AI Dynamics

Global AI News Aggregator

About

SCoPE Uses Ray-Space Position Encoding for Video World Models

What if video world models stopped thinking in pixel grids? The University of Hong Kong, HKUST, and Tencent ARC Lab present SCoPE. SCoPE (Sightline-Coordinate Positional Encoding) replaces the standard tensor-grid positional encoding in video DiTs with ray-space coordinates.

→ View original post on X — @jiqizhixin