[2G4-OS-47a] OS-47

1:45 PM - 2:00 PM JST(4:45 AM - 5:00 AM UTC)

[2G4-OS-47a-02] World Handoff Consistency for Multi-view Driving Video Generation

〇Bum Jun Kim¹, Makoto Kawano¹, Yusuke Iwasawa¹, Yutaka Matsuo¹ (1. The University of Tokyo, Graduate School of Engineering)

Keywords:

World Model,Physical AI,Video Generation,Multi-View,Autonomous Driving

Multi-view driving video generation must preserve object continuity across front and rear views, not only visual realism. This paper proposes World Handoff Consistency, WHC, a self-supervised metric set that measures temporal, worldline, epipolar, appearance, and reprojection consistency for front to rear handoff. On 57 generated videos, MatchPair averages 0.481; MagicDrive-V2 and InstaDrive exceed 0.72 while DriveDreamer2 is 0.17, showing that visually plausible videos still suffer from missing or misplaced handoffs. These results indicate that WHC complements Physical AI evaluation of worldline coherence.