Genie 3

Open original

Genie 3 is a DeepMind research prototype that turns a text description into a photorealistic 3D environment you can explore in real time. The article frames it as a step toward training AI agents in varied, safe simulated worlds, though it still has clear limits on accuracy and duration.

  • Genie 3 generates interactive environments at 720p and 20-24 frames per second, frame by frame from a text prompt and user actions.
  • It keeps worlds consistent for several minutes and recalls changes from specific interactions for up to a minute, but can't sustain hours of interaction.
  • Genie is grounded in Google Street View data, and "promptable world events" let users change weather or add objects mid-simulation to test how agents handle surprises.

Genie 3 is a DeepMind research prototype that generates explorable 3D worlds from a text prompt. You describe a place, anything from a desert coastline to a fantastical animated scene or Ancient Rome, and the model produces photorealistic environments at 720p that you can move through in real time, at roughly 20 to 24 frames per second. The key technical point is that it's auto-regressive: each frame is generated based on the world description plus your actions, rather than being rendered from a fixed 3D model like NeRFs or Gaussian Splatting. That makes the worlds more dynamic and detailed, but it also makes consistency hard. DeepMind says environments stay largely coherent for several minutes, and the model can recall changes from specific interactions for up to about a minute, so if you revisit a spot, the model has to remember what was there. Genie is also grounded in Google Street View data, which lets people generate scenes anchored to real locations. The article frames Genie 3 as a step toward AGI, and the more concrete pitch is for training agents. Genie 3 can introduce "promptable world events," such as changing the weather or adding objects and characters, to give agents practice with unexpected situations. In one demo, DeepMind pairs the model with SIMA, its agent that carries out tasks in virtual environments: SIMA gets a goal, Genie 3 simulates the consequences of its actions without knowing the goal, and researchers can probe where the agent fails. The company also points to education (letting students walk through historical eras) and to safe training for autonomous vehicles as applications beyond gaming. The limitations are stated plainly, which is useful. Genie 3 can't simulate real-world locations with perfect accuracy. Text in the generated world is only legible when it was already in the prompt. Sessions support a few minutes of continuous interaction, not hours. Modeling multiple independent agents interacting in a shared space is still an open research problem, and the range of actions agents can take is limited. The article also says the team worked with its Responsible Development & Innovation Team on safety, since real-time, open-ended generation creates new risks, but it doesn't spell out specific mitigations. The page closes with a long acknowledgements list naming the research and engineering contributors.

Ask about this article