Google has announced a new version of its virtual world generation model, Genie 3. This neural network is capable of creating interactive environments in real time based on a simple text description or an image. The maximum quality of the generated worlds reaches 720p at 24 frames per second.

Among the key improvements is a significantly expanded model "memory": while the previous version could only maintain the integrity of a generated world for 10 seconds, Genie 3 remembers locations for several minutes. Additionally, it is now possible to modify worlds in real time with text commands, adding new objects or changing the weather with realistic physics.

Despite its impressive capabilities, Genie 3 has limitations. The model cannot recreate real-world locations and sometimes produces "hallucinations," such as incorrect character movements. Currently, access to Genie 3 is restricted, but Google plans to make it available to a limited number of experts and researchers for further refinement.

ORIENT