Genie: Generative Interactive Environments
AI summary of the linked articleGenie, a foundation world model from Google's Genie Team, is trained on publicly available Internet videos, mainly 2D platformer games and robotics, and can generate playable, action-controllable worlds from a single prompt image such as a photograph or sketch. The model learns controls without action labels, inferring latent actions that behave consistently across generated environments. The team says a smaller 2.5B model trained on RT1 robotics videos showed similar consistency, and that the approach could help train generalist AI agents. The article gives no quantitative benchmark results beyond these examples. |
Recommended by 2 curators |
5 comments That's awesome - soon, everyone will be able to create their own mods. Given that two of the world's most popular games (League of Legends & Counter-Strike) were fan-made mods, it might unlock many new gaming ideas. frotograf.ethSubmitter two things here: - i'd scroll down and look at the drawings being brought to life, this TBH is more interesting for me from a designers perspective, than the above examples - I'd go with a hot take that this solution is not viable in the long run (eg. in the early stage creating a whole game based on 1 drawing is gonna be a very low or no quality) BUT might be interesting with a smart gaming mechanic. Imagine a game level generated in this way that can be accessed with a special NFT ;) Token-gated levels and skins would be nice, especially if they're fan-made and the gaming company would share % of revenue with them :) |
More from sites.google.com on Kiwi News |
