Google DeepMind has crossed a new threshold in building real-time virtual universes with AI. The company's newly announced Genie 4 is a world model that produces photorealistic simulations, steerable either through keyboard commands or natural language. For now, it is available only as a research preview.
Built on the foundation of Genie 3, which was shown roughly a year ago, the new model promises unprecedented levels of photorealism and interactivity. The sample content shared so far backs up that claim.
Editing the simulation while it runs
What sets Genie 4 apart from rival tools and earlier versions is the ability to intervene in the environment while the simulation keeps running. Users can switch the weather to rain, view the scene from a different camera angle, or add new people and objects to the world.
All of this happens without pre-built static 3D scenes or hand-coded interactions. As the user moves through the world, frames are generated on the fly. Ambient sounds, including wind and human voices, are also produced live.
Two minutes of consistency and better physics
The model can maintain consistency within a single scene for up to two minutes — exactly double Genie 3's limit. While Genie 3 was announced with 720p resolution and 24 fps as fixed specs, no official resolution has been given for Genie 4; however, the first demo videos reached viewers at FHD resolution and a 24 fps frame rate.
Beyond visual quality, there is notable progress in physics. According to DeepMind, interactions such as collisions and aerodynamics are simulated significantly better than in its predecessor. Human characters added to a scene both move more realistically and respond naturally to environmental changes and user commands.
Talking characters and Gemini orchestration
On the sensory side, built-in audio integration stands out. Genie 4 can generate waves, wind, ambient noise and music in sync with the world. Virtual characters can also hold real-time, interactive dialogues based on user input and events in the world — turning them from walking puppets into speaking actors.
The mental architecture behind the simulation is run by Gemini models. In the workflow Google described, Gemini acts like a film director: breaking tasks into stages, continuously reviewing generated frames, and keeping the scene's logical framework intact. In one shared technical example, Gemini 3.8 Flash served as a routing agent that watched the simulation nonstop. In another experiment, Gemini 4 Argon broke complex tasks into steps and had Genie 4 run the simulation in real time. In a test combining both models, a scenario in which Apptronik's Apollo humanoid robot played checkers was successfully simulated.
Limitations and access plans
The DeepMind team openly acknowledges that the model is at an earlier stage and contains flaws. As simulations run longer, objects can drift and degrade, and too many complex commands issued in quick succession can break continuity.
So who gets to use it, and when? Genie 4 is currently open only to Google's designated "trusted testers." In the coming weeks, subscribers to Google AI Ultra — which starts at $100 per month — will gain access through Project Genie, the interface that lets users create interactive worlds from text and images.
Comments (0)
No comments yet. Be the first!