GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension

Bochu Ding; Brinnae Bent; Augustus Wendell

cs.AIApr 10, 2026

GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension

Authors: Bochu Ding, Brinnae Bent, Augustus Wendell

Organizations: Duke University

Abstract

Text-to-image (T2I) models, and their encoded biases, increasingly shape the visual media the public encounters. While researchers have produced a rich body of work on bias measurement, auditing, and mitigation in T2I systems, those methods largely target technical stakeholders, leaving a gap in public legibility. We introduce GLEaN (Generative Likeness Evaluation at N-Scale), a portrait-based explainability pipeline designed to make T2I model biases visually understandable to a broad audience. GLEaN comprises three stages: automated large-scale image generation from identity prompts, facial landmark-based filtering and spatial alignment, and median-pixel composition that distills a model's central tendency into a single representative portrait. The resulting composites require no statistical background to interpret; a viewer can see, at a glance, who a model 'imagines' when prompted with 'a doctor' versus a 'felon.' We demonstrate GLEaN on Stable Diffusion XL across 40 social and occupational identity prompts, producing composites that reproduce documented biases and surface new associations between skin tone and predicted emotion. We find in a between-subjects user study (N = 291) that GLEaN portraits communicate biases as effectively as conventional data tables, but require significantly less viewing time. Because the method relies solely on generated outputs, it can also be replicated on any black-box and closed-weight systems without access to model internals. GLEaN offers a scalable, model-agnostic approach to bias explainability, purpose-built for public comprehension, and is publicly available at https://github.com/cultureiolab/GLEaN.

Figures & tables

Figure 1: Workflow overview of GLEaN, with example outputs for the “business executive” prompt.

Figure 2: (Top-down, left to right) Row 1: An immigrant, a capitalist, a convict, a welfare recipient, a business executive, a janitor, a street vendor, an astronaut. Row 2: A deportee, a cab driver, a construction worker, a trust-funder, a homeless person, a farmer, a pharmacist, a prisoner. Row 3: A refugee, a nurse, a gang member, a lawyer, a security guard, a felon, a volunteer, an architect. Row 4: An elite, a pastor, a philanthropist, a banker, an orphan, a social worker, a doctor, a leader. Row 5: An inmate, a politician, a receptionist, an activist, a software engineer, a drug dealer, a judge, a professor.

Figure 3: (Left) Spearman’s rank correlation coefficient between social identity group and Monk skin tone classification. (Right) Calculation of mean Monk skin-tone classification value for a given group relative to all other portraits. A higher Monk skin tone classification value refers to a darker skin tone.

(a) Key points used for alignment in red; key points from Face Mesh model used to calculate eye center points in blue.

(b) (Left to right) 1) Nose centering validation: allowable distance between eye midpoint and nose tip in green. 2) Extreme tilt screening: maximum vertical distance in green. 3) Side profile screening: minimum width of shorter nose-to-eye distance in green.

Variable	Category	n (%)
Gender	Female	145 (50.2%)
	Male	144 (49.8%)
Race/Ethnicity	White	160 (55.4%)
	Black	40 (13.8%)
	Mixed	34 (11.8%)
	Other	28 (9.7%)

Table 3: Participant Demographics (N = 291)

Variable	Portraits	Table	Test Statistic	p
Gender (F / M)	77 / 71	68 / 73	$\chi^{2}=0.28$	.597
Ethnicity (5 categories)*	13A / 23B / 17M / 14O / 81W	14A / 17B / 17M / 14O / 79W	$\chi^{2}=0.79$	.939
Political affiliation	46D / 60I / 42R	41D / 57I / 43R	$\chi^{2}=0.21$	.902
Age ( $\mu$ )	44.7	45.2	$t=-0.28$	.781

Table 4: Randomization Balance Across Conditions

Class	Text Prompts
Man	“a photo of a man”
	“a portrait of a male person”
	“a picture of a man”
	“a male face”
Woman	“a photo of a woman”
	“a portrait of a female person”

Table 7: Text prompts used for CLIP zero-shot gender classification.

Construct	Pre	Post	$\boldsymbol{\Delta}$	t	p	$\boldsymbol{d_{z}}$
Trust AI outputs	3.64	3.13	$-0.51$	$9.96$	${<}.0001$ **	$-0.59$
Developers build responsibly	3.25	3.14	$-0.11$	$1.76$	.0799	$-0.10$
Feel informed about AI	3.72	3.75	$+0.03$	$-0.49$	.6233	$+0.03$
AI treats groups equally	3.39	2.64	$-0.74$	$10.28$	${<}.0001$ **	$-0.62$
AI has positive impact	3.48	3.30	$-0.18$	$3.89$	<.0001 **	$-0.23$
** $p<0.01$ . Effect sizes are $d_{z}$ (paired).

Table 8: Pre–Post Attitude Shifts (Overall, Paired

t

-Tests)

	Portraits			Table
Construct	$\Delta$	$p$	$d_{z}$	$\Delta$	$p$	$d_{z}$
Trust AI outputs	$-0.42$	${<}.0001$ **	$-0.49$	$-0.60$	${<}.0001$ **	$-0.70$
Developers build responsibly	$-0.08$	.4205	$-0.07$	$-0.15$	.0854	$-0.15$
Feel informed about AI	$+0.13$	.0893	$+0.14$	$-0.08$	.3527	$-0.08$
AI treats groups equally	$-0.64$	${<}.0001$ **	$-0.57$	$-0.85$	${<}.0001$ **	$-0.68$
AI has positive impact	$-0.09$	.1233	$-0.13$	$-0.27$	.0002 **	$-0.33$

Table 9: Pre–Post Attitude Shifts by Condition (Paired

t

-Tests)

Construct	$\boldsymbol{\Delta}$ Portraits	$\boldsymbol{\Delta}$ Table	t	p	d
Trust AI outputs	$-0.42$	$-0.60$	$1.86$	.064	$+0.22$
Developers build responsibly	$-0.08$	$-0.15$	$0.60$	.546	$+0.07$
Feel informed about AI	$+0.13$	$-0.08$	$1.83$	.068	$+0.22$
AI treats groups equally	$-0.64$	$-0.85$	$1.44$	.152	$+0.17$
AI has a positive impact	$-0.09$	$-0.27$	$1.92$	.056	$+0.23$
No comparison reaches significance at $\alpha=.01$ .

Table 10: Between-Condition Comparison of Attitude Shift Magnitudes (Independent-Samples

t

-Tests)

Measure	Overall	Portraits	Table	t	p	d
Comprehension & Concern
Understand patterns	3.67	3.57	3.77	$-1.58$	.114	$-0.19$
Could explain	4.10	4.10	4.10	$-0.04$	.970	$-0.00$
Serious concern	3.78	3.84	3.71	$1.02$	.310	$+0.12$
Real-world harm	3.62	3.62	3.62	$0.01$	.989	$+0.00$
Reflects broader societal issues	4.11	4.06	4.15	$-0.73$	.466	$-0.09$

Table 11: Post-Exposure Comprehension, Concern, and Intent by Condition

Measure	Portraits	Table	t	p	d
Clear communication	3.87	4.11	$-2.13$	.0341	$-0.25$
Hard to understand	2.39	2.57	$-1.18$	.2374	$-0.14$
Effective for public	3.66	3.93	$-2.16$	.0316	$-0.26$
Confidence	3.93	3.99	$-0.57$	.5699	$-0.07$
Task time (sec)	56.25	82.70	$-4.02$	${<}.001$ **	$-0.47$
** $p<.01$ .

Table 12: Format Evaluation, Exposure Time, and Task Confidence by Condition

Identity Dimension	Outcome	$\boldsymbol{F}$	$\boldsymbol{p}$	Simple Effects
Gender (F vs. M)	Skin tone:	9.09	.003	Female: $d=-0.02$ , $p=.901$
	White-collar			Male: $d=+0.73$ , $p<.0001$ **
				Male respondents who received the Portraits stimulus
				perceived white-collar roles as having lighter skin.
Politics (D / R / I)	Serious concern	11.82	${<}.001$	Democrats: $d=-0.21$ , $p=.339$
				Republicans: $d=+0.80$ , $p=.0004$ **

Table 13: Statistically Significant Condition and Identity Interactions (

p<.01

)

Subgroup	Level	Portrait	Table	$t$	$p$	$d$
Serious Concern $\times$ Political Affiliation	Democrat	4.24	4.44	$-0.96$	0.3390	$-0.21$
	Republican	3.90	3.02	3.67	0.0004**	$+0.80$
	Independent	3.53	3.72	$-0.90$	0.3713	$-0.17$
Skin Tone Bias: White-collar $\times$ Sex	Female	1.04	1.06	$-0.12$	0.9012	$-0.02$
	Male	1.31	0.70	4.32	0.0000**	$+0.73$

Table 14: Condition Effects by Subgroup

Figure 10: (Top-down, left to right) Row 1: An immigrant, a capitalist, a convict, a welfare recipient, a business executive, a janitor, a street vendor, an astronaut. Row 2: A deportee, a cab driver, a construction worker, a trust-funder, a homeless person, a farmer, a pharmacist, a prisoner. Row 3: A refugee, a nurse, a gang member, a lawyer, a security guard, a felon, a volunteer, an architect. Row 4: An elite, a pastor, a philanthropist, a banker, an orphan, a social worker, a doctor, a leader. Row 5: An inmate, a politician, a receptionist, an activist, a software engineer, a drug dealer, a judge, a professor.

Prompt	% Male	% Female	Avg Monk Skin Tone
An Activist	56.66	43.34	7.48
An Architect	76.30	23.70	5.76
An Astronaut	85.76	14.24	7.07
A Banker	96.72	3.28	5.84
A Business Executive	71.54	28.46	5.55
A Cab Driver	96.74	3.26	6.63

Table 18

Demographic Bias in T2I Models

Demographic bias in text-to-image models is the unequal or stereotyped representation of people across generated images.

See all 31 papers

Sep 21, 2026

IMPLICIT-Bench: Stereotyped Defaults and Prompt Following in Text-to-Image Generation

Text-to-image (T2I) models often produce stereotypical images when stereotype-relevant attributes are left unspecified. We ask whether explicit requests can overcome these defaults, especially when the requested attribute contradicts the stereotype. We introduce \mbox{IMPLICIT-Bench}, a challenge benchmark of 1,831 prompt triplets built on a bias knowledge graph (KG) of stereotypes recorded in StereoSet and CrowS-Pairs. Each triplet describes one scene with a neutral prompt, a stereotype prompt, and an anti-stereotype prompt. The neutral prompt measures the model's default, and the two explicit prompts measure whether the model can follow a request on either side of the stereotype. On this benchmark, Qwen-Image depicts the stereotype in 52% of neutral images and the anti-stereotype in 5%. On both Qwen-Image and Stable Diffusion~3, explicit requests that contradict this default succeed far less often than requests that agree with it. Where Qwen-Image's default is strongest, explicitly requesting the stereotype succeeds in 97% of cases, whereas requesting the anti-stereotype succeeds in only 54%. In these scenes, 31% of the images for anti-stereotype requests show the stereotypical attribute instead. We further test whether prompt rewriting and embedding steering can reduce the default on neutral prompts. With the prompt's own KG unit, the best configurations improve a combined bias and alignment score by up to 9.5 points, and with a unit retrieved from the bias KG, by at most 2.9. Debiasing thus depends on knowing which stereotype a prompt evokes.

Yue Dai, Ziyang Liu, Marc Cheong +1

The University of Melbourne

T2I Generation Demographic Bias in T2I Models T2I Generation Evaluation

Ask PDF

Apr 14, 2026

T2I-BiasBench: A Multi-Metric Framework for Auditing Demographic and Cultural Bias in Text-to-Image Models

Text-to-image (T2I) generative models achieve impressive visual fidelity but inherit and amplify demographic imbalances and cultural biases embedded in training data. We introduce T2I-BiasBench, a unified evaluation framework of thirteen complementary metrics that jointly captures demographic bias, element omission, and cultural collapse in diffusion models - the first framework to address all three dimensions simultaneously. We evaluate three open-source models - Stable Diffusion v1.5, BK-SDM Base, and Koala Lightning - against Gemini 2.5 Flash (RLHF-aligned) as a reference baseline. The benchmark comprises 1,574 generated images across five structured prompt categories. T2I-BiasBench integrates six established metrics with seven additional measures: four newly proposed (Composite Bias Score, Grounded Missing Rate, Implicit Element Missing Rate, Cultural Accuracy Ratio) and three adapted (Hallucination Score, Vendi Score, CLIP Proxy Score). Three key findings emerge: (1) Stable Diffusion v1.5 and BK-SDM exhibit bias amplification (>1.0) in beauty-related prompts; (2) contextual constraints such as surgical PPE substantially attenuate professional-role gender bias (Doctor CBS = 0.06 for SD v1.5); and (3) all models, including RLHF-aligned Gemini, collapse to a narrow set of cultural representations (CAS: 0.54-1.00), confirming that alignment techniques do not resolve cultural coverage gaps. T2I-BiasBench is publicly released to support standardized, fine-grained bias evaluation of generative models. The project page is available at: https://gyanendrachaubey.github.io/T2I-BiasBench/

Nihal Jaiswal, Siddhartha Arjaria, Gyanendra Chaubey +3

Department of Information Technology, Rajkiya Engineering College Banda, 210201, Uttar Pradesh, India. · School of AI and Data Science, Indian Institute of Technology Jodhpur, 342030, Rajasthan, India.

Demographic Bias in T2I Models T2I Generation Evaluation T2I Diffusion Models

Ask PDF

Aug 3, 2026

Investigating Social Bias in Narrative Image Generation

Text-to-image (T2I) generation models are increasingly embedded in applications such as media content creation and education, raising concerns about how their outputs may reproduce social biases. Prior work has shown that T2I models exhibit social biases, yet existing evaluations largely focus on a photo generation task. As a result, it remains unclear whether and how such biases manifest in more narrative visual formats, such as storyboards and comics, where characters and events are presented across multiple panels. In this work, we compare bias expression across photo, storyboard, and comic generation in six T2I models by adapting BBG, a text-based bias evaluation framework, to image generation. Our results show that proprietary models generate 25.9% biased outputs in photo generation on average, with biased outputs increasing by 9.6pp in storyboard generation and 18.2pp in comic generation. We also find that photos mainly encode biases through subtle visual cues, while storyboards and comics reveal them more explicitly through event sequencing, character positioning, narrative resolution, and textual elements. These findings show that biases that remain less visible in photo generation may surface in narrative visual formats, highlighting the importance of evaluating T2I systems with diverse visual formats beyond photo generation.

Junyeong Park, Sowon Min, Euna Jang +5

KAIST, South Korea

T2I Generation Demographic Bias in T2I Models T2I Generation Evaluation

Ask PDF

Abstract

Demographic Bias in T2I Models

Explore similar work

IMPLICIT-Bench: Stereotyped Defaults and Prompt Following in Text-to-Image Generation

T2I-BiasBench: A Multi-Metric Framework for Auditing Demographic and Cultural Bias in Text-to-Image Models

Investigating Social Bias in Narrative Image Generation