Wizard:透過智能體控制即時操控任意世界
Wizard: Manipulate Any World in Real Time via Agentic Control
Wizard 提出一套透過智能體控制即時操控遊戲世界的方法,將角色動畫、場景操控及遊戲邏輯與視覺生成解耦。研究構建了涵蓋 750 個場景及 178 個角色、共 50.3 小時的代理-影片配對資料。透過因果適配及少步蒸餾,系統可在 24 FPS 下互動,並零樣本泛化至未見場景、角色及動作。
🪄 Wizard: Manipulate Any World in Real Time via Agentic Control!
Existing game world models produce realistic visuals but remain constrained by familiar environments and predefined motions. Control largely focuses on locomotion, with limited support for articulated actions. Can a world model follow an unseen character swinging a sword, spreading its wings, or transforming mid-battle?
Wizard decouples character animation, scene manipulation, and gameplay logic from visual generation. Agents control articulated skeletons and scene geometry through structured visual proxies, while explicit world states support adaptive boss behavior and mechanics-consistent combat.
We construct 50.3 hours of aligned proxy–video pairs spanning 750 scenes and 178 characters. Character-aware adaptive masking preserves reference appearance, while causal adaptation and few-step distillation enable interaction at 24 FPS, with zero-shot generalization to unseen scenes, characters, and motions.
From world exploration and character summoning to transformation and scene manipulation, Wizard gives players greater freedom to create and control their own game worlds.
Code coming soon!
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co