Genie 4: A new generation of world simulation
Google DeepMind unveils Genie 4, a multimodal world model that generates photorealistic, interactive simulations with real-time audio, physics, and natural language control for research and development applications.
Genie 4 introduces a multimodal world model capable of generating interactive simulations with synchronized audio, ambient sounds, and dialogue alongside visuals. The model integrates capabilities from the Gemini family to anchor simulations in real-world behavior while maintaining coherence and continuity during user-driven events. Users can initiate actions through natural language prompts, such as altering weather or introducing objects, and the system adapts dynamically to maintain realistic progression. The model is currently available as a research preview for trusted testers, with plans to expand access to Google AI Ultra subscribers and early partners in Project Genie and the Gemini API.
Genie 4 enhances photorealism through advanced physical simulations, including aerodynamics, collisions, and natural character movements that respond organically to environments. The model supports real-time modifications via text prompts, enabling users to steer scenes while maintaining spatial and temporal continuity. Demonstrations include cinematic documentary-style scenes with controlled lighting, camera movements, and ambient sounds produced entirely by the system. The platform emphasizes naturalistic rendering, avoiding artificial enhancements like HDR or dramatic contrast.
The system allows users to set initial conditions using reference images or detailed prompts, such as recreating specific environments with precise visual elements. Examples include a floating platform with a saxophone player in a misty lakeside setting or a sun-drenched meadow with a flowing pink satin stream. Genie 4 simulates these scenes with realistic physics, including fluid movement of fabrics and natural wind interactions. Users can navigate freely within the environment using standard controls while the model maintains consistent environmental and character behaviors.
Genie 4 is positioned as a step toward safe, scalable simulations for applications like robotics training and interactive experiences. The model’s ability to generate cohesive, real-time worlds with native audio and dynamic responses underscores its potential for research and development. Access is currently limited to a research preview, with broader release planned for select partners and subscribers in the coming phases.