Source URL: https://www.theregister.com/2024/08/28/google_doom_ai/
Source: The Register
Title: Google trains a GenAI model to simulate DOOM’s game engine in real-ish time
Feedly Summary: The proof of concept shows promise despite big limitations
A team from Google and Tel Aviv University have developed a generative AI game engine capable of simulating the cult classic DOOM at more than 20 frames per second because research.…
AI Summary and Description: Yes
Summary: The text discusses a newly developed generative AI game engine called GameNGen, created by a collaboration between Google and Tel Aviv University. It proves significant for AI professionals as it explores the use of reinforcement and diffusion models for real-time game simulation, with implications for future AI applications beyond gaming.
Detailed Description: The development of GameNGen marks a notable advancement in the intersection of generative AI and gaming technology. This project aims to innovate the traditional methods used in game engine creation by integrating AI methodologies. Here are the major points addressed in the text:
– **Generative AI Model**: GameNGen uses reinforcement learning and diffusion models, showcasing how these technologies can be employed to generate game engines that operate in real-time.
– **Training on DOOM**: While it was trained specifically on the cult classic DOOM, the approach is versatile and can be adapted for various other games, indicating a broad application potential.
– **Comparison with Traditional Engines**: Unlike conventional game engines that rely on manual coding and predictable loops, GameNGen dynamically generates game frames based on player actions and previously generated frames.
– **Learning Process**: The initial phase involved creating a reinforcement learning agent that played DOOM. The resulting data fed into a custom diffusion model (based on Stable Diffusion v1.4), which subsequently rendered the frames.
– **Performance Metrics**: Running on a TPU v5, GameNGen achieved approximately 20 frames per second (FPS), with theoretically achievable FPS going up to 50 when compromising on quality.
– **Visual Quality**: Although it produced frames comparable to lossy JPEG compression, the capacity for human raters to distinguish between generated and original clips was near random chance, demonstrating a significant advancement in visual fidelity for generated content.
– **Limitations**: The model currently faces notable restrictions, particularly in memory storage, allowing only a 3-second gameplay retention and incomplete coverage of the original game’s environment, resulting in potential inaccuracies.
Overall, GameNGen serves as a proof of concept, highlighting innovative uses of AI and potential avenues for further research and applications, particularly for professionals interested in the future of AI in gaming and beyond.