2026-05-19 バージニア工科大学（Virginia Tech）
A series of AI-generated images from Virginia Beach, Richmond, and Washington, D.C. These images are closer to the cities they are supposed to represent, getting more accurate as the population of each city increases. Some details, such as the incorrect traffic light in the Washington, D.C., are still wrong. Images courtesy of Jungwhan Kim.
＜関連情報＞
- https://news.vt.edu/articles/2026/05/cnre-research-AI-cities.html
- https://www.sciencedirect.com/science/article/abs/pii/S0160791X26001491
都市を想像する：DALL·E 2を用いたAI生成都市景観における視覚的リアリズムとアイデンティティの評価 Imagining the city: Evaluating visual realism and identity in AI-generated cityscapes using DALL·E 2
Junghwan Kim, Nami Jain, Kee Moon Jang, Xinyue Ye
Technology in Society Available online: 15 April 2026
DOI:https://doi.org/10.1016/j.techsoc.2026.103360
Highlights
- Conducted one of the first human-centered evaluations of AI-generated urban imagery.
- Assessed GenAI’s effectiveness in conveying visual realism and identity.
- Found that longer-term residents tended to evaluate GenAI-generated images more critically.
- Identified that GenAI performs better for larger cities than for smaller towns.
- Discussed ethical considerations for the application of GenAI in urban planning and design.
Abstract
Despite rapid advances in Generative Artificial Intelligence (GenAI) and growing interest in its use for urban planning and design, its ability to authentically represent urban environments remains poorly understood. Most existing studies rely on online materials or expert evaluations, overlooking human-centered approaches that draw on the nuanced, place-based insights of local residents. This study aims to systematically evaluate the performance of OpenAI’s DALL·E 2 in conveying both visual realism and identity in urban contexts. We conducted an online survey of 129 respondents, who assessed AI-generated images of one small town and three large cities in Virginia, each emphasizing one of Kevin Lynch’s five urban elements. For each image, we quantified an AI-generated photo alignment score capturing two interrelated dimensions: visual realism and representational fidelity to the city’s visual identity. Results show that images of landmarks received lower alignment scores than those of other elements, such as districts and edges, indicating GenAI’s limitations in capturing culturally and historically significant urban features. Longer-term residents provided more critical evaluations, reflecting the role of place-based knowledge in shaping perceptions of AI imagery. Additionally, larger metropolitan areas received higher alignment scores than the small town of Blacksburg, revealing an uneven geography in GenAI’s capabilities. This study delivers one of the first systematic, human-centered evaluations of AI-generated urban imagery, introducing a novel methodological framework that integrates local knowledge with place-based evaluation. The findings highlight critical ethical considerations for applying GenAI in urban planning and design, particularly in contexts where equitable and authentic representation is essential.
