Google’s DeepMind division has developed the Genie artificial intelligence model that can transform images into video games. It allows creating game worlds for platformers in just a few simple steps.
The Genie model is relatively small with 11 billion parameters. It has been trained on more than 200 thousand hours of video of people playing two-dimensional platformers. Such games are quite formulaic, so it is not surprising that Genie has figured out the mechanics and physics of the actions associated with them. The training was quite effective even though the video streams did not contain information about when a button or control was pressed.
As a result, Genie takes a single image (a photo, a sketch, or an AI-generated image) and turns it into a game that can be played by responding to user controls. In fact, the image is transformed into a rudimentary interactive environment in one step.
I am really excited to reveal what @GoogleDeepMind's Open Endedness Team has been up to 🚀. We introduce Genie 🧞, a foundation world model trained exclusively from Internet videos that can generate an endless variety of action-controllable 2D worlds given image prompts. pic.twitter.com/TnQ8uv81wc
— Tim Rocktäschel (@_rockt) February 26, 2024
So far, we should not expect high quality games from the model. Genie — is a research project, not a final product. The model was trained on videos with an ultra-low resolution of 160×90 pixels and a frequency of only 10 frames per second. Consequently, it generates «games» with the same low resolution that run for only 16 seconds at 1 frame per second.
However, the basic concept has been proven, and there is every indication that Genie will improve significantly with scaling. It just needs to use longer, higher-resolution videos and additional computing power.
Source: newatlas