Diffusion for World Modeling: Visual Details Matter in Atari

https://diamond-wm.github.io/

DIAMOND: Atari 100k mean HNS 1.46. CS:GO world model from 87 h of human play, about 10 fps on an RTX 3090, trained 12 days on an RTX 4090; allows repeated jumps the real game forbids.

Diffusion and autoregressive neural game engines