Google DeepMind's Genie 2 Generates Playable 720p 3D Worlds from Single Image or Text Prompt

Dec 4, 2024, 03:52 PM

Google DeepMind has unveiled Genie 2, its most advanced large-scale foundation world model to date. Capable of generating an endless variety of action-controllable, playable 3D environments from a single image or text prompt, Genie 2 produces immersive worlds in 720p resolution that can be navigated using keyboard and mouse inputs for up to a minute. The model simulates virtual worlds, including the consequences of actions such as jumping or swimming, by leveraging a large-scale video dataset. Genie 2 supports both human and AI agent interactions and can generate diverse outcomes from the same starting point. Developed by the Open-Endedness Team at Google DeepMind, this breakthrough builds upon the progress of Genie 1, launched nine months prior. Genie 2 could unlock new capabilities for training and evaluating embodied AI agents and holds the potential to revolutionize video game development by enabling prompt-to-game generation without the need for traditional game engines.