Google has introduced a new AI that generates 3D reality from descriptions.

Google DeepMind has introduced a new version of its AI virtual world model called Genie 3. It can generate 3D game spaces in real time that can be interacted with by both users and AI agents.

Genie 3 is an improved version of the previously released Genie 2. It allows for longer interaction with the created world and remembering the location of objects. This was made possible by the model's expanded "visual memory," which now works for up to a minute. The resolution of the worlds is 720p at 24 frames per second.

AI virtual world models allow for the creation of fully generated locations — like video games, but without manual drawing and modeling. Simply provide a text description, and the space appears in seconds.

Genie 3 also supports "scripted events." Using text, you can, for example, change the weather or add characters. This makes the interaction even more flexible and lively.

However, not everyone will be able to try the technology yet. Genie 3 is being released as a limited research preview, available only to a small group of researchers and creators. This is to test risks and understand how to scale the system safely.

The company is actively developing the "world modeling" direction — earlier in December, it showcased Genie 2 and also created a special team led by the former co-leader of OpenAI Sora.

Source: The Verge