Google DeepMind has launched Genie 4, an AI model capable of creating interactive virtual worlds in real-time, enhancing realism and user interaction.
Google DeepMind has introduced Genie 4, a new artificial intelligence model that can create interactive virtual worlds in real-time, complete with sound. Developers report enhancements in realism, physics modeling, and character behavior.
Genie 4 is categorized as a world model, which differs from traditional video generators by simulating environmental changes based on user actions. Users can navigate the virtual space using a keyboard and manage the camera, while text commands can alter scenes by adding objects or characters and changing weather conditions. The Gemini system adapts these requests to the ongoing scenario and ensures a coherent world progression.
According to TechCrunch, Genie 4 can maintain a consistent scene for up to two minutes, doubling the capability of its predecessor.
New Features
A key update in Genie 4 is its integrated sound generation. It produces ambient noise, music, and dialogue simultaneously with visuals. Characters engage in conversations, respond to user actions and events, and adjust their lines as the scene unfolds.
Developers have also improved aerodynamic modeling, collision detection, and other physical interactions. The most significant advancement includes more natural character movements and real-time interaction with the environment.
Demonstrations showcased scenarios such as a hot air balloon flight over Antarctica and a rural Japanese station from the 1980s, alongside dragon racing and a stained-glass world.
Genie 4 builds on the Gemini family of models, with Gemini Omni ensuring realistic object appearance and behavior, while other Gemini models help maintain simulation coherence. For complex multi-step events, Google employs Gemini 3.8 Flash as a managing agent to oversee the scene, break tasks into steps, analyze new frames, and verify the execution of each phase. If the world changes or a user intervenes, the agent adjusts the plan accordingly.
Google is also exploring the combination of Genie 4 with Gemini 4 Argon for research in Gemini Robotics. In one demonstration, Argon broke down a task while Genie 4 simulated how the humanoid robot Apollo from Apptronik plays checkers in real-time.
Researchers aim for such environments to train robots and AI agents, testing their behavior in simulations before real-world deployment.
Applications and Limitations
Genie 4 is currently being tested by professionals in film, journalism, education, game development, and engineering. They are utilizing the model for film pre-visualization, game prototyping, educational and historical scenarios, and generating situations for simulations.
Diego Rivas, an engineer at DeepMind, noted that real-time alterations are particularly beneficial in the early stages of creative work, such as brainstorming and storyboarding.
However, Genie 4 remains an experimental model. Google’s team acknowledged that over time, details and objects may change in appearance or position, especially during lengthy sessions or after complex edits.
Further work is needed on complex physical interactions, sound fidelity, text readability, and spatial accuracy. The company did not provide quantitative quality metrics in its announcement.
To label AI-generated content, all videos and audio created with Genie 4 will feature digital watermarks from SynthID.
Currently, Genie 4 is available to a limited group of testers in a research preview. Following this phase, it will be accessible to subscribers of Google AI Ultra in Project Genie and partners in the early access program for Gemini API and Google AI Studio. According to TechCrunch, this rollout will occur in the coming weeks, with a waiting list open for Project Genie. Pricing details for API usage remain undisclosed.
