Google DeepMind: AI generation of video games presented

There has been impressive progress in the field of artificial intelligence over the past year. AI tools have become an integral part of modern life and are finding their way into browsers and other programs. ChatGPT was considered the top dog here, but there are now a whole range of strong alternatives.

Just recently, OpenAI presented the impressive generative AI model “Sora”, which can generate pretty impressive videos from short texts. Google DeepMind is now taking the next step: generative AI can even be used to create video games from sketches, images or real photos.

Google DeepMind creates video games using AI

Google DeepMind has presented the next step in the field of generative artificial intelligence. The new AI tool is called “Genie”, which stands for “Generative Interactive Evironments”. According to DeepMind, the tool is able to create an endless variety of playable worlds.

Genie can be provided with images that the tool has never seen before, without the AI having to be trained beforehand. How does this work? Genie was fed with a huge database of publicly accessible videos of games from the Internet.

According to the corresponding paper published on arXiv, Genie has watched almost 200,000 hours of gaming video material for training purposes. 30.000 hours of gameplay from 2D platformer games laid the foundation for building the worlds within the AI games.

The AI internalized all by itself which actions lead to the character changing its position on the screen. Genie then generates further images based on the previous actions. In the game, each action is therefore performed gradually, resulting in animations and movement sequences.

Genie only needs a single image to create a complete interactive environment. The tool recognizes hand-drawn sketches by children as well as real photos or images generated with the help of other AI tools.

Currently limited to 2D platformers

However, Genie by Google DeepMind is still in its infancy. So far, the tool is only able to develop 2D platformer titles in the style of Super Mario Bros. In addition, the titles currently run very slowly, with only one frame per second, as the generative AI has to calculate each action individually.

However, the results are already impressive. Especially as the method should be generally valid and easily scalable, which would enable higher frame rates and games in completely different genres and dimensions in the future.

Genie is currently only an internal research project for Google DeepMind and is not intended for publication. However, the team does not want to completely rule out the possibility that the tool could find its way into game development in an adapted and expanded form at some point.