Bilawal Sidhu demonstrates Project Genie, the web experience built around Google DeepMind's Genie 3 world model. A creator separately describes an environment and a character, chooses a first-person or third-person view, refines the Nano Banana-generated starting image, and then enters a world whose pixels are generated live as the user moves.
The model often learns useful visual affordances without explicit programming. A GPS display can stay aligned with movement, scenes can produce atmospheric light and reflections, and familiar environments can be extended beyond the reference image. The same tests also reveal unstable faces, mutable geometry, inconsistent controls, and limited interaction, showing why the system is not yet a conventional game engine.
Genie 3 co-lead Jack Parker-Holder and product manager Diego Goberna describe the release as a way to discover uses that the team has not anticipated. Immediate possibilities include rapid game-world prototyping, virtual camera scouting for filmmakers, educational simulations, and stepping into personal photographs or artwork. The model's unpredictability can become part of the creative experience because two users may discover different behaviors from the same starting prompt.
The current 60-second limit is a product choice rather than a fundamental model boundary. Longer sessions could use more context, external spatial memory, or simple continuation from a final frame. DeepMind is also exploring more dynamic scenes, richer interactions, stronger real-world simulation, and combinations with other generative models, while user feedback determines which product directions matter most.
Watch on YouTube
