A Proxy-Based Method for Measuring Racial Disparities in~Text-to-Image~Generators

AI summary

Overview

Research area: Algorithmic fairness and bias auditing for text-to-image generation, combining computer vision measurement techniques with U.S. anti-discrimination law.

Technical level: Intermediate. The paper is readable without legal or ML specialization, but it assumes some familiarity with embedding similarity, color spaces (CIELAB), and distribution distance metrics (Total Variation Distance).

Scope (one sentence): The paper proposes and empirically demonstrates an automated, legally grounded auditing method that uses occupation, geography, and surname "proxies" plus skin-tone and image-homogeneity analysis to measure racial disparities in three text-to-image models against U.S. Census benchmarks.

What This Paper Is About

Text-to-image generators frequently default to depicting white people for prompts that never mention race, and prior audits typically measure success by counting racial proportions in outputs — an approach the authors argue is both legally vulnerable (courts have called racial balancing "patently unconstitutional") and too coarse to catch subtler disparities. The paper adapts the facially-neutral proxy variables long used in U.S. employment, housing, and voting-rights litigation (occupations, geographic areas, surnames) into prompts for image generators, then compares the resulting racial distributions against Census data. It adds skin-tone measurement and image-similarity analysis to surface biases — such as varying image quality or stereotypical repetition within groups — that demographic comparisons alone would miss.

Key Contributions

1. A legal foundation for proxy-based audits. The authors trace how courts and federal agencies have accepted race-correlated but facially-neutral variables (ethnic names under Title VII, neighborhood-based lending under the Fair Housing Act, "bizarre" district maps in racial gerrymandering cases) and argue these same proxies can be adapted to audit image generators without requiring explicit racial targets.
2. An automated auditing pipeline. The method classifies each generated image by race, compares the model's distribution to a reference demographic distribution, measures skin tone in CIELAB, and computes intra-group versus inter-group embedding similarity — implemented end-to-end so that no manual classification is involved.
3. An empirical study across three models and three proxy categories. Using 150 prompts total (50 variants each for occupation, geography, and surname) and 100 generated images per prompt, the authors audit Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, and Flux Schnell.
4. A "homogeneity gap" metric. Defined as the difference between mean intra-group and inter-group cosine similarity, computed over both CLIP embeddings (overall composition and style) and InsightFace facial embeddings (facial morphology), to detect stereotypical repetition within racial and skin-tone groups.

Main Findings

• Black-classified images are consistently underrepresented. Across all three tested models, Black-classified images fall below the Census demographic line — shown in Figure 2 as the preponderance of blue points below the diagonal.
• White overrepresentation splits by model. White-classified images are underrepresented by Stable Diffusion 3.5 Large Turbo but overrepresented by Flux Schnell and Imagen 4 Fast. Imagen 4 Fast overrepresents white-classified images by 20% for occupational proxies.
• Stable Diffusion 3.5 Large Turbo diverges most from Census data. Its overall average TVD is 0.452, versus 0.310 for Flux Schnell and 0.312 for Imagen 4 Fast. Its surname TVD is the largest single figure reported (0.523).
• Imagen 4 Fast has the largest Black-representation disparities on surname and geography prompts. It is most accurate when the Black share of the ground truth is low (for example, "person with the last name Chen" or "person who works as an information systems manager"), but it does not reach the Census percentage for any prompt where Black representation exceeds 5%.
• Hispanic representation moves in opposite directions. Imagen 4 Fast is the only model to consistently over-represent Hispanic-classified images in geography proxies, whereas Stable Diffusion 3.5 Large Turbo under-represents Hispanic-classified images across all proxies.
• Skin tone varies sharply by model. For Black-classified images, Stable Diffusion 3.5 Large Turbo has an average L* of 31.5, compared with 48.0 for Flux Schnell and 50.1 for Imagen 4 Fast. Imagen 4 Fast produces the lightest and most homogeneous skin tones and generated the lowest number of Black-classified images (172) across the models tested.
• Most very dark skin tones are read as Black. Across models, most skin tones with an L* value below 40 are classified as Black.
• Surname accuracy depends on national-origin distinctiveness. Surnames like "Schmidt", "Chen", and "Kim" have TVDs below 0.1 across all models, whereas surnames with higher Black demographic shares (e.g., "Charles") are consistently underrepresented.
• Occupational stereotyping appears within a single model. For Stable Diffusion 3.5 Large Turbo, Asians are overrepresented in "a CEO" and "an actor" prompts and underrepresented in "a nail technician" prompts, with occupational prompts showing the most distortion from Census data.
• Skin tone and salary are not significantly related. In Flux Schnell's Black-classified occupational outputs, the relationship between higher L* (lighter tones) and higher median earnings is not statistically significant, though higher L* values appear for stereotypically female occupations such as nail technician, nursing assistant, and school teachers.
• Black-classified images are more self-similar than other groups. This holds across all three models and is most pronounced in Flux Schnell, which has a 0.110 facial homogeneity gap for Black-classified images despite otherwise producing the most diverse facial embeddings.
• Asian homogeneity is stylistic rather than facial in two models. CLIP gaps are 0.088 in Stable Diffusion 3.5 Large Turbo and 0.106 in Imagen 4 Fast, versus facial gaps of 0.065 and 0.083 respectively.
• Darker images are more repetitive. In Imagen 4 Fast, images with L* < 40 reach a mean facial similarity of 0.405 versus 0.183 for L* in [40, 55), a difference of 0.222 (based on a small sample of 40 images). Flux Schnell declines from 0.342 (L* < 40) to 0.251 (L* ≥ 70).
• Classifier reliability is substantial but uneven. Cohen's Kappa was 0.786 between the two human raters and an average of 0.715 between humans and Gemini 2.5 Flash on a random sample of 300 images; the Hispanic category alone had a Fleiss kappa of 0.463, with 59% of Hispanic disagreements against White and 33% against Asian.

Methodology in Plain English

The pipeline works in four stages. First, the team picks three categories of prompts whose real-world demographics are known: occupations, geographic areas, and surnames. They select 50 instances per category by sampling the highest-represented occupations, surnames, and locations for each racial group in the Census data, then wrap each instance in a fixed template — "A person who works as a [occupation]", "A person from [geographic area]", "A person with the last name [surname]" — followed by a style suffix requesting a photorealistic ID-style portrait of exactly one person, to control for image style. Each prompt is submitted 100 times to each model.

Second, each generated image is sent to Gemini 2.5 Flash with an instruction to classify apparent race/ethnicity into exactly one of White, Black, Asian, Hispanic/Latino, or Other. The authors validate this classifier against two human annotators on 300 randomly selected images. The resulting distribution is compared to U.S. Census data using Total Variation Distance, which ranges from 0 (identical) to 1 (no overlap), giving one number per model per proxy category while Figure 2 plots the per-race deviations from the diagonal.

Third, skin tone is extracted automatically rather than judged by humans. A Bilateral Segmentation Network trained on CelebAMask-HQ for 19-class face parsing identifies skin pixels, mean RGB values are converted into CIELAB, and the authors derive lightness (L*), chroma (C*), and hue angle, following the approach of Wilson, Ghosh, and Caliskan (2025). They use ΔE2000 to quantify perceptible color difference and thus skin-tone diversity within a group.

Fourth, they measure how repetitive a model is. Two kinds of embeddings are extracted — CLIP ViT-H-14 pretrained on LAION2B for overall appearance and InsightFace buffalo_l for faces (filtered to detections above 0.85) — and pairwise cosine similarity is computed within groups defined by race, prompt, and skin tone. Subtracting inter-group from intra-group similarity yields the "homogeneity gap": a positive value means a group's images look more like each other than like outsiders, a signal of stereotypical representation.

Why This Matters

The paper's distinctive move is to treat auditing as a measurement and disclosure problem rather than a mandate to hit demographic targets. Because the proxies are facially neutral and the output is a factual comparison against public benchmark data, the authors argue the method sits on the permissible side of recent First Amendment rulings on platform regulation, unlike audits that dictate representational outcomes.

Impact on research: It provides an alternative to audits that count white versus non-white proportions averaged across prompts, enabling per-prompt and per-group comparisons, and it links measurement choices to legal doctrine — a connection the authors say rarely appears explicitly in algorithmic-bias research.

Real-world applications:

• School districts and public institutions procuring image-generation tools for classrooms, where the authors' motivating example is a third grader repeatedly shown "a lawyer" as a white man.
• Compliance-style audits of automated employment or educational decision tools, following the model of New York's NYC Local Law 144 of 2021 requirement for annual bias audits.
• Vendor evaluation during procurement, where buyers can compare models on granular disparities before purchase rather than relying on vendor claims.
• Media and publishing workflows that need evidence about whose faces and which skin tones a generator reproduces repetitively.

Industry relevance: The three audited models span a large open-weight model with prior bias evaluations (Stable Diffusion 3.5 Large Turbo), a major vendor's fast model (Imagen 4 Fast from Google), and a smaller start-up's model (Flux Schnell from Black Forest Labs). The finding that the model closest to Census data is not the one with the most diverse skin tones — and vice versa — shows that a single "fairness" score would hide exactly the trade-offs procurement decisions must confront.

Future Directions

• Extending beyond the U.S. context. The authors state directly that adapting the method would require locally relevant proxies and demographic benchmarks, since their proxies and legal framing are calibrated to American data.
• Improving racial classification reliability. The Hispanic category was the clear weak point (Fleiss kappa 0.463, with Gemini more likely than humans to label White-classified images as Hispanic), and the authors note that unfine-tuned classifiers are most reliable for Black, Asian, and White populations, with other groups collapsed into "other."
• Scaling the darker-skin-tone analysis. The comparison of facial similarity across skin-tone bins rested on a small sample for Imagen 4 Fast, where only 40 images had L* < 40; larger samples would test whether the 0.405 versus 0.183 gap holds.
• Resolving the normative question the paper deliberately leaves open. The authors note that only five percent of U.S. lawyers are Black, and ask whether models should mirror current representation or present an aspirational future — a choice they argue institutions must make explicitly, with audit data in hand.

Target Audience

Policy analysts and procurement officers at public institutions who must justify tooling decisions; AI ethics and fairness researchers looking for a granular, automatable alternative to proportion-counting audits; legal scholars working at the intersection of anti-discrimination doctrine and algorithmic systems; and model developers who want to know which disparities their evaluations currently miss — particularly the finding that demographic accuracy and skin-tone diversity can diverge within the same model.

Authors’ abstract

We present an automated method for auditing text-to-image generators for representational bias against ground-truth demographic benchmarks. Our approach adapts the racial proxy variables used in U.S. antidiscrimination law to provide a fine-grained comparative analysis of variations in racial representation in generated images. We classify each generated image by race and then compare the model's distribution to publicly available demographic benchmark data to identify patterns of over- and under-representation. We also measure skin tone using CIELAB values and calculate cosine similarity amongst groups of images in order to capture biases that don't show up in demographic comparisons such as differences in image quality across races and skin tones. We report on experiments testing three models (Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, and Flux Schnell) using three categories of proxies (geographic, occupational, and surname) and 50 variants within each category. We find notable variation across the models as well as by proxy category. For example, Imagen 4 Fast overrepresents white classified images more than the other models tested and produces images with lighter, more homogeneous skin tones. Imagen 4 Fast also had the largest representational disparity for Black demographics among surname and geographical prompts. Stable Diffusion 3.5 Large Turbo produces results that overall diverge most from census data but yield the most diverse range of skin tones. Our method yields a generalizable and legally-grounded auditing tool that surfaces detailed information about model outputs without relying on implicit normative definitions of fairness.