Google DeepMind’s Genie 4 doubles scene consistency to two minutes

Control is the new feature

Google introduced Genie 3 in research mode a little over a year ago. Genie 4 improves the visual realism and physics, makes characters look more human, and lets users change a scene as it runs: they can alter the weather or camera angle, or add people and objects.

That live control may matter more than the visual upgrade. Google DeepMind engineer Diego Rivas said it is especially useful in creative work that depends on continuity, such as developing ideas and storyboarding, where people need to try variations quickly.

Genie 4 also works more closely with Gemini. Gemini plans actions and checks frames while Genie generates the scene. In an experiment connected to Gemini Robotics, Gemini 4 Argon broke tasks into steps that Genie 4 simulated in real time. The demonstration points to a way to test robot scenarios in simulation before they happen in the physical world.

A crowded field, and an unfinished product

Genie 4 arrives months after video AI startup Runway released its second world model, GWM Worlds 2, which supports gameplay, responds to user actions and generates audio. The wider market is also drawing acquisition interest: World Labs, founded by Fei-Fei Li, is buying AMD; Amazon is in talks to buy Decart; and Rivian acquired Third Dimension to use its simulation technology for autonomous-vehicle testing.

Google is positioning Genie 4 for games, interactive learning and robotics. One example of the learning use case is stepping into a historical scene depicted in a painting and talking with the people there about the event. Trusted testers have also identified film prototyping as a possible application.

Google already has Veo, its model for generating video from text, image and video prompts, which produces high-resolution clips with synchronized audio and is gaining popularity among filmmakers. Rivas said users may eventually be able to turn prototypes into creative material with Gemini Omni, a multimodal large language model that generates video from images, audio and text. He cautioned that those capabilities are still a long way off.

I think the announcement’s most important omission is a clear account of what Genie 4 is meant to become. Rivas said Google has not decided what the finished product will look like. The team also acknowledged that details and objects can change as a simulation continues, and that it has not yet found a reliable way to keep audio accurate or text readable.

Genie 4 is available now to trusted testers in a research version. In the coming weeks, it will begin rolling out to Google AI Ultra subscribers through Project Genie, Google’s tool for creating interactive worlds from text or image prompts. Gemini API and Google AI Studio partners will also get early access.

The launch gives Google a stronger demonstration of controllable world simulation, but the gap between a convincing prototype and a dependable creative tool remains the real test.