Google has introduced AI technology called GameNGen, which creates the shooter “Doom” on a TPU in real time, completely without an engine. The resulting AI “Doom” is interactive and playable. It is apparently hard for people to tell the difference from the original game based on video footage.
Advertisement
The game runs at 20 fps, but according to Google, GameNGen can also generate 50 frames per second with no loss of quality.
At its core, GameNGen version 1.4 has a pumped-up version of the static diffusion image generator, which generates frames based on previous images and user input. A neural network manages the state of the game – i.e. remembers how much ammunition has been used, how many life points the character has, which doors have been opened and whether enemies have been killed.
Images, videos, and games from AI
Google’s AI technology creates “Doom” in real time, taking into account player input. First, an AI agent was trained to play the original “Doom” using reinforcement learning. These attempts were recorded. Finally, the diffusion model was trained to generate new frames from previous images and user input. This is reminiscent of Nvidia’s DLSS3 frame generation technology, which is already used in some modern games and inserts an AI frame between two traditionally calculated frames.
Graphic from the Google paper: First, an AI agent plays “Doom” before using video content to train a generative model.
(Image: Google)
Researchers from Google and Tel Aviv University Present your GameNGen model in a paper and released videos of AI-generated games. Their experiments show that neural networks can basically simulate high-quality video games in real time. They conclude in their paper that AI can generate not only images and videos, but also games.
GameNGen only works with “Doom”
But the technology is not perfect. GameNGen can only remember about three seconds of the past. It cannot be guaranteed that game elements such as ammunition, life points and opponents’ positions will remain stable and useful over long sessions. In addition, the AI model was trained specifically on “Doom” and cannot be easily applied to other games, according to the researchers. More modern graphics than “Doom,” released in 1993, can also drastically increase computing costs.
Google writes that people can hardly tell the difference between the AI-generated “Doom” and the original. But it’s worth taking a closer look: in the published video materials, random fluctuations in ammunition levels can be clearly seen. In addition, enemies look very blurry when moving.
There is still a fundamental limitation: GameNGen needs content from an existing video game to be able to recreate it. You cannot develop completely new games with the current approach. Nevertheless, Google says that development studios can use such technology to speed up the development of their games and save costs. You can also use AI technology to change the behavior of existing games and create mods. “For example, we may be able to convert a set of frames into a new playable level or create a new character from sample images without writing code.”
(There)
