Genie 4 Lets Users Steer Generated Worlds as They Unfold

Google DeepMind presents Genie 4 as a world model in which a user can enter a generated environment and keep changing it while moving through it. The demonstrations combine text prompts with standard keyboard controls for navigation and camera movement. The user can steer what happens next without specifying the entire scene in advance.

That distinction is clearest in the snowy landscape. A crowned crane begins walking beneath a dusk sky. The user adds a shooting star, then prompts the crane to approach a white piano and press a key. The sequence demonstrates successive instructions changing an already unfolding scene: prompts are not limited to describing the world at its start.

Other scenes show the same interaction applied to events, characters and atmosphere. In a dusk moped scene, the user prompts a character to look back and another to ask whether they will make it. Further prompts add explosions and gigantic alien tentacles to the background. In a separate dragon-riding scene, the user enters a first-person view of riders heading toward a castle; spoken lines include a call to attack and a warning that it is a trap. These examples present prompting as a way to direct both the environment and developments within it.

DeepMind also shows a picture being turned into a place to move through. The input is an image of Van Gogh’s The Starry Night, with a prompt asking for an explorable world that maintains visual consistency. The resulting simulation includes a stone path and walking figures. DeepMind says the transformation includes ambient sound. Together, the examples point to a system intended to support interaction across more than one mode: users can move, change the camera, direct details in text, and hear the resulting world. The footage is labeled “Results for illustrative purposes. Sequences shortened.”

After the video moves through other environments, it returns to the snowy landscape, where the crane is standing beside the piano keyboard. DeepMind frames this as the original world continuing to evolve while the user spends time in other worlds. The later view shows the crane beside the object it had been prompted to approach.

The montage broadens the kinds of scenes in which users can move, including a train, a skateboard, a floating astronaut, a field, a snowplow and a triceratops. Keyboard controls appear throughout. Rather than specifying a single kind of environment, the demonstrations frame Genie 4 as supporting varied interactive scenarios, with text prompts available to reshape them as they unfold.

DeepMind proposes using Genie 4 to simulate scenarios in which robots can practice tasks virtually, with the possibility that they might one day carry out work in the physical world. The video shows a grid of simulated feeds featuring robotic arms performing laboratory tasks, then focuses on one arm working with a pipette and a microplate.

An on-screen reasoning panel says the arm successfully inserted the pipette tip into a microplate well and is advancing to liquid aspiration or retraction. In the example, the task is represented as a sequence: one event is completed, and the simulation moves on to the next. That gives the robot demonstration a different emphasis from the character scenes. It presents simulation as a setting for practicing steps in a task, rather than only as a world for a person to explore.

The physical-world use remains a proposed application: the video shows the task in simulation, while DeepMind describes virtual practice as a way robots might eventually prepare for work outside it.

DeepMind says Genie 4 is rolling out as a research preview for trusted testers. It is expected to come to Google AI Ultra subscribers through Project Genie, and to early-access partners using the Gemini API and Google AI Studio. The announcement describes a staged preview, rather than general availability.