Genie 3 by Google

Genie 3 by Google DeepMind is a world model that can generate fully interactive 2D game environments directly from images, sketches, or text prompts. Trained on large-scale video data, it learns the underlying physics, dynamics, and controls of diverse environments without task-specific supervision. The model produces playable game scenes where users can control a character and interact with objects, enabling rapid prototyping of game mechanics and interactive simulations.

Key capabilities include transforming a single static image into a controllable environment, generating new levels consistent with a given visual style, and supporting responsive character movement and environment transitions. For researchers, Genie 3 provides a testbed for studying world models, embodied agents, and planning in learned environments.