Genie is a series of generative world models developed by Google DeepMind. The first model, unveiled in February 2024, was trained on internet gameplay videos and produced controllable 2D environments from a single image; Genie 2, released in December 2024, extended this to 3D environments. Genie 3, announced in August 2025, lets users navigate environments created from text prompts in real time and alter them with further text. A successor, Genie 4, was announced in 2026.
Project Genie is a research prototype that lets the public create environments with Genie models and explore them from a first- or third-person viewpoint. Unlike video generation models, Genie is defined by interactivity, rendering the next frames in response to user actions, and it is also used to generate environments for training and evaluating AI agents.