Google DeepMind's Genie 4 gives generated worlds sound and keeps them coherent for two minutes
The new world model builds interactive environments frame by frame, now with audio and steering while you play.
Google DeepMind on Tuesday presented Genie 4, the successor to the world model that rattled gaming stocks last year. Such a model generates a virtual environment in real time from a text prompt, without pre-built levels. According to TechCrunch, a scene now stays coherent for up to two minutes, twice as long as with Genie 3, and the model produces matching audio on the fly, from music to dialogue.
Users can change the weather or camera angle during the simulation, or add people and objects. Ars Technica writes that the improvements rely partly on Gemini, which acts as a planner breaking down tasks while Genie renders the images. DeepMind sees uses in gaming, education and training robots in endless virtual environments.
For now only selected testers can use the model. Google AI Ultra subscribers and developers with early access to the Gemini API will follow in the coming weeks, TechCrunch reports. Google acknowledges that details still drift the longer a simulation runs.