Google DeepMind has unveiled Project Genie, an experimental AI tool that generates virtual worlds users can explore and manipulate based on text and images, opening a new chapter for researching and testing world-model technology.

TechCrunch and a Google blog said on Jan. 29 that Google DeepMind opened access to Project Genie for Google AI Ultra subscribers aged 18 and over in the United States. The release follows a research preview of Genie 3, a general-purpose world model introduced in August last year. The goal is to let users experience world models more interactively while providing feedback and learning data. Through this, DeepMind plans to validate the usefulness and performance of world models in real user environments.

Project Genie is a web-based experimental application that combines Genie 3 with the image-generation model Nano Banana Pro and the language model Gemini. Users can create a world sketch using a text prompt or a generated or uploaded image, set characters, movement methods and viewpoint (first-person or third-person), and enter a virtual world. Users can preview and revise the visual composition through Nano Banana Pro. Based on that, Genie 3 generates routes and environments in real time, and the world expands according to user actions.

Google DeepMind highlighted world sketch, world exploration and world remix as core functions of Project Genie. Users can freely explore environments generated based on movement and interaction. They can add new interpretations based on prompts from existing worlds, or use curated worlds and random-generation tools to experience varied environments. The created worlds and exploration process can also be downloaded as video files, allowing users to record and share personal creations.

DeepMind aims through this project to develop general-purpose AI that can handle diverse situations and changes in the real world, beyond AI agents specialized for a single environment.

Genie 3 was designed with possible uses in mind such as robot training, animation and fiction production, and exploration of real locations or historical spaces. It generates routes in real time based on user actions, simulates physics and interactions, and maintains a high level of consistency.

Project Genie is a research prototype, with generation and exploration time limited to up to 60 seconds. Some generated worlds may not fully reflect real-world physics and prompts, or may experience character control lag and unstable controls. Google DeepMind said it will continue to improve these limitations and, over the long term, expand access to more users and regions.