Google DeepMind Previews Genie 4 World Model - Dataconomy
Google DeepMind released Genie 4 as a research preview on Tuesday. It is a world model that creates interactive environments with real-time audio. It also lets users change scenes while a simulation is running.
Trusted testers can use it now. In the coming weeks, Genie 4 will become available to Google AI Ultra subscribers through Project Genie, which creates interactive worlds from text or image prompts. It will also reach early-access partners using the Gemini API and Google AI Studio.
Genie 4 creates music, background sounds and interactive conversations. During a simulation, users can type instructions to change the weather or camera angles, or add people and objects.
The model can keep a scene consistent for up to two minutes, twice as long as Genie 3, which was released as a research preview a little over a year ago. Genie 4 also creates more realistic-looking images, handles physics better and produces characters that behave more like humans.
Google's Gemini AI model is more closely connected to Genie 4. Gemini can plan steps and check individual frames while Genie creates the visuals. In a research experiment called Gemini Robotics, a Gemini 4 Argon model broke tasks into smaller steps, which Genie 4 then simulated in real time. This showed one possible way to test robot situations in a simulation before trying them in the real world.
Google research director Tejas Kulkarni said during a briefing that future world models could combine different strengths. These hybrid models could use Gemini to generate code and text while Genie simulates the real world in real time, at different levels of detail.
Possible business uses include video games, robotics and interactive learning. One imagined learning example would let someone step inside a painting of a historical event and talk to the people shown in it about what is happening in the scene.
Google's trusted testers have also identified making early film prototypes as a possible use. DeepMind engineer Diego Rivas said during a press briefing that real-time features were especially useful for keeping creative ideas flowing while planning scenes and storyboards.
Google also offers Veo, a tool that creates high-definition video clips with matching audio from text, image or video prompts. It has been becoming more popular among filmmakers.
Rivas said that, in the future, users might be able to turn Genie 4 prototypes into creative materials using Gemini Omni. This is a multimodal large language model that generates video from images, audio and text. However, Rivas said these features are still some way off.
The release comes a couple of months after Runway launched GWM Worlds 2, its second world model. It can be played like a game, responds to user actions and creates audio.
AMD is acquiring Fei-Fei Li's World Labs, while Amazon is in talks to buy Decart. Rivian acquired Third Dimension so it could bring simulation technology in-house for testing self-driving systems.
Rivas stressed that Genie 4 is still at an early stage, and it is not yet clear what the complete final product will look like. The team acknowledged several problems: details and objects can gradually change during longer simulations, audio accuracy is still an unresolved issue, and text inside generated worlds may be difficult to read.