Computer Science
Diffusion for World Modeling: Visual Details Matter in Atari
E. Alonso, A. Jelley, et al.
DIAMOND (DIffusion As a Model Of eNvironment Dreams) trains reinforcement learning agents inside a diffusion-based world model that preserves richer visual detail, boosting performance and safety. It achieves a new world-model best on Atari-100k with a mean human-normalized score of 1.46 and can function as an interactive neural game engine after training on Counter-Strike: Global Offensive. The research was conducted by Authors present in <Authors> tag. Code, agents, videos and playable models are released at https://diamond-wm.github.io
Related Publications
Explore these studies to deepen your understanding of the subject.
