“Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation.”
GameNGen runs DOOM at over 20 frames per second on a single TPU, with no game engine, just Stable Diffusion v1.4 predicting the next frame. The training data came from an RL agent playing the real game, because Google cannot collect human gameplay at scale. Raters judged short clips, not full play sessions. The neural engine still needed the original engine to exist first.