Google DeepMind previews Genie 4 with real-time world editing and explorable environments

Google DeepMind has introduced Genie 4 in a research preview. The AI world model generates explorable virtual environments that respond to user interactions, with higher visual fidelity and improved consistency than Genie 3. The model also adds real-time audio, including background music, ambient sounds, and interactive dialogue, while allowing users to modify running simulations through natural-language prompts to change weather, camera angles, characters, or objects.

Genie 4 can maintain consistent environments for up to two minutes, doubling its predecessor's limit, and offers improved physics simulations and more humanlike characters. However, Google acknowledges that objects and scene details can still drift during longer simulations.

The model also features deeper Gemini integration, with Gemini planning actions and checking frames while Genie 4 handles the simulation. Genie 4 is expected to become available to Google AI Ultra subscribers soon, with a waitlist already open through Project Genie.