CoCode-World: Generating two-player-view videos with a shared 3D proxy
What happened
Qirui Li and co-authors propose the CoCode-World framework, which separates structural control from visual synthesis to generate videos that stay consistent across two players' viewpoints. A programmable proxy specifies scene layout, motion and occlusion, while RGB references guide the appearance of characters and scene, so the two kinds of information no longer crowd into one stage and constrain each other. At generation time, paired proxy videos, appearance references and target viewpoints are concatenated under a unified spatial layout, and a pretrained video model jointly produces both views in a single generation pass, learning through LoRA to follow the proxy and to keep appearance corresponding across views. The authors report that they validated the framework experimentally.
AI-written from the coverage · updated 9 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game AIPickCoCode-World proposes shared 3D proxy generation framework for two-player world models
CoCode-World proposes a shared 3D proxy framework that separates structural control from visual synthesis to generate videos with consistent viewpoints for two players. Programmable proxies specify scene layout, motion and occlusion, while an RGB reference guides character and scene appearance; paired proxy videos, appearance references and target viewpoints are tiled in a unified spatial layout, and a pretrained video model jointly produces both views in a single generation, with LoRA used to learn proxy adherence and cross-view appearance correspondence.