Interactive World Models
Generative models that produce explorable, action-conditioned 3D/video environments in real time from a prompt.
Unlike video generators, interactive world models render frames conditioned on a user’s actions and the history of prior frames, keeping a navigable environment spatially consistent in real time.
They emerged as a serious research frontier in 2025 (e.g. Genie 3 at 720p/24fps for minutes of consistency) for embodied agents, simulation, and training-data generation.