Google researchers have developed GameNGen, an AI system capable of simulating the classic video game DOOM in real-time, achieving over 20 frames per second with visuals nearly indistinguishable from the original.
This breakthrough raises significant questions about the future of video game development.
GameNGen operates on a single chip, predicting each frame using a diffusion model, and produces playable gameplay at 20 frames per second. Trained on 900 million frames of gameplay data, the AI generates three-second clips that playtesters found almost identical to the actual game.
However, the researchers acknowledge the challenge of collecting human gameplay data at scale. To address this, they initially trained a reinforcement learning agent to play the game, recording its actions and observations to create the training data for the generative model.
Running on a single TPU, GameNGen manages DOOM’s 3D environments and fast-paced action without relying on traditional game engine components. In tests, human raters struggled to differentiate between AI-generated clips and the original game.
“GameNGen is the first AI model that can generate a complex and playable video game in real-time without any underlying real game engine,” the researchers noted. “We’re at a fascinating time where soon, AI will be able to create entire games on the fly, personalized to each player.”