GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity in Text-to-Image Street-View Generation
This research tested text-to-image models to see if they could generate specific road segments rather than just generic city scenes. The GeoFidelity-Bench benchmark compared generated images against real street views, finding that models struggle to distinguish the exact target segment from nearby local alternatives, even when given street and neighborhood names.