DeepMind Unveils Genie 4, Adding Sound and Live Steering to Simulated Worlds

Google DeepMind on Tuesday announced Genie 4, the latest version of its world model, a system that generates interactive, navigable scenes from a prompt and responds to a user’s input as the scene runs. Google describes it as its most realistic model of the kind so far and is releasing it as a research preview to a limited group of trusted testers.

The main additions over Genie 3, which appeared as a research preview a little over a year ago, are sound and live control. Genie 4 generates audio to match each world, including ambient noise, music and dialogue, and a user can type a prompt in the middle of a simulation to change the weather, move the camera or add people and objects. Google says physics and the movement of human characters are also improved.

Consistent scenes now last up to two minutes, twice Genie 3’s limit. Google acknowledges that details and objects can still drift in longer sessions, that the accuracy of generated audio is unresolved, and that readable text inside generated scenes is not yet reliable. The company calls the model early and says the shape of any finished product is unclear.

One experiment is aimed at robotics. A Gemini model planned a task in steps while Genie 4 simulated each step in real time, a way to test how a robot might behave before trying it in the physical world. Tejas Kulkarni, a Google DeepMind research director, described such pairings as hybrid world models, according to TechCrunch. Testers are also trying the system for film pre-visualization and storyboarding.

Access will widen slowly. Google says Ultra subscribers using Project Genie and early partners in the Gemini API and Google AI Studio will get access in the coming weeks, though no firm date was given. Google says it will apply its SynthID watermark to all video and audio the model produces.

The announcement lands in a crowded field. TechCrunch notes that Runway released a playable world model of its own, GWM Worlds 2, a couple of months ago and that the sector is consolidating, with AMD reportedly acquiring World Labs and Amazon reportedly in talks to buy Decart.

Supporters of world models argue they could make robot testing safer and cheaper, since a mistake in a simulation costs nothing and can be repeated endlessly. Skeptics point to the limits the company itself lists: two minutes of coherence is far short of the long, reliable simulations that safety testing would demand, and a simulated world that drifts can teach a robot the wrong lessons.

For now, Genie 4 is a demonstration of progress rather than a product, and its value will depend on how much of the realism holds up when outsiders test it. The watermarking plan and the limited rollout suggest Google expects scrutiny about how convincing generated audio and video can be.