Google DeepMind introduced Genie 4 on Tuesday as a research preview for selected testers. The AI system, known as a world model, creates virtual environments that people can explore and change as they are generated. Google says it offers more realistic scenes, human characters, and physical movement than Genie 3.
Users can type instructions during an experience to change the weather, add people or objects, or adjust the camera angle. The model also generates matching audio, including music, background sounds, and dialogue. Gemini 3.8 Flash helps guide the process by planning steps and checking the images that Genie produces.
Genie 4 can maintain a consistent scene for about two minutes, compared with one minute for its predecessor. Potential uses include planning film scenes, developing early game ideas, and creating training environments for robots. In a robotics experiment, another Gemini model planned tasks that Genie 4 then simulated.
However, objects and details can change unexpectedly during longer sessions, and physical behavior is not always accurate. The team also faces challenges with accurate sound and readable text within scenes. Google plans to expand access to Google AI Ultra subscribers through Project Genie, plus selected partners through the Gemini API and Google AI Studio.
