Google DeepMind has announced the launch of its enhanced AI “world” model named Genie 3, designed to create 3D environments for real-time interaction by both users and AI agents. This new iteration promises a more immersive experience, allowing users to engage with these created worlds for extended periods and ensuring that the model retains visual memory, even when users momentarily look away.
World models, a branch of AI technology, serve to replicate environments for various applications, including education, entertainment, and training for robots or AI agents. Users can provide prompts that lead to the generation of dynamic spaces akin to video games, with the distinction that these environments are constructed through AI rather than traditional 3D assets. Google has recently ramped up its efforts in this area, having showcased Genie 2 in December—an earlier model that could generate interactive worlds based on images—and is building a specialized world models team under the leadership of a former co-lead from OpenAI’s Sora video generation tool.
Genie 2, however, exhibited several limitations, including a gameplay duration restricted to just a minute. Users have found experiences with other AI-driven interactive video projects, notably one supported by Pixar’s co-founder, to render environments similar to a blurry Google Street View, where unexpected changes occurred as users explored.
Genie 3 appears to represent a significant advancement over its predecessor. The updated model allows users to create worlds with prompts that can sustain interactive experiences for a few minutes, a considerable increase from the mere 10 to 20 seconds previously offered by Genie 2, as detailed in a blog post. Moreover, Genie 3 is capable of maintaining visual memory for approximately one minute, enabling users to return to previously viewed objects—such as wall decorations or text on chalkboards—without losing their original locations. The worlds produced will feature a resolution of 720p at 24 frames per second.
Additionally, DeepMind is introducing what they term “promptable world events” into Genie 3. This feature allows users to modify aspects of the environment, such as adjusting weather conditions or incorporating new characters, based solely on user prompts.
Nonetheless, access to Genie 3 will be limited, as it will launch as a “limited research preview” intended for a select group of academics and creators. This approach aims to help developers identify risks associated with its use and devise strategies to mitigate them, as stated by Google. Users will face constraints on interactions with the generated worlds, and clear text will frequently only appear when included in the input description of the environment. Google has expressed interest in expanding access to additional testers in the future.
