Reproduce “GameDirector: Decoupling Gameplay Logic from Rendering for Player-Configurable Game World Models” on OpenResearch

About this paper

Zijun Lin, Zhiyang Deng, Yuzhe Wu, Bihan Wen, Yeying Jin

GameDirector separates combat rules and state tracking from video generation, using visual perception and a language-model planner to control configurable fights while a video model renders them. In reported evaluations, it achieved 98.5% and 99.6% mechanics-fidelity scores and outscored the strongest tested baselines on boss action decisions; its 100% state-alignment result reflects consistency with the same hit detections used by its tracker.

Key insights

• Separating rule enforcement from rendering lets a generative video model provide visual flexibility without being responsible for maintaining exact gameplay state.
• The planner can adapt boss choices to player distance and orientation, while deterministic checks constrain its proposals to actions allowed by the configured mechanics.
• The reported 100% state alignment is conditional: the evaluation and state tracker rely on the same AttackNet hit signals, rather than independent ground-truth game state.

Summary by alphaXiv.