Anthony Hu, Václav Volhejn, Adrien Ramanana Rahary, Chris Mulder, Aditya Makkar, Amélie Royer, Manu Orsini, Alyx Liao et al.
This paper proposes a latent diffusion model that generates real-time multiplayer game scenes in complex physical environments, moving beyond single-player world models by modeling the interactions of multiple agents.
Existing single-player world models treat other agents as part of the environment, making it difficult to accurately model the dynamics of complex multiplayer environments where players interact rapidly and tightly.
A 5-billion-parameter latent diffusion model is developed, trained on 10,000 hours of gameplay. The study systematically investigates key design choices: the video codec, the generative objective, and the multiplayer conditioning scheme.
The model generates four-player matches in real time at 20 frames per second on a single Nvidia B200 GPU, with distributional quality holding steady out to five minutes, far beyond the training horizon. The dataset, full training and inference codebase, and a live demo are released to support further research.