Powered by Gemini 3.8 Flash, Genie 4 generates two-minute interactive worlds with sound, and accepts natural language instructions during simulation. Access reserved for the moment to testers.
Google DeepMind has presented Genie 4, the new version of its world model capable of generating interactive environments from a description. Four new features are highlighted: a more realistic rendering, the ability to modify the experience in real time with a simple written instruction, the generation of a sound accompaniment, and the support of Gemini 3.8 Flash to control the simulation. Sessions can now last about two minutes, compared to only one with Genie 3.
Google is targeting uses such as cinema previews, video game prototyping or the creation of training scenarios for robots. The company acknowledges, however, that on the longest simulations, the details and permanence of objects can drift.
Genie 4 is first offered as a "search preview" to a circle of trusted testers. It is expected to arrive "soon" for Google AI Ultra subscribers, and a waiting list is open to try it in Project Genie.
Read the full article on the original source
Go to the source →