Google DeepMind has introduced Genie 4, a research-preview AI model that converts text prompts into photorealistic, three-dimensional virtual worlds that users can explore and modify while they are running.
Unlike conventional AI video systems, Genie 4 generates interactive environments rather than fixed clips. Users can type instructions during a simulation to alter the camera position, create weather effects such as rain or add new characters without stopping and restarting the world.
Live sound and longer simulations
The system also produces audio in real time to correspond with events inside the generated environment. This can include ambient sounds, music and virtual characters speaking or responding to user inputs.
Genie 4 can run simulations for up to two minutes, double the duration offered by Genie 3 last year. Google has also said the updated model delivers more realistic physics and characters with more human-like behaviour.
Gemini helps plan the virtual environments
Genie 4 works with Google’s Gemini AI, which functions as a planning layer for the system. Gemini can determine the logic and sequence of actions, while Genie 4 generates and updates the visual simulation.
Tejas Kulkarni, a research director at Google, said during a briefing that future AI systems could increasingly combine models in this way, with Gemini producing text or code and Genie simulating the results in real time.
The approach is also being explored for robotics. In one experiment, a Gemini 4 Argon model planned tasks that Genie 4 subsequently simulated, demonstrating how robots could potentially be trained in virtual environments before being deployed for real-world testing.
Potential applications
Google expects Genie 4 to have applications in gaming, interactive education and robotics. Google DeepMind engineer Diego Rivas also identified filmmaking as a possible use, particularly for rapid storyboarding that allows creators to test ideas without waiting for lengthy video renders.
Google already provides Veo for producing high-definition video clips. In the future, developers may be able to use Gemini Omni, Google’s multimodal model, to convert Genie 4 prototypes into finished creative assets.
Early limitations and access
The technology remains at an early stage. Rivas said extended simulations can cause the generated world to “drift”, with objects losing detail or disappearing. Google’s team is also continuing to improve audio synchronisation and the rendering of readable text within the virtual environments.
Genie 4 is currently restricted to a small group of trusted testers. Google plans to make it available in the coming weeks to Google AI Ultra subscribers through Project Genie, which costs $100 a month. Early-access partners will also receive access through the Gemini API and Google AI Studio.
Follow us
