Diffusion Models Are Real-Time Game Engines

Researchers from Google Research and Tel Aviv University present GameNGen, the first game engine powered entirely by a neural model, enabling real-time interaction in complex environments. GameNGen can simulate the classic game DOOM at over 20 frames per second on a single TPU, with visual quality comparable to lossy JPEG compression. The engine is trained using both reinforcement learning and a diffusion model, with impressive stability in long simulations.

Visit Original Article →