Genie 3 is a general-purpose world model developed by Google DeepMind and announced on August 5, 2025. It generates interactive virtual environments from text descriptions, producing new visual frames in response to navigation and other inputs. At its announcement, DeepMind reported real-time operation at 24 frames per second and 720p resolution, with environments remaining broadly consistent for several minutes. It was initially introduced as a limited research preview. (deepmind.google)
Development
Genie 3 belongs to a series of systems for generating environments rather than merely producing fixed images or video clips. The original Genie, described in a February 2024 research paper, learned interactive environments from video without action annotations. Its research architecture combined a video tokenizer, a latent-action model and a dynamics model. (arxiv.org)
Genie 2, announced on December 4, 2024, extended the series to diverse, action-controllable environments with three-dimensional appearances. Starting from an image, it generated subsequent observations in response to keyboard and mouse inputs. DeepMind described it as an autoregressive latent diffusion model, with consistent generations lasting up to approximately one minute. Those architectural details describe Genie 2 and should not automatically be attributed to Genie 3. (deepmind.google)
Operating principle and capabilities
Genie 3 generates its environments autoregressively: new frames depend on the world description, earlier generated observations and the user’s actions. Rather than following an already completed video sequence, the visual trajectory develops during interaction. Maintaining continuity requires the system to recover relevant information when previously visited locations return to view. DeepMind reports memory of particular interactions extending back approximately one minute. (deepmind.google)
The model also supports promptable world events—text instructions that alter an environment during exploration, such as changing the weather or introducing objects and characters. These interventions are distinct from actions performed directly by a controlled character. (deepmind.google)
This makes Genie 3 a form of interactive generative artificial intelligence, not simply a system for rendering a static scene. Its apparent spatial consistency emerges through frame generation rather than requiring a supplied explicit three-dimensional scene representation. Nevertheless, errors can accumulate as generation continues. (deepmind.google)
Project Genie
Project Genie is the experimental user-facing application through which Google made Genie 3 more widely accessible; it is not another name for the underlying model. Launched on January 29, 2026, for adult Google AI Ultra subscribers in the United States, the web application combines Genie 3 with Nano Banana Pro and Gemini. (blog.google)
Its initial workflow included three main functions:
- World sketching: creating an initial environment and character using text and generated or uploaded images.
- World exploration: navigating while the system generates the unfolding environment.
- World remixing: modifying existing prompts to create alternative interpretations of a world.
Users could also download recordings of their explorations. At launch, sessions were limited to 60 seconds, and promptable world events were not included. These application restrictions were separate from the longer interaction capabilities demonstrated for the research model. (blog.google)
On May 19, 2026, Google announced a global rollout to eligible adult Google AI Ultra subscribers and introduced Street View grounding. This feature allowed starting locations to be based on imagery of places in the United States, with optional stylistic transformations. Grounding an initial scene in real imagery did not constitute a guarantee of geographically exact simulation. (blog.google)
Research applications
One intended application is generating environments for AI agents. DeepMind demonstrated a version of its SIMA agent pursuing goals inside Genie 3 environments: the agent supplied navigation actions, while the world model generated the resulting observations without being given the agent’s goal. This demonstrated compatibility with agent interaction, rather than establishing that arbitrary skills learned in generated worlds transfer reliably to reality. (deepmind.google)
DeepMind has also identified potential uses in education, creative exploration and the training or evaluation of autonomous systems. Such uses depend on the accuracy and controllability required by the task; visually convincing output alone does not establish a validated simulation. (deepmind.google)
Limitations
DeepMind identifies several limitations: a restricted range of directly executable actions, difficulty simulating interactions among independent agents, imperfect geographical accuracy, inconsistent text rendering and interaction durations measured in minutes rather than hours. (deepmind.google)
Project Genie’s launch documentation additionally noted that generated worlds could depart from prompts, source images or real-world physics, and that character control could be unreliable or exhibit latency. Consequently, its generated environments should be distinguished from simulations whose physical behavior and spatial structure have been independently verified. (blog.google)
References
- Genie 3: A new frontier for world modelsdeepmind.google
- Genie: Generative Interactive Environmentsarxiv.org
- Genie 2: A large-scale foundation world modeldeepmind.google
- Genie 3 — Google DeepMinddeepmind.google
- Project Genie: Experimenting with infinite, interactive worldsblog.google
- Simulate real-world places with Project Genie and Street Viewblog.google