Google DeepMind unveils Genie 2, generating playable 3D worlds from a single image

Google DeepMind/Genie world models

Google DeepMind introduced Genie 2, a world model that turns a single image into a playable 3D environment controllable with keyboard and mouse. Worlds stay consistent for up to a minute with emergent abilities like object interaction, NPC behavior and physics simulation. The aim is an endless supply of training and test environments for AI agents.

• Generates playable 3D worlds from single image; controllable with keyboard and mouse
• Worlds consistent for up to one minute; majority of examples last 10-20 seconds
• Emergent abilities: object interactions, character animation, physics, NPC behavior, water/smoke/lighting effects
• Use cases: agent training, rapid prototyping of interactive experiences from concept art or text