Tencent's Wizard is a research paper on a game world model that generates video from character motion controls in real time, and it claims to work on scenes and characters it never saw in training.
Method: Skeleton and geometry proxies condition a video model, with adaptive masking to keep the video aligned with them.
Speed and data: The paper claims 24 FPS, trained on 50.3 hours of proxy-video pairs across 750 scenes and 178 characters.
Game logic: An agentic framework tracks game states so players can configure worlds and control characters.
Missing: The code is "coming soon."
Swords, wings and mid-battle transformations, all on the table.
Sign in to comment.
TL;DR of the paper, generated automatically.
Tencent's Wizard is a research paper on a game world model that generates video from character motion controls in real time, and it claims to work on scenes and characters it never saw in training.
Swords, wings and mid-battle transformations, all on the table.