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

Bruno Lepri; Ciro Beneduce; Massimiliano Luca

REVIEW 3 major objections 6 minor 16 references

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

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper shows that FLUX 1-schnell and Stable Diffusion 3.5-L encode implicit geographic knowledge of the United States, yet collapse the generic 'USA' prompt into a metropolitan stereotype that underrepresents rural, frontier, desert…

desk verdict The diversity-deficit result is probably real and worth publishing as a diagnostic; the finer 'geographic knowledge' claim is under-built because it rests on singular covariance estimates and eyeballed clustering. read the letter →

arxiv 2506.16898 v2 pith:EV5OGJRS submitted 2025-06-20 cs.AI cs.CVcs.CY

Ciro Beneduce , Massimiliano Luca , Bruno Lepri This is my paper · ORCID

classification cs.AIcs.CVcs.CY

keywords diffusionmodelsgeographicknowledgebiasurbanimagerytext-to-imagegenerationFIDDINO-v2representation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tests whether two open diffusion image models, FLUX 1-schnell and Stable Diffusion 3.5-L, have internal geographic knowledge of the United States and whether they use it fairly. The authors generate 150 street-view images for each U.S. state, each state capital, and a generic 'USA' prompt, embed them with DINO-v2, and measure pairwise Fréchet Inception Distance (FID). They find that geographically neighboring states and capitals cluster together in FID space, evidence of implicit geographic structure. But when prompted generically with 'USA', both models collapse onto a metropolitan stereotype, leaving frontier, desert, tropical, rural, and small-city environments far from the 'USA' reference. The paper concludes that the models know more geography than they deploy in national-level prompts.

What carries the argument

The argument is carried by the pairwise FID distance matrix built from DINO-v2 ViT-S/14 embeddings of the generated images. FID, a measure of dissimilarity between two sets of image embeddings, compares each location's embedding distribution through its mean and covariance; low FID means visually similar street scenes. Hierarchical clustering of these distances shows whether visual similarity mirrors real geography, while each location's FID to the generic 'USA' images quantifies how much that location is represented in the model's national stereotype.

What would settle it

Recompute pairwise FID with a shrinkage-regularized covariance estimator, or replace FID with a nonparametric distance such as maximum mean discrepancy on the same DINO-v2 embeddings; if geographically neighboring states no longer cluster together, the reported geographic knowledge is an artifact of unstable covariance estimation. Alternatively, shuffle location labels across image sets and show that random pairings produce clustering as strong as real geographic neighbors.

Watch

Extended reading notes

Core claim

The central claim is that FLUX 1-schnell and Stable Diffusion 3.5-L encode fine-grained implicit geographic knowledge of the United States, visible as clustering of geographically proximate states and capitals in FID space, while simultaneously reproducing a narrow metropolitan stereotype for the generic 'USA' prompt. Small state capitals with European-sounding names (Frankfort, Montpelier, Pierre, Dover, Olympia, Bismarck) are systematically misgenerated as European cityscapes, which the authors attribute to toponymic confusion and data sparsity. The result is a demonstrated gap between the geographic diversity the models can produce when asked for specific locations and the diversity they actually produce when asked for a broad region.

Load-bearing premise

The FID computation assumes the 150 image embeddings for each location form a Gaussian distribution and estimates a full 384x384 covariance matrix from just 150 samples without regularization, so the clusters that reveal 'geographic knowledge' could partly reflect estimation noise rather than true visual structure.

Editorial extensions

If this is right

If the models indeed encode geographic structure, researchers can use image-generation outputs and embedding distances as a probe of what a model has learned about place, without needing labels or external geographic data.
Any downstream urban-analysis pipeline that queries a model with a country-level prompt will inherit the metropolitan bias; generated scenario images for 'USA' will not represent the diversity of American built environments.
The systematic misgeneration of small capitals implies that place names with strong foreign-language associations need explicit geographic context in prompts to avoid visual cross-country confusion.
The FID-to-USA ranking provides a simple, reusable audit: compute the distance of every region's generated images to the generic country prompt to detect underrepresentation.
The knowledge-diversity gap suggests that dynamic prompting (adding region, biome, or city-size hints) could make models display geographic knowledge they already have.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same FID pipeline could test whether the 'USA' stereotype is driven by training-data overrepresentation rather than by prompt ambiguity, by comparing with a neutral prompt such as 'a street in a country' across multiple countries.
The clustering patterns might align more with climate and ecoregion boundaries than with state borders; if so, the 'geographic knowledge' may actually be environmental knowledge, and the paper's state-level framing would be one way to see it.
A natural extension is to measure whether this metropolitan collapse also occurs at other scales (continents, global regions) and whether it shrinks when the model is asked for 'rural USA' or 'small town USA' explicitly.
The small-capital misgeneration suggests a testable mitigation: adding the state name to the prompt ('Frankfort, Kentucky') should move those capitals toward correct American clusters; if it does not, the issue is visual training data rather than toponymic priors.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read it this morning. Quick take: the headline finding — both FLUX 1-schnell and SD 3.5-L collapse the generic "USA" prompt into a metropolitan stereotype, even though they show some per-state differentiation — is likely right and useful. The weaker link is the paper's other main claim, that the models encode fine-grained geographic knowledge. That claim leans on clustering of FID distances that are computed from singular covariance estimates and then interpreted by eye.

What's genuinely new: the per-state and per-capital FID analysis of two current open models, and the use of the same "USA" reference to expose which states and capitals are left out of the national stereotype. That's a reasonable, incremental contribution to the geographic-bias evaluation literature. The diversity-deficit table is the most convincing part: the gaps are large (Hawaii at 2594 vs. New Jersey at 331 for FLUX; Arizona at 3352 vs. Illinois at 413 for SD 3.5-L). Effects of that size are unlikely to be pure covariance noise, even if the point estimates for any specific rank are unstable.

The soft spot is Section 3.1. Each location is summarized by 150 DINO-v2 embeddings of dimension 384, so each empirical covariance is singular and high-variance; the FID trace term is ill-conditioned without regularization. No bootstrap, confidence interval, or alternative distance is reported. The clusters in Figure 2 are described qualitatively — "Mountain West," "Desert Southwest" — with no linkage criterion, no algorithm, and no quantitative match against known regions. The small-capital misgeneration section is basically anecdotal ("Olympia notably characterized by features of ancient Greek architecture"). No artifacts are released, so others can't re-run or sanity-check. Also, the SD 3.5-L reference points to the 2022 LDM paper, which is the wrong source for that model. That's sloppy in a short paper.

Those are real flaws, but they don't sink the core message. The metropolitan-stereotype finding is robust against the methodological caveats because it's a relative comparison against one reference and the effect sizes dwarf the likely noise. The fine-grained clustering should be treated as suggestive, not established. A serious referee would ask for regularized covariance or bootstrap intervals, a pre-registered clustering analysis with external validation against U.S. Census regions, and at least a sample of generated images released.

Who's this for? People evaluating geographic bias in text-to-image models, and the urban-AI community that actually uses these tools to generate scenario imagery. It's a workshop-grade empirical diagnostic, not a definitive study. I'd send it to peer review rather than desk-reject it; with moderate revisions (robustness checks, reframing the knowledge claim, fixing the citation) it would be a fine short paper. I'd bring it to a reading group if we had one that cared about generative-model evaluation, but I wouldn't treat its clustering results as evidence of structured geographic knowledge yet.

Referee Report

3 major / 6 minor

Summary. The paper investigates whether two open text-to-image diffusion models (FLUX 1-schnell and Stable Diffusion 3.5-Large) encode geographic knowledge of the United States and exhibit a national-scale representational bias. For each of the 50 states, their capitals, and a generic "USA" prompt, the authors generate 150 street-view images, embed them with DINO-v2 ViT-S/14, and compute pairwise Fréchet Inception Distances (FID) between all location pairs. They report that geographically proximate states and capitals cluster together in FID space, that small capitals with European-sounding names are systematically misgenerated as European cityscapes, and that the generic "USA" prompt yields images that are much closer in FID to large metropolitan states than to frontier, desert, tropical, or rural states. The paper concludes that the models possess detailed latent geographic knowledge but deploy a narrow metropolitan stereotype when prompted at the national scale.

Significance. If the central findings are supported, the paper makes a useful empirical contribution to the study of geographic bias in generative models, with implications for urban analytics, fairness, and the evaluation of text-to-image systems. The authors are to be credited for a systematic experimental design: 101 prompts, 150 images per prompt per model, a fixed prompt template, two current open models, and a public embedding model. The headline diversity-deficit result is qualitatively plausible and is partially supported by Table 1, which shows that states such as Hawaii, Alaska, and Arizona have much larger FID to the "USA" prompt than states like New Jersey or Illinois. However, the more ambitious claim of implicit geographic knowledge rests on visual inspection of unvalidated clusters computed from ill-conditioned covariance estimates, and the reported FID numbers are not accompanied by uncertainty intervals. The paper is therefore best treated as a valuable preliminary study whose central knowledge claim requires stronger statistical grounding before it can be taken as established.

major comments (3)

[§2.2 and §2.3, Eq. (1)] The FID computation uses each location's 150 DINO-v2 embeddings of dimension 384 to estimate a full 384×384 covariance matrix. With n=150 and d=384, each empirical covariance is singular and high-variance, and the matrix-square-root term in Eq. (1) is ill-conditioned. No regularization, shrinkage, dimensionality reduction, bootstrap, or confidence interval is reported. Because the pairwise FID matrix is the basis for the clustering in Figure 2 and the rankings in Table 1, the apparent geographic structure could partly reflect estimation noise. I request a robustness analysis: for example, regularized covariance estimation, PCA projection to fewer dimensions, or bootstrap confidence intervals around the FID values and cluster assignments.
[§3.1, Figure 2] The claim that geographically proximate states and capitals "cluster together" is based on visual inspection of dendrograms, but the manuscript does not specify the clustering algorithm, linkage criterion, cutoff, or any quantitative validation against known U.S. regions. Without a test such as comparing within-region versus between-region FID distances, a Mantel-style correlation between FID distance and geographic distance, or at least a reproducible cluster assignment, the groupings described in the text (e.g., the Mountain West cluster, Desert Southwest cluster, New England cluster) cannot be distinguished from arbitrary thresholding or subjective reading of the dendrogram. Please add a defined clustering procedure and a quantitative evaluation of its agreement with geography.
[§3.3, Table 1] The diversity-deficit ranking is more robust than the clustering claim because it compares every location to the same "USA" reference, but it inherits the same covariance-estimation instability. Table 1 reports only point estimates ranked by FID; no bootstrap intervals or significance tests are given, so the reader cannot tell whether the rank ordering of, say, New Jersey versus Hawaii is statistically meaningful. Additionally, the table mixes states and capital cities in a single ranking, which may confound two different prompt types; the interpretation would be cleaner if states and capitals were analyzed separately or explicitly modeled as distinct conditions. Please add uncertainty estimates and clarify whether the combined ranking is appropriate.

minor comments (6)

[§1, Introduction] The sentence "While it is fundamental to examine the geographic knowledge and biases that models encode is crucial" contains a grammatical error and should be split into two clauses; there is also a typo in "spaital" (spatial).
[References] Reference [13] is identical to reference [12] and both cite Rombach et al. 2022, but Stable Diffusion 3.5-Large is not introduced by that paper; the authors should cite the appropriate Stable Diffusion 3.5 technical report or model card.
[§2.1] The sentence "For each prompt, we generated images" does not state the number; the abstract says 150 images per prompt, but the main text should repeat that number explicitly for clarity.
[Figure 2] The cluster numbers referenced in the text (e.g., "Cluster 10" and "Cluster 11") are not easily identifiable in the figure; please annotate the dendrograms with the cluster labels used in the text.
[§3.2] The claim that misgenerated capitals "resemble European cities" is based on visual inspection of clustered images; since this is a secondary finding, it should be framed as an illustrative observation rather than a quantitative result, or supported by a content-based evaluation.
[General] No statement of data availability or code release is provided; sharing the generated-image lists, FID matrices, and analysis scripts would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical measurements of generated-image embeddings with no fitted parameters, target-derived quantities, or load-bearing self-citations.

full rationale

The paper's derivation chain is empirical and self-contained: it generates images from fixed prompts, embeds them with a pretrained DINO-v2 model, computes pairwise FID values via Equation (1), and then interprets the resulting clusters. There is no parameter fitted to any subset of the data and then renamed as a prediction; the FID distances are computed directly from the generated images and are not defined in terms of the geographic labels they are claimed to correlate with. The central claim, that geographically proximate states cluster in FID space, is a post-hoc interpretation of an independently computed distance matrix, not a quantity that is forced by construction. The only self-citations are references [1] and [2], which are cited in the introduction as examples of urban AI applications and are not load-bearing for the methodology, the FID computation, or the clustering interpretation. The statistical fragility of the covariance estimates (n=150 embeddings in a 384-dimensional space) is a validity concern, not a circularity concern, because it does not make any claimed result equivalent to its inputs by definition. Therefore the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The study is an empirical measurement with no fitted parameters. It relies on standard assumptions about the representational validity of DINO-v2 embeddings and the Gaussian modeling of embedding distributions, plus the representativeness of 150 samples per location.

assumptions (4)

domain assumption DINO-v2 ViT-S/14 embeddings capture visual semantics relevant to geographic and urban scenes.
The analysis treats distances between these embeddings as meaningful for place identity; this is assumed in Section 2.2 without validation on real street-view images.
domain assumption FID between embedding distributions is a valid measure of visual similarity between two sets of generated images.
FID is applied to DINO-v2 embeddings rather than the Inception features it was originally designed for; the paper does not justify this substitution (Section 2.3).
domain assumption The embedding distribution for each location is approximately multivariate Gaussian.
Equation (1) models each location as a Gaussian defined by mean and covariance; this is a standard FID assumption but is not tested.
domain assumption 150 generated images per prompt are a representative sample of the model's output distribution for that prompt.
Section 2.1 fixes 150 images per location, but there is no convergence analysis to show that this sample size stabilizes the FID estimates.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI's Blind Spots: Geographic Knowledge and Diversity Deficit in Generated Urban Scenario." pith.science (2026). https://pith.science/paper/EV5OGJRS

@misc{pith2026250616898,
  author       = {Pith},
  title        = {Pith review of: AI's Blind Spots: Geographic Knowledge and Diversity Deficit in Generated Urban Scenario},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EV5OGJRS}},
  note         = {Machine review of arXiv:2506.16898}
}

read the original abstract

Diffusion-based text-to-image models are increasingly used for urban analysis and scenario generation, but their geographic knowledge and representational biases remain poorly understood. We evaluate FLUX 1-schnell and Stable Diffusion 3.5-Large in the United States by generating 150 street-view images for each state, each state capital, and a generic "USA" prompt. Images are embedded with DINO-v2 ViT-S/14 and compared with Fr\'echet Inception Distance (FID). Pairwise FID clustering shows that geographically proximate states and capitals often group together, indicating implicit geographic structure. However, the generic ``USA'' prompt collapses this diversity into a metropolitan stereotype: frontier, desert, tropical, rural, and small-city environments are underrepresented or distant in FID space. These results show that diffusion models can encode fine-grained geography while still reproducing narrow national-scale visual stereotypes.

Figures

Figures reproduced from arXiv: 2506.16898 by the authors.

**Figure 2.** Hierarchical clustering of FID for states and capitals for FLUX 1-schnell (top) and Stable Diffusion 3.5-L.(bottom) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 canonical work pages

[1]

Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Large language models are zero-shot next location predictors.IEEE Access(2025)

work page 2025
[2]

Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Urban Safety Percep- tion Through the Lens of Large Multimodal Models: A Persona-based Approach. arXiv preprint arXiv:2503.00610(2025)

arXiv 2025
[3]

Melissa Hall, Samuel J Bell, Candace Ross, Adina Williams, Michal Drozdzal, and Adriana Romero Soriano. 2024. Towards geographic inclusion in the evaluation of text-to-image models. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. 585–601

work page 2024
[4]

Melissa Hall, Candace Ross, Adina Williams, Nicolas Carion, Michal Drozdzal, and Adriana Romero Soriano. 2024. DIG In: Evaluating Disparities in Image Generations with Indicators for Geographic Diversity. arXiv:2308.06198 [cs.CV] https://arxiv.org/abs/2308.06198

arXiv 2024
[5]

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems30 (2017)

2017
[6]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

2020
[7]

Reddy, and Sunipa Dev

Akshita Jha, Vinodkumar Prabhakaran, Remi Denton, Sarah Laszlo, Shachi Dave, Rida Qadri, Chandan K. Reddy, and Sunipa Dev. 2024. ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation. arXiv:2401.06310 [cs.CV] https://arxiv.org/abs/2401.06310

arXiv 2024
[8]

Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

2024

Show all 16 references

[9]

Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. 2022. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. arXiv:2108.01073 [cs.CV] https://arxiv.org/abs/2108.01073

2022 arXiv
[10]

Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2022. GLIDE: Towards Photo- realistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv:2112.10741 [cs.CV] https://arxiv.org/abs/2112.10741

2022 arXiv
[11]

Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabba...

2024 arXiv
[13]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

2022
[14]

Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. 2022. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv:2209.14792 [cs.CV] https://arxiv.or...

2022 arXiv
[15]

Yixin Wan, Arjun Subramonian, Anaelia Ovalle, Zongyu Lin, Ashima Suvarna, Christina Chance, Hritik Bansal, Rebecca Pattichis, and Kai-Wei Chang. 2024. Sur- vey of Bias In Text-to-Image Generation: Definition, Evaluation, and Mitigation. arXiv:2404.01030 [cs.CV] https://arxiv.o...

2024 arXiv
[16]

Qingyi Wang, Yuebing Liang, Yunhan Zheng, Kaiyuan Xu, Jinhua Zhao, and Shenhao Wang. 2025. Generative AI for Urban Planning: Synthesizing Satellite Imagery via Diffusion Models. arXiv:2505.08833 [cs.CV] https://arxiv.org/abs/ 2505.08833

2025 arXiv
[17]

Yuxin Yang, Pengfei Zhu, Mengshi Qi, and Huadong Ma. 2024. Uncovering the human motion pattern: Pattern Memory-based Diffusion Model for Trajectory Prediction. arXiv:2401.02916 [cs.CV] https://arxiv.org/abs/2401.02916

2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Large language models are zero-shot next location predictors.IEEE Access(2025)

work page 2025

[2] [2]

Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Urban Safety Percep- tion Through the Lens of Large Multimodal Models: A Persona-based Approach. arXiv preprint arXiv:2503.00610(2025)

arXiv 2025

[3] [3]

Melissa Hall, Samuel J Bell, Candace Ross, Adina Williams, Michal Drozdzal, and Adriana Romero Soriano. 2024. Towards geographic inclusion in the evaluation of text-to-image models. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. 585–601

work page 2024

[4] [4]

Melissa Hall, Candace Ross, Adina Williams, Nicolas Carion, Michal Drozdzal, and Adriana Romero Soriano. 2024. DIG In: Evaluating Disparities in Image Generations with Indicators for Geographic Diversity. arXiv:2308.06198 [cs.CV] https://arxiv.org/abs/2308.06198

arXiv 2024

[5] [5]

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems30 (2017)

2017

[6] [6]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

2020

[7] [7]

Reddy, and Sunipa Dev

Akshita Jha, Vinodkumar Prabhakaran, Remi Denton, Sarah Laszlo, Shachi Dave, Rida Qadri, Chandan K. Reddy, and Sunipa Dev. 2024. ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation. arXiv:2401.06310 [cs.CV] https://arxiv.org/abs/2401.06310

arXiv 2024

[8] [8]

Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

2024