Google DeepMind has formally introduced Genie 2, a "foundation world model" that can generate playable 3D environments based on a single prompt image. According to Google, Genie 2 represents a massive improvement over Genie 1, which could generate 2D worlds. Simulating a 3D world also includes actions such as jumping and swimming.
With complex character interaction, physics, lighting, and more, it's an impressive demonstration that Google says could be used for creative workflows and creative workflows for prototyping interactive experiences. A game developer could create detailed concept art or even a pre-rendered image and then use Genie 2's model to see how it might look as a playable experience.
Even more interesting is the diverse trajectories; the simulation and what you see will change depending on the input or movement. For example, taking a different path at a fork in the road in a first-person fantasy Genie 2 creation will lead to a different environment.
And unlike the recent AI-generated Minecraft tech demo we reported on, Genie 2 can remember where you've been and what's no longer in view, so they can be accurately rendered if you retread familiar ground. And yes, it can handle various genres and perspectives, including first-person action, racing games, and isometric affairs.
You can even interact with objects and characters. In one example, the player shoots a red explosive barrel in a Counter-Strike-like world. Granted, the physics and explosion look weird in that AI-generated way. Still, it's impressive, especially when it can handle real-time reflections and create playable experiences from real-world photos.
"Genie 2 shows the potential of foundational world models for creating diverse 3D environments and accelerating agent research," Google writes. "This research direction is in its early stages, and we look forward to continuing to improve Genie's world generation capabilities in terms of generality and consistency."
Check out Google DeepMind's full Genie 2 article for more information and demonstrations of Genie 2's capabilities.
- Read more: Oasis AI and a single NVIDIA H100 GPU has created a playable 'AI Minecraft' at 720p 20 FPS
- Read more: AI creates a playable version of the original Doom, generating each frame in real-time
- Read more: Google CEO fires one off at Microsoft by saying its ready for the AI race 'any day, any time'
- Read more: Google announces generative video model 'Veo' to compete with OpenAI's impending Sora AI model