DeepMind Releases Genie 3: New Model Creates Interactive Worlds in Real Time
Credit: Google
Google DeepMind has unveiled the third generation of its Genie “world model.” The new Genie 3 can generate interactive simulations in real time based on a text query or image. The model creates detailed virtual worlds with a resolution of 720p at 24 frames per second.
Genie 3 looks set to be a significant step up. The developers claim that users will be able to create worlds that support “several” minutes of continuous interaction. Genie 3 can hold generated objects in memory for about a minute, allowing the user to return to them and find them in the same place.
Genie 3 introduces what it calls “suggested world events,” which allow you to change weather conditions or add new characters in real time. Google says that there are currently limited ways to interact with generated worlds, and readable text “often only appears if it’s present in the original world description.”
For now, Genie 3 remains a closed research tool. DeepMind plans to open access to a select group of experts and researchers to further refine the technology, with the prospect of more widespread access in the future.
