Google Research has unveiled a “game engine” called GameNGen that creates games using multiple generative neural networks. The program has already recreated levels from the first DOOM.
GameNGen is a neural model-based game engine capable of creating complex, high-quality interactive scenes.
GameNGen can interactively simulate Doom gameplay at 20 fps. When predicting the next frame, the peak signal-to-noise ratio reaches 29.4 (comparable to lossy JPEG compression).
GameNGen was trained in two stages: the first involved an AI agent playing Doom (sessions were recorded), and the second involved training a diffusion model to generate the next frame based on the sequence of previous frames and input commands. In other words, GameNGen does not generate a game on the fly, but only reproduces what it has already seen. Instead of rendering, the neural model creates a sequence of frames that changes based on the player’s actions.
GameNGen is currently being demonstrated in the game DOOM, but the creators plan to apply the technology to other games and interactive software systems, highlighting the potential for wider use in the gaming industry.