Genie 2

A world model you can play

Genie 2 is a foundation world model from Google DeepMind, announced on 4 December 2024 ^[1]. Given a single prompt image, it generates a playable 3D environment ^[1]. A world model, in the general sense, is a system that learns how a scene tends to behave, so that the scene can be continued forward as things happen inside it rather than merely rendered once. That is the difference between a generated video, which you can only watch, and a generated world, which you can act in. The word "foundation" points at the same idea from another direction: one general model meant to sit underneath many downstream uses, rather than a system built for a single task.

Controls and consistency

A person or an AI agent plays the generated world with keyboard and mouse inputs, and Genie 2 keeps that world consistent for up to a minute ^[1]. Consistency is the hard part: every action changes what should be visible next, and small errors compound from one step to the next. Genie 2 also keeps long-horizon memory of parts of the world that are no longer in view ^[1], so a scene does not reset itself the moment the camera turns away.

What it simulates

The model reproduces interactions between objects — opening doors ^[1], bursting balloons, shooting barrels of explosives ^[1] — as well as physics effects such as water, smoke and gravity ^[1]. These are the ingredients of a small game-like space rather than a still picture: things react, fall, spread and break.

For every example, Genie 2 was prompted with a single image generated by Imagen 3 ^[1]. The starting point need not be a photograph or a hand-built level; a generated picture is enough to seed a world.

Agents inside the world

DeepMind put a SIMA agent into environments generated by Genie 2, and the agent followed natural-language instructions there ^[1]. This is the practical promise of a world model: an agent can be trained or evaluated in scenes that are cheap to produce and endlessly varied, instead of only in worlds built by hand or in the real world, which is slow and hard to repeat.

Where it sits

A window of about a minute defines the scale at which these worlds were shown: acting coherently over long stretches, where the world has to keep track of many earlier choices, is a different problem from acting for a few dozen steps.

Genie 2 was succeeded by Genie 3 in August 2025 ^[2], the first in the series to allow real-time interaction ^[2]. According to DeepMind, Genie 3 generates worlds that can be navigated in real time at 24 frames per second at 720p resolution ^[2], and its environments remain largely consistent for several minutes ^[2] — a longer window than the minute-scale consistency of Genie 2 ^[1]. Placed side by side, the two mark the step from a generated world you can interact with to one that answers at the speed of play.