IT Home On October 28, a blog post by Google DeepMind team (dated October 27) announced the launch of the Genie 4 world model, described as a “major research milestone.” It achieves key breakthroughs in native audio, real-time natural language rewriting, and more realistic physics and characters, transforming the world model from a one-time generated image into a dynamic simulation world that can be controlled in real time, heard, and reasoned about.
For users, the biggest feature of Google Genie 4 World Model is the real-time “director” of the world. During simulation runs, users can use natural language text commands to change the world, such as changing the weather, moving the camera, adding new characters or objects.
Unlike mainstream AI that re-generates a video, Genie 4 combines the current scene context to achieve dynamic changes, more like a “interactive virtual sandbox” or “AI director system”. Users can set up an event, then observe how the world evolves, while being able to intervene and adjust at any time.
In other aspects, Genie 4 is DeepMind’s first model that generates audio in real-time world simulations, with the audio and visuals being produced simultaneously. Additionally, Google claims that Genie 4 performs better in physical scenarios such as aerodynamics and collisions; the characters’ movements are also more natural, capable of organic responses to the environment, and it supports real-time control. IT Home has attached the related demonstration video below:
Google CEO Sundar Pichai posted a tweet on his personal account today, saying Genie 4 is Google’s most realistic and interactive world model so far.
In an interview with TechCrunch, Google Research Director Tejas Kulkarni said, in the future, Gemini will be responsible for text, code, and task planning, while Genie 4 will convert these elements into an interactive world in real time, creating a “hybrid world model”.
DeepMind engineer Diego Rivas admitted that Genie 4 is still in its early stages; long-term simulations may result in detail drift, and the audio accuracy and text readability within the video have not been fully resolved yet.
Genie 4 is currently available only in a research preview form for “trusted testers”; it will be expanded in the coming weeks to Project Genie for Google AI Ultra subscribers, as well as early partners of the Gemini API and Google AI Studio.
Related Reads: