We present GameNGen, the first game engine powered entirely by a neural modelthat enables real-time interaction with a complex environment over longtrajectories at high quality. GameNGen can interactively simulate the classicgame DOOM at over 20 frames per second on a single TPU. Next frame predictionachieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters areonly slightly better than random chance at distinguishing short clips of thegame from clips of the simulation. GameNGen is trained in two phases: (1) anRL-agent learns to play the game and the training sessions are recorded, and(2) a diffusion model is trained to produce the next frame, conditioned on thesequence of past frames and actions. Conditioning augmentations enable stableauto-regressive generation over long trajectories.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)