Google Research has made significant strides in AI with the development of GameNGen, a groundbreaking AI-based game engine capable of generating original Doom gameplay. Using the Stable Diffusion model, GameNGen processes player inputs and previous frames to create new, visually coherent gameplay elements in real-time.
This achievement marks a significant step forward in AI-driven game development. The AI-generated Doom can be played like the original video game, with players able to turn, strafe, fire weapons, and navigate dynamically generated levels. Even the ammo count remains remarkably accurate. While the current version runs at 20 FPS, it’s often difficult to distinguish it from actual Doom gameplay in short clips.
How they did it
To train GameNGen, the Google team had an AI agent play Doom at various difficulty levels and simulate different player skill levels. The agent was rewarded for collecting power-ups and completing levels but punished for taking damage or dying. This process provided hundreds of hours of visual training data for GameNGen to learn from and recreate.
While GameNGen’s Doom is not yet perfect, it’s a remarkable demonstration of AI’s capabilities. There are still visual glitches and inconsistencies, but the overall experience is surprisingly immersive. The levels generated are dynamic and unpredictable, and the AI’s ability to simulate Doom gameplay is impressive.
What’s next?
GameNGen’s success could pave the way for AI-generated content in other games. By demonstrating the ability to create coherent gameplay using AI, this project could inspire developers to explore similar techniques for generating levels, characters, or even entire game worlds. This could lead to a future where games are not just designed by humans but co-created with AI, offering endless possibilities for new experiences.
