DeepMind connected its Genie 3 world model to Street View imagery so users can drop a pin on a map

DeepMind connected its Genie 3 world model to Street View imagery so users can drop a pin on a map and generate a walkable, AI-built environment that begins from real photos.

On May 20, 2026, DeepMind demonstrated a prototype that pairs its Genie 3 world model with Street View imagery to produce explorable, AI-generated environments tied to real locations. The link lets users start from actual map coordinates and immediately enter a walkable scene grounded in Street View photos rather than an entirely synthetic map, a step DeepMind says matters for training agents in realistic settings.

The demo workflow is user-driven: someone drops a pin on a map, optionally selects an aesthetic such as "Ocean World," "Desert Sands," "Stone Age," or "B&W film," and describes a character. Genie 3 then constructs a navigable world whose starting point is rooted in the corresponding Street View footage, using an interface DeepMind calls Maps Imagery Grounding to anchor model outputs to real imagery.

DeepMind emphasizes that the underlying advantage is the Street View archive itself. Years of imagery collection across roads, interiors, waterways and remote areas provide both training material and a visual anchor for generative worlds, which DeepMind and its spokespeople present as a competitive edge for producing more realistic, location — specific simulations than purely synthetic environments.

The team positions the feature primarily as a research and training tool rather than a consumer game. DeepMind says the world model will help agents learn to navigate, reason and act in realistic settings; its agent SIMA 2 already uses Genie for training, and Waymo has applied the platform to simulate street scenarios for self-driving tests. Connecting to Street View allows those simulation runs to be pinned to specific, real-world coordinates.

Public demos and early tests have shown impressive variety alongside obvious rough edges. Former Google product manager Bilawal Sidhu shared test prompts on May 19 that included a flooded Golden Gate Bridge, a 1920s — styled Fort Worth Stockyards, a branded Formula 1 car racing down the Las Vegas Strip, riding past the Palace of Fine Arts as a squirrel on a scooter, and walking through the White House using indoor Street View. Observers noted visible artifacts such as soft textures, unstable geometry and surreal transitions in some scenes.

Access is currently gated: the feature is rolling out to Google AI Ultra subscribers (priced at $200/month, 18+), and real-location generation works only for U.S. spots for now, with plans to expand. DeepMind labels the release an experimental research prototype and says the team is working on refinements — an important caveat for builders who may test these anchored simulations in agent, robotics, or self-driving development pipelines.