Genie is an 11-billion-parameter generative interactive environment from Google DeepMind, described by its creators as the first foundation world model: it turns a single image, photo, or sketch into a playable, controllable two-dimensional environment, and it learned to do so entirely from unlabelled Internet gameplay videos. Genie was introduced in the paper "Genie: Generative Interactive Environments," posted to arXiv on 23 February 2024, and it won a Best Paper Award at ICML 2024.[1][2][5] The release described here is the original model, sometimes referred to informally as "Genie 1" to distinguish it from the later Genie 2 and Genie 3; this article covers that first version.

The work was carried out by DeepMind's open-endedness group together with collaborators at the University of British Columbia.[3] At 11 billion parameters, the authors argue Genie can be considered a foundation world model, by analogy with the foundation models that had reshaped language and image generation.[1] It was presented as a research result rather than a public product, and was never offered as a consumer tool or playable demo.

What is Genie?

Genie is a learned system that generates action-controllable virtual worlds from a prompt, then lets a user "play" them one frame at a time. In the words of the paper's abstract, "We introduce Genie, the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual worlds described through text, synthetic images, photographs, and even sketches."[1] Unlike a video generator, which produces a fixed clip, Genie responds to a control signal at every step, so the same prompt image can unfold differently depending on the actions a player chooses.[1][2]

A world model is a learned system that predicts how an environment will change in response to actions, letting an agent "imagine" future states instead of acting only in the real world. The idea has a long history in reinforcement learning, including DeepMind's own Dreamer line of model-based agents. What set Genie apart was the source and form of its supervision. Earlier interactive world models typically required environments with known action sets, or video paired with logged controller inputs. Genie instead learned from raw, unlabelled gameplay footage scraped from the public Internet, inferring both the visual dynamics of a game and a notion of "actions" without ever being told which button a player pressed.[1]

That choice mattered because action-labelled video is scarce, while unlabelled video is abundant. By learning a controllable interface from passive footage, Genie pointed toward a way of building interactive simulators, and ultimately training environments for embodied agents, at the scale of Internet data rather than hand-built games.[1]

How does Genie work?

Genie is built from three components that are trained largely on the same backbone, a spatiotemporal transformer (ST-transformer) that attends across both space within a frame and time across frames.[1]

Video tokenizer. A spatiotemporal VQ-VAE compresses raw video frames into a grid of discrete tokens, with the ST-transformer in both encoder and decoder so the codes capture motion rather than treating each frame in isolation. In the paper this tokenizer has roughly 200 million parameters and a codebook of 1,024 entries.[1][4]

Latent action model. The core trick is an unsupervised latent action model (LAM) that looks at consecutive frames and learns a small set of discrete "actions" that best explain the transition from one frame to the next. The vocabulary is deliberately tiny: the authors restrict it to eight latent actions, which forces the model to discover a compact, meaningful control space. The LAM is used only during training to provide action targets; at inference the user supplies one of these latent actions each step.[1][4]

Dynamics model. A MaskGIT-style transformer takes the past frame tokens together with a chosen latent action and predicts the tokens of the next frame, trained autoregressively. This is by far the largest part of the system, on the order of 10 billion parameters, which accounts for most of Genie's 11 billion total.[1][4]

At generation time the pipeline runs in a loop. A prompt image is tokenized, the user picks a latent action, the dynamics model predicts the next frame's tokens, and the tokenizer's decoder renders them back into pixels. Repeating this turns a still image into a frame-by-frame playable sequence in which the same latent action produces consistent effects, such as moving left, jumping, or scrolling the scene.[1]

Component	Approx. parameters	Role
Video tokenizer (ST-transformer VQ-VAE)	~200M	Compress frames to discrete tokens
Latent action model	~300M	Infer 8 discrete latent actions (training only)
Dynamics model (MaskGIT transformer)	~10B	Predict next-frame tokens from action and past
Total	~11B	Foundation world model

How does Genie learn from videos?

Genie's main model was trained on a filtered set of about 30,000 hours of publicly available Internet gameplay videos drawn from hundreds of different 2D platformer games.[1] DeepMind began from a much larger pool of clips and trimmed it down using a learned classifier, ending with roughly 6.8 million sixteen-second clips. The footage was standardized to a low 160 by 90 resolution at 10 frames per second to keep training tractable at this scale.[4] The 11B model was trained for 125,000 steps on 256 TPUv5p accelerators.[1][4]

Crucially, none of these videos carried action labels. The platformer genre was a useful testbed because side-scrolling games share a fairly consistent grammar of movement, which made the latent actions easier to discover. To show the approach was not limited to games, the authors also trained a separate 2.5-billion-parameter model on a robotics dataset (RT-1 style robot demonstrations), where it again recovered distinct, consistent action representations from video alone.[1]

What can Genie do, and what are its limits?

Given a single starting frame, Genie can generate an open-ended variety of playable 2D scenes, and it accepts prompts well outside its training distribution, including text-to-image outputs, real photographs, and rough sketches.[1][2] Because the learned latent actions are consistent across different generated worlds, a person (or another model) can "play" the environment in a controllable way. The paper also showed that latent actions learned purely from Internet video could be used to infer policies in unseen reinforcement-learning environments, suggesting a path toward training generalist agents from passive video.[1]

The original Genie was firmly a research prototype. It generated at a very low resolution and ran extremely slowly, on the order of one frame per second, which the team noted was roughly 20 to 30 times slower than interactive play would require.[3] It could only hold about 16 frames of context, so worlds drifted and lost coherence over longer horizons, and the imagery was blocky and prone to artifacts. It was confined to 2D side-scrolling dynamics rather than full 3D scenes. These constraints, especially speed, consistency, and resolution, became the explicit targets of the successor models.

How was Genie received?

Genie drew considerable attention as an early demonstration that controllable, interactive environments could be learned from unlabelled video at foundation-model scale. At the Forty-first International Conference on Machine Learning (ICML 2024), the paper received a Best Paper Award, one of two such awards given that year to DeepMind's open-endedness team.[5][6] Lead authors including Jake Bruce and senior author Tim Rocktaschel publicly noted the recognition.[6] Press coverage emphasized both the novelty of turning a single image into a playable world and the practical limitations of the first version.[2][3]

What came after Genie 1?

DeepMind extended the line in two further releases. Genie 2, announced on 4 December 2024, moved beyond 2D platformers to generate action-controllable 3D environments from a single image, simulating physics, lighting, and object interactions, though it held coherence only for tens of seconds to about a minute.[7] Genie 3, unveiled on 5 August 2025, was presented as the first real-time interactive general-purpose world model, generating 720p worlds at 24 frames per second from a text prompt and maintaining consistency for several minutes with a visual memory reaching back roughly one minute.[8] The broader effort is closely tied to DeepMind's work on embodied and instruction-following agents such as SIMA, which the world-model environments are intended to help train and evaluate.

Version	Announced	Output	Key advance
Genie 1	23 Feb 2024	2D platformer worlds, ~1 frame/sec	First 11B foundation world model from unlabelled video
Genie 2	4 Dec 2024	3D worlds from a single image	Physics, lighting, object interactions; tens of seconds of memory
Genie 3	5 Aug 2025	720p worlds at 24 fps from text	First real-time general-purpose world model; ~1 minute memory

References

Editing: Genie (DeepMind)