HKU Researchers Propose SCoPE for Ray-Space Video World Models
机器之心 · wechat · 2026-08-24
HKU, HKUST, and Tencent ARC Lab introduced SCoPE (Sightline-Coordinate Positional Encoding), shifting Video DiTs from pixel grids (u,v,t) to a ray-space coordinate system based on camera rays.
Core Mechanism:
- Ray Coordinates: Pre-calculates observation rays (Plücker coordinates) for each token based on camera trajectory and injects them into attention queries/keys.
- Geometry + Content: The model simultaneously checks content similarity and ray intersection to implicitly establish 3D correspondences.
Results:
- Interactive Control: Generates consistent camera movements from a single image using WASD commands without 3D reconstruction.
- Revisit Consistency: Significantly improves consistency when the camera returns to the starting point (LPIPS Recovery boosted).
- Efficiency: Adds <0.1% parameters; performance gains scale with model size.
More from Research
- Trained two 16M-param models to do generative CAD with real physics — debreuil · 2026-08-24
- Claude model helps discover complex structure on S^6, solving 60-year-old math problem — Singularitarian · 2026-08-24
- Study: Agents read instructions/notes 60.5% of the time, rarely touch API docs — dair_ai · 2026-08-24
- Claude Verifies 43 Lean Modules autonomously, Tackling Theoretical Physics — Tkaraletsos · 2026-08-24
- AI fakes memory: why it gets confidently wrong without forgetting — PrajwalTomar_ · 2026-08-24
- Google's AI research agents discover 66 novel biomarkers in automated biomedical study — imjustnewatai · 2026-08-24