Google DeepMind has announced Genie 3, a new world model that generates real-time, interactive, and high-consistency virtual environments from text prompts. Genie 3 enables users to navigate dynamic worlds at 24 fps and 720p resolution, preserving coherence for several minutes. Compared to Genie 1 and Genie 2, it delivers a major leap in realism and interactivity, bringing instant control to world simulation.
Deep physics and environmental interactions
Genie 3 can model natural phenomena such as water, lighting, wind, and complex surface interactions. It maintains physical consistency across challenging scenarios like driving a rover over volcanic terrain, walking along a coastal road during a hurricane, or flying a drone through narrow canyons.
Natural and fictional worlds
The model spans a wide spectrum—from vibrant ecosystems and dense foliage to animated, fantastical characters and scenes. It can generate and let users explore settings such as Zen gardens, Alpine gorges, the canals of Venice, or the Palace of Knossos, in real time.
Long-horizon consistency and real-time performance
Auto-regressive frame-by-frame generation is prone to compounding errors. Genie 3 mitigates this by maintaining a visual memory that can refer back roughly a minute, keeping scenes logical and physically plausible for several minutes. While techniques like NeRFs and Gaussian Splatting enable consistent 3D navigation via explicit 3D representations, Genie 3 builds more dynamic worlds frame by frame based on prompts and user actions.
“Promptable world events”
Beyond navigation controls, the model supports text-triggered events that can alter the generated world—such as changing weather or introducing new objects and characters. This expands counterfactual “what if” scenarios, helping experience-driven agents handle unexpected situations.
Fueling agent research
DeepMind tested compatibility with generalist 3D agents like SIMA. Thanks to improved consistency and real-time control, agents can now execute longer action sequences to accomplish more complex goals. The team sees this as pivotal for both generative media and AI research on the path toward AGI.
Responsible rollout and access
Given the open-ended, real-time nature of the technology, Genie 3 is being released as a limited research preview to select academics and creators. DeepMind plans a gradual expansion informed by feedback and responsible development practices. Potential applications include education and training, robotics and autonomous systems, performance evaluation, and weakness analysis.