Generative AI for Smart Cities: Mapping Models to Urban Applications - A Three-Axis Taxonomy

Generative artificial intelligence — generative adversarial networks, variational autoencoders, diffusion models, and large language models — has moved from research curiosity to applied tool across urban systems in under five years, now synthesizing mobility trajectories, generating household energy profiles, producing land-use designs, and functioning as spatio-temporal reasoning agents for city tasks. Despite this breadth, no existing survey classifies the field along model family, application task, and urban domain simultaneously, leaving practitioners without a principled basis for selecting a generative model for a given urban problem. This paper proposes a three-axis taxonomy — model family, task, and urban domain — documents a systematic multi-database literature search (9,763 records identified across four databases, 6,764 remaining after deduplication), presents a bibliometric keyword co-occurrence analysis corroborating the taxonomy’s structure, identifies five knowledge gaps, and formulates five testable hypotheses for evaluation through comparative experimentation.