HKU and Tencent Introduce SCoPE for 3D-Aware Video World Models
jiqizhixin · x · 2026-08-29
HKU, HKUST, and Tencent ARC Lab present SCoPE (Sightline-Coordinate Positional Encoding), improving positional encoding in video DiTs.
Core Improvement:
- Replaces standard tensor-grid positional encoding with ray-space coordinates.
- Computes observation rays based on camera motion for each token, enabling the model to reason about 3D spatial relationships instead of 2D pixel neighborhoods.
- Adds less than 0.1% new parameters.
Results: Significantly improves camera control, cross-view consistency, and revisit consistency. As model scale grows from 5B to 14B, SCoPE's lead widens, with translation error advantage growing from 12% to 25% and FVD improvement from 23% to 47%.
More from Research
- STEMBench launches: 6 STEM domains, zero full-target successes across all runs — mariofilhoml · 2026-10-15
- OpenAI gives $5M to Simons Foundation's $15M push to help mathematicians navigate AI-era results — OpenAI · 2026-10-15
- DeepPhylo infers phylogenetic trees without unrealistic evolutionary assumptions — anshulkundaje · 2026-10-15
- Wander preprint explores graph foundation models via shared objective and random-walk interface — mmbronstein · 2026-10-15
- CHI Reviewer: AI Slop Is Real, but Unusually Rigorous Papers Still Getting Through — IanArawjo · 2026-10-15
- Into the Unknown: CMU benchmark finds 114 of 135 AI-written research papers fail integrity checks — arankomatsuzaki · 2026-10-15