Google's Genie 4 World Model Adds Audio, More Realistic Worlds

Google DeepMind has unveiled its latest world model, Genie 4, in a research preview, marking a significant leap in generating photorealistic and interactive environments. Unlike its predecessor, Genie 3, Genie 4 offers even more advanced features. Users can now type prompts at any point during the simulation to modify the scene, adding to the flexibility and interactivity of the model.

Additionally, Genie 4 generates real-time audio, incorporating music, ambient noises, and interactive dialogue that aligns with the simulated world. This is a substantial improvement over Genie 3, which lacked this audio capability.

One of the most notable advancements in Genie 4 is its ability to allow freeform edits. Users can now alter the weather, camera angle, add new people or objects, and maintain a consistent scene for up to two minutes - a substantial increase from the previous generation's limit. The model also boasts improved physics and more human-like characters, making the simulated environments more believable.

Tejas Kulkarni, Google's research director, commented on the development, stating that we'll increasingly see hybrid world models in the future. These will involve a combination of Google's Gemini AI writing code and text, and Genie ingesting this information in real time to simulate reality at different levels of abstraction. Genie 4 is currently available for select testers, with a wider rollout planned for Google AI Ultra subscribers in the coming weeks. A demo of Genie 4 is available on YouTube.