Google's DeepMind division has developed an artificial intelligence model, Genie, that can transform images into video games. It allows you to create platformer game worlds in just a few simple steps.
The Genie model is relatively small, with 11 billion parameters. It is trained on more than 200 thousand hours of video of people playing 2D platformers. These types of games are pretty formulaic, so it's no surprise that Genie figured out the mechanics and physics behind them. The training was quite effective even though the video streams did not contain information about when the button or control was pressed.
As a result, Genie takes a single image (photo, sketch, or AI-generated image) and turns it into a game that can be played in response 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
You shouldn't expect high quality games created from the model just yet. Genie is a research project, not a final product. The model was trained on ultra-low resolution video of 160×90 pixels and a frame rate of only 10 frames per second. So it generates “games” with the same low resolution that only run for 16 seconds at 1 frame per second.
Book illustration. Visualize the stories, characters and world of books. Registration for the course
However, the basic concept has been proven, and there is every indication that Genie will improve significantly as it scales up. All you need to do is use longer, higher resolution videos and more processing power.