songs, set artwork, publish full releases — suno · 2026-10-06
- Deemos publishes SIGGRAPH 2026 paper ACT: unified rigging, skinning and animation for 3D — sidahuj · 2026-10-06