AI's Blind Spots: Geographic Knowledge and Diversity Deficit in Generated Urban Scenario

Ciro Beneduce, Massimiliano Luca, Bruno LepriarXiv 2025 · arXiv · OpenReview

Abstract

As image generation models like FLUX 1 and Stable Diffusion 3.5 become integral to urban design and architectural visualization, understanding their internal "geographic maps" is crucial. This study analyzes 30,300 synthetic images generated for U.S. states and capitals. While these models demonstrate sophisticated regional knowledge, they exhibit a profound "representative bias." When prompted for a generic "USA" street view, models default to dense metropolitan imagery, effectively marginalizing rural and non-urban identities. Furthermore, models struggle with toponymic disambiguation, often confusing U.S. capitals with their European namesakes.

Introduction

Generative AI is no longer just a creative tool; it is increasingly used for simulating future urban environments and social scenarios. However, if the training data is skewed toward global hubs, the resulting "mental map" of the AI may erase the visual diversity of actual human settlements. This research probes whether state-of-the-art diffusion models understand the nuances of American geography or if they rely on a narrow set of urban stereotypes.

Methodology

The researchers systematically tested FLUX 1-schnell and Stable Diffusion 3.5-Large (SD-3.5-L) using a standardized prompt: “A photorealistic high-resolution street-view photo of {LOCATION}”. The locations included all 50 U.S. states, their capitals, and a general "USA" baseline.

To quantify visual similarity, the team used DINO-v2 ViT-S/14 to generate feature embeddings for each image. They then calculated the Fréchet Inception Distance (FID) between locations. In this context, a lower FID indicates that the model perceives two locations as visually similar, serving as a proxy for the model's internal categorization of space.

Geographic Knowledge vs. The "USA" Stereotype

The models display a surprising level of latent geographic knowledge. Clustering FID scores revealed that the models naturally group states with similar landscapes, such as the Mountain West or New England. However, this knowledge is often bypassed by a strong metropolitan default when specific constraints are removed.

Comparison of specific city prompts vs. the generic "USA" prompt across FLUX 1 and Stable Diffusion 3.5. Note how "USA" defaults to iconic skyscrapers regardless of the model.

As shown in the figure above, while the models can accurately depict the mountainous backdrop of Juneau or the historic brickwork of Boston when specifically asked, the generic "USA" prompt triggers a stereotypical image of a dense metropolis dominated by skyscrapers. The FID distance between "USA" and urban-heavy states like New Jersey or Illinois was minimal, whereas the distance to states like Alaska, Hawaii, or Arizona was up to eight times higher. To these AIs, the default visual identity of America is a big city.

Toponymic Confusion and the "European Glitch"

A significant finding was the models' struggle with entity disambiguation. Small U.S. capitals with names shared by European locations often triggered "Old World" architectural styles. For example:

  • Bismarck and Pierre were frequently depicted as traditional German or French towns.
  • Olympia, Washington often resulted in imagery featuring ancient Greek ruins or neoclassical architecture.
  • Frankfort and Montpelier clustered together in the models' latent space due to shared European visual features rather than their actual American geography.

Conclusion

While current diffusion models possess the latent capacity to represent diverse geographic regions, they suffer from a "diversity deficit" in their default outputs. The tendency to equate "America" with "Big City" creates a representative bias that could have real-world implications if these models are used for planning or social representation. Future work must focus on strategies to ensure more equitable and diverse visual representations, preventing the visual erasure of rural and non-metropolitan identities by AI stereotypes.

Summary updated

Abstract

Image generation models are revolutionizing many domains, and urban analysis and design is no exception. While such models are widely adopted, there is a limited literature exploring their geographic knowledge, along with the biases they embed. In this work, we generated 150 synthetic images for each state in the USA and related capitals using FLUX 1 and Stable Diffusion 3.5, two state-of-the-art models for image generation. We embed each image using DINO-v2 ViT-S/14 and the Fr\'echet Inception Distances to measure the similarity between the generated images. We found that while these models have implicitly learned aspects of USA geography, if we prompt the models to generate an image for "United States" instead of specific cities or states, the models exhibit a strong representative bias toward metropolis-like areas, excluding rural states and smaller cities. {\color{black} In addition, we found that models systematically exhibit some entity-disambiguation issues with European-sounding names like Frankfort or Devon.
BibTeX
@misc{beneduce2025ai,
  author = {Ciro Beneduce and Massimiliano Luca and Bruno Lepri},
  title = {{AI's Blind Spots: Geographic Knowledge and Diversity Deficit in Generated Urban Scenario}},
  year = {2025},
  eprint = {2506.16898},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.16898},
}
Figure