Pith. sign in

REVIEW 2 major objections 2 minor 75 references

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

T0 review · 2 major / 2 minor · reviewed 2026-05-10 · grok-4.3

Challenge this review Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A pipeline that generates many portraits from a prompt, aligns them, and averages their pixels creates single images that let non-experts see biases in text-to-image models as clearly as data tables do.

desk verdict GLEaN gives a clean visual shortcut for showing T2I biases to non-experts, but the user-study evidence is still thin on methods and the median composites may not be as faithful as claimed. read the letter →

arxiv 2604.09923 v1 submitted 2026-04-10 cs.AI cs.CV

Bochu Ding , Brinnae Bent , Augustus Wendell This is my paper · ORCID
classification cs.AIcs.CV
keywords text-to-imagebiasdetectionexplainabilitygenerativeAIuserstudymediancompositionfaciallandmarkspublicunderstanding
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

The authors present GLEaN, a three-stage process for turning text-to-image model outputs into visual bias summaries. Large sets of images are created from prompts involving social or occupational identities, faces are aligned using landmarks, and a median composition is formed to represent the typical output. This visual format requires no statistical training to interpret, allowing viewers to directly observe associations the model makes. A controlled study with hundreds of participants confirmed that these portraits communicate bias information with the same accuracy as numerical tables while requiring less time to review. The technique depends only on generated images, enabling its use with any text-to-image system regardless of whether its internal details are accessible.

What carries the argument

The median-pixel composition step, which follows landmark-based filtering and spatial alignment of large-scale generated faces to create a single representative portrait of the model's typical output for a given prompt.

What would settle it

A direct comparison where the GLEaN composite differs substantially from the most frequent visual features observed across the full set of generated images, or where participants detect biases less accurately with the portraits than with equivalent data tables.

Watch

Extended reading notes

Core claim

GLEaN distills the central tendency of a text-to-image model's responses to identity prompts into median-pixel portraits by generating images at scale, applying facial landmark filtering for alignment, and composing the median across pixels. These portraits reproduce established biases and highlight additional patterns, such as links between skin tone and expressed emotion, and a user study demonstrates that they convey bias information to the public as effectively as conventional tables but in significantly shorter viewing times.

Load-bearing premise

The median pixel values after landmark alignment accurately capture the model's intended central representation rather than being altered by generation noise, prompt wording, or the specific alignment rules chosen.

Editorial extensions

If this is right

  • Biases in model outputs become immediately visible to people without technical expertise.
  • The method can be applied to proprietary or black-box models using only their generated images.
  • New bias patterns can be discovered through visual inspection of the composites.
  • Communication of model behaviors to the public can be streamlined compared to tabular data presentations.

Reading between the lines

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

  • Extending the alignment and composition process to other image domains, such as objects or scenes, could broaden bias auditing beyond faces.
  • Integration into public dashboards might allow ongoing monitoring of model updates for bias shifts.
  • Comparing composites across different models could provide a standardized visual benchmark for bias levels.
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, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces GLEaN, a three-stage pipeline (large-scale generation from identity prompts, facial landmark filtering/alignment, and median-pixel composition) to produce representative portraits that visualize biases encoded in text-to-image models for non-technical audiences. Demonstrated on Stable Diffusion XL across 40 social/occupational prompts, the composites reproduce documented biases and surface new skin-tone/emotion associations. A between-subjects user study (N=291) claims the portraits communicate biases as effectively as data tables while requiring significantly less viewing time. The method is presented as model-agnostic and replicable on black-box systems.

Significance. If the user-study results and the fidelity of the median composites hold, GLEaN would provide a scalable, accessible tool for public comprehension of T2I biases, filling a gap between technical auditing methods and lay audiences. Strengths include its black-box applicability, open-source release, and potential to support broader discourse on AI fairness without requiring statistical expertise.

major comments (2)
  1. [Abstract / User Study] Abstract and User Study section: The central claim that GLEaN portraits 'communicate biases as effectively as conventional data tables' with significantly less viewing time rests on a between-subjects study (N=291) but reports no details on design, controls, randomization, statistical tests, effect sizes, or power analysis. This omission prevents evaluation of whether the equivalence conclusion is supported.
  2. [§3 (GLEaN Pipeline)] §3 (GLEaN Pipeline, median-pixel composition step): The assumption that the final median-pixel portrait after landmark filtering and alignment accurately represents the model's central tendency is load-bearing for all downstream claims, yet no validation is provided against the full distribution of generated images. Common T2I artifacts (warped faces, inconsistent lighting) or prompt sensitivity could systematically distort the composite away from the true bias distribution.
minor comments (2)
  1. [Discussion / Limitations] The manuscript would benefit from explicit discussion of limitations, including sensitivity of results to the specific landmark detector, alignment method, and number of samples used in the median computation.
  2. [Figures / Code Availability] Figure captions and the GitHub link should include version numbers or commit hashes to support reproducibility of the exact pipeline used for the reported composites.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment below and have revised the manuscript to incorporate additional details and validation as requested.

read point-by-point responses

  1. Referee: [Abstract / User Study] Abstract and User Study section: The central claim that GLEaN portraits 'communicate biases as effectively as conventional data tables' with significantly less viewing time rests on a between-subjects study (N=291) but reports no details on design, controls, randomization, statistical tests, effect sizes, or power analysis. This omission prevents evaluation of whether the equivalence conclusion is supported.

    Authors: We agree that the original manuscript did not provide sufficient detail on the user study to allow full evaluation of the claims. In the revised version, we have expanded the User Study section to include a complete account of the between-subjects design, randomization procedures for condition assignment and prompt order, controls for attention and demographics, the statistical tests employed (independent-samples t-tests for viewing time and two one-sided tests for equivalence on comprehension accuracy), effect sizes, and a post-hoc power analysis. We also report participant exclusion criteria and make the anonymized data and analysis scripts available. These additions directly support evaluation of the reported findings. revision: yes

  2. Referee: [§3 (GLEaN Pipeline)] §3 (GLEaN Pipeline, median-pixel composition step): The assumption that the final median-pixel portrait after landmark filtering and alignment accurately represents the model's central tendency is load-bearing for all downstream claims, yet no validation is provided against the full distribution of generated images. Common T2I artifacts (warped faces, inconsistent lighting) or prompt sensitivity could systematically distort the composite away from the true bias distribution.

    Authors: The referee correctly identifies that explicit validation of the median composite against the full image distribution was absent. While the median pixel operation is robust by construction, we have added a new validation subsection in §3 that compares each median portrait to the corresponding mean image and to a random sample of 50 individual generations using pixel-wise variance maps and perceptual similarity metrics (SSIM and LPIPS). The results show that landmark filtering substantially reduces artifact-induced variance and that the median better approximates the central tendency than the mean. We also discuss remaining limitations due to prompt sensitivity. This revision provides the requested empirical grounding. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: procedural pipeline validated by external user study

full rationale

The paper describes a three-stage procedural pipeline (large-scale generation from identity prompts, landmark-based filtering/alignment, median-pixel composition) to produce representative portraits, then validates communication effectiveness via an independent between-subjects user study (N=291) comparing GLEaN portraits to data tables. No mathematical derivations, equations, fitted parameters, or self-citations appear in the load-bearing claims; the central result is an empirical finding about viewer time and comprehension rather than any reduction of outputs to inputs by construction. The method is explicitly model-agnostic and black-box compatible, with no self-referential definitions or uniqueness theorems invoked.

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

Abstract-only review reveals no explicit free parameters, unstated mathematical axioms, or newly invented entities; the contribution is a practical image-processing pipeline relying on standard generative models and facial analysis tools.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension." pith.science (2026). https://pith.science/paper/2604.09923

@misc{pith2026260409923,
  author       = {Pith},
  title        = {Pith review of: GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2604.09923}},
  note         = {Machine review of arXiv:2604.09923}
}
read the original 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

Figures reproduced from arXiv: 2604.09923 by the authors.

Figure 1
Figure 1. Workflow overview of GLEaN, with example outputs for the “business executive” prompt. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
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 … view at source ↗
Figure 3
Figure 3. (Left) Spearman’s rank correlation coefficient between social identity group and Monk skin tone classification. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Methods used in portrait filtering and alignment. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: (Left) A selection of images generated from SDXL from "a business executive" prompt. (Right) Post-processed [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The green mask visualizes the area of the face used to approximate the Monk scale skin-tone category. 2 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Correlation between Monk skin-tone classification and probability of angry being the prediction dominant emotion. 2.5. LAION-2B extraction [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Results of a query for "nurse" from the LAION-2B dataset using similar filters deployed to train SDXL such as text faithfulness, high aesthetic scores, and low watermark and NSFW probability. 4 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: * represents statistical significance of p < 0.01 for a one-sample t-test against µ = 0. 7 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. 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 pharmaci…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 75 canonical work pages

  1. [1]

    Abuzuraiq and Philippe Pasquier

    Ahmed M. Abuzuraiq and Philippe Pasquier. Explainability-in-action: Enabling expressive manip- ulation and tacit understanding by bending diffusion models in ComfyUI, 2025. 1, 4

    work page 2025

  2. [2]

    Adobe firefly: The next evolution of creative AI is here | adobe blog, 2026

    Adobe Firefly Team. Adobe firefly: The next evolution of creative AI is here | adobe blog, 2026. 1

    work page 2026

  3. [3]

    AI- generated faces influence gender stereotypes and racial homogenization

    Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. AI- generated faces influence gender stereotypes and racial homogenization. 15(1):14449, 2025. 2

    work page 2025

  4. [4]

    Do mass mediated stereotypes harm members of negatively stereotyped groups? a meta-analytical review on media-generated stereotype threat and stereotype lift

    Markus Appel and Silvana Weber. Do mass mediated stereotypes harm members of negatively stereotyped groups? a meta-analytical review on media-generated stereotype threat and stereotype lift. 48(2):151–179,

    work page

  5. [5]

    Cultivating intergroup emotions: An intergroup threat theory ap- proach

    Anita Atwell Seate, Rong Ma, Hsin-Yi Chien, and Dana Mastro. Cultivating intergroup emotions: An intergroup threat theory ap- proach. 21(2):178–197, 2018. _eprint: https://doi.org/10.1080/15205436.2017.1381262. 1

    work page doi:10.1080/15205436.2017.1381262 2018

  6. [6]

    Neutral prompts, non-neutral people: Quantifying gender and skin-tone bias in gemini flash 2.5 image and GPT image 1.5, 2026

    Roberto Balestri. Neutral prompts, non-neutral people: Quantifying gender and skin-tone bias in gemini flash 2.5 image and GPT image 1.5, 2026. version: 1. 1, 2

    work page 2026

  7. [7]

    How well can text-to-image generative models understand ethical natural language interven- tions?, 2022

    Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interven- tions?, 2022. 1, 3

    work page 2022

  8. [8]

    Easily accessible text-to-image generation amplifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tat- sunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493–1504....

    work page 2023

Show all 75 references
  1. [9]

    Bond and Brandon Miller

    Bradley J. Bond and Brandon Miller. From screen to self: The relationship between television exposure and self-complexity among lesbian, gay, and bisexual youth. 11:19–19, 2017. 1

    2017

  2. [10]

    Ross, Roberto Cuomo, and Warren M

    Jevan Cevik, Bryan Lim, Ishith Seth, Foti Sofiadellis, Richard J. Ross, Roberto Cuomo, and Warren M. Rozen. Assessment of the bias of artificial intelligence generated images and large language models on their depiction of a surgeon. 94(3):287–294, 2024. _eprint: https://onlin...

    2024 doi

  3. [11]

    Identifying race and gender bias in stable dif- fusion AI image generation

    Aadi Chauhan, Taran Anand, Tanisha Jauhari, Arjav Shah, Rudransh Singh, Arjun Rajaram, and Rithvik Vanga. Identifying race and gender bias in stable dif- fusion AI image generation. In2024 IEEE 3rd Inter- national Conference on AI in Cybersecurity (ICAIC), pages 1–6, 2024. 1, 2, 3

    2024

  4. [12]

    TIBET: Identifying and evaluating biases in text-to-image generative models, 2024

    Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. TIBET: Identifying and evaluating biases in text-to-image generative models, 2024. version: 2. 1, 3

    2024

  5. [13]

    DALL- eval: Probing the reasoning skills and social biases of text-to-image generation models, 2023

    Jaemin Cho, Abhay Zala, and Mohit Bansal. DALL- eval: Probing the reasoning skills and social biases of text-to-image generation models, 2023. 1, 2

    2023

  6. [14]

    OpenBias: Open-set bias detection in text-to-image generative models, 2024

    Moreno D’Incà, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. OpenBias: Open-set bias detection in text-to-image generative models, 2024. 1, 3

    2024

  7. [15]

    Dixon and Cristina L Azocar

    Travis L. Dixon and Cristina L Azocar. Priming crime and activating blackness: Understanding the psycho- logical impact of the overrepresentation of blacks as lawbreakers on television news. 57(2):229–253, 2007. 1

    2007

  8. [16]

    Interactive dis- covery and exploration of visual bias in generative text-to-image models, 2025

    Johannes Eschner, Roberto Labadie-Tamayo, Matthias Zeppelzauer, and Manuela Waldner. Interactive dis- covery and exploration of visual bias in generative text-to-image models, 2025. version: 1. 3

    2025

  9. [17]

    ‘person’ == light- skinned, western man, and sexualization of women of color: Stereotypes in stable diffusion

    Sourojit Ghosh and Aylin Caliskan. ‘person’ == light- skinned, western man, and sexualization of women of color: Stereotypes in stable diffusion. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 6971–6985. Association for Computational Linguistics...

    2023

  10. [18]

    A large scale analysis of gender biases in text-to-image generative models, 2025

    Leander Girrbach, Stephan Alaniz, Genevieve Smith, and Zeynep Akata. A large scale analysis of gender biases in text-to-image generative models, 2025. version:

    2025

  11. [19]

    Peter Glick and Susan T. Fiske. Hostile and benevolent sexism: Measuring ambivalent sexist attitudes toward women. 21(1):119–135, 1997. 7

    1997

  12. [20]

    Face landmark detection guide

    Google AI Edge. Face landmark detection guide. 4

  13. [21]

    Iden- tifying implicit social biases in vision-language models,

    Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, and Marzyeh Ghassemi. Iden- tifying implicit social biases in vision-language models,

  14. [22]

    Social perception of faces in a vision- language model

    Carina I Hausladen, Manuel Knott, Colin F Camerer, and Pietro Perona. Social perception of faces in a vision- language model. InProceedings of the 2025 ACM Con- ference on Fairness, Accountability, and Transparency, pages 639–659. Association for Computing Machinery,

    2025

  15. [23]

    OpenFace 3.0: A lightweight multi- task system for comprehensive facial behavior analysis,

    Jiewen Hu, Leena Mathur, Paul Pu Liang, and Louis- Philippe Morency. OpenFace 3.0: A lightweight multi- task system for comprehensive facial behavior analysis,

  16. [24]

    ComfyGPT: A self-optimizing multi- agent system for comprehensive ComfyUI workflow generation, 2025

    Oucheng Huang, Yuhang Ma, Zeng Zhao, Mingrui Wu, Jiayi Ji, Rongsheng Zhang, Zhipeng Hu, Xiaoshuai Sun, and Rongrong Ji. ComfyGPT: A self-optimizing multi- agent system for comprehensive ComfyUI workflow generation, 2025. 4 9

    2025

  17. [25]

    Bodenhausen

    Kurt Hugenberg and Galen V. Bodenhausen. Facing prejudice: Implicit prejudice and the perception of facial threat. 14(6):640–643, 2003. 7

    2003

  18. [26]

    Reddy, and Sunipa Dev

    Akshita Jha, Vinodkumar Prabhakaran, Remi Denton, Sarah Laszlo, Shachi Dave, Rida Qadri, Chandan K. Reddy, and Sunipa Dev. ViSAGe: A global-scale anal- ysis of visual stereotypes in text-to-image generation. InProceedings of the 62nd Annual Meeting of the Asso- ciation for Com...

    2024

  19. [27]

    Stable bias: Analyzing societal representations in diffusion models, 2023

    Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing societal representations in diffusion models, 2023. 1, 3

    2023

  20. [28]

    Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A

    Vongani H. Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A. Efros, Angjoo Kanazawa, and Devin Guillory. Studying bias in GANs through the lens of race, 2022. 3

    2022

  21. [29]

    Nila Masrourisaadat, Nazanin Sedaghatkish, Fatemeh Sarshartehrani, and Edward A. Fox. Analyzing qual- ity, bias, and performance in text-to-image generative models, 2024. version: 1. 2

    2024

  22. [30]

    Dana E. Mastro. A social identity ap- proach to understanding the impact of televi- sion messages. 70(2):98–113, 2003. _eprint: https://doi.org/10.1080/0363775032000133764. 1

    2003 doi

  23. [31]

    The monk skin tone scale, 2023

    Ellis Monk. The monk skin tone scale, 2023. 4, 6

    2023

  24. [32]

    Uncovering bias in face generation models, 2023

    Cristian Muñoz, Sara Zannone, Umar Mohammed, and Adriano Koshiyama. Uncovering bias in face generation models, 2023. 3

    2023

  25. [33]

    Social biases through the text-to-image generation lens, 2023

    Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens, 2023. 1, 2, 3

    2023

  26. [34]

    Face the facts: Using face averaging to visualize gender-by-race bias in facial analysis algo- rithms

    Kentrell Owens, Erin Freiburger, Ryan Hutchings, Mat- tea Sim, Kurt Hugenberg, Franziska Roesner, and Ta- dayoshi Kohno. Face the facts: Using face averaging to visualize gender-by-race bias in facial analysis algo- rithms. InProceedings of the 2024 AAAI/ACM Con- ference on AI...

    2024

  27. [35]

    Pasquini, Brian Kennedy, Jeffrey Gottfried, Monica Anderson, and Colleen McClain

    C. Pasquini, Brian Kennedy, Jeffrey Gottfried, Monica Anderson, and Colleen McClain. How the u.s. public and AI experts view artificial intelligence, 2025. 1

    2025

  28. [36]

    SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis, 2023. 3

    2023

  29. [37]

    Breaking language barriers or reinforcing bias? a study of gender and racial disparities in multilingual con- trastive vision language models, 2025

    Zahraa Al Sahili, Ioannis Patras, and Matthew Purver. Breaking language barriers or reinforcing bias? a study of gender and racial disparities in multilingual con- trastive vision language models, 2025. version: 4. 2

    2025

  30. [38]

    Rains, Dana Mastro, and Stewart Coles

    Muniba Saleem, Ian Hawkins, Stephen A. Rains, Dana Mastro, and Stewart Coles. Media depictions of minority groups: A meta-analytic review exam- ining the effects of positive and negative portrayals on outgroup evaluations. 0(0):1–31, 2025. _eprint: https://doi.org/10.1080/1521...

    2025 doi

  31. [39]

    DiffusionWorldViewer: Exposing and broadening the worldview reflected by generative text-to-image models, 2024

    Zoe De Simone, Angie Boggust, Arvind Satyanarayan, and Ashia Wilson. DiffusionWorldViewer: Exposing and broadening the worldview reflected by generative text-to-image models, 2024. 7

    2024

  32. [40]

    ComfySearch: Autonomous exploration and reasoning for ComfyUI workflows, 2026

    Jinwei Su, Qizhen Lan, Zeyu Wang, Yinghui Xia, Hairu Wen, Yiqun Duan, Xi Xiao, Tianyu Shi, Yang Jingsong, and Lewei He. ComfySearch: Autonomous exploration and reasoning for ComfyUI workflows, 2026. 4

    2026

  33. [41]

    The effect of prime time television ethnic/racial stereo- types on latino and black americans: A longitudinal national level study

    Riva Tukachinsky, Dana Mastro, and Moran Yarchi. The effect of prime time television ethnic/racial stereo- types on latino and black americans: A longitudinal national level study. 61(3):538–556, 2017. _eprint: https://doi.org/10.1080/08838151.2017.1344669. 1

    2017 doi

  34. [42]

    Quantifying bias in text-to-image generative models, 2023

    Jordan Vice, Naveed Akhtar, Richard Hartley, and Aj- mal Mian. Quantifying bias in text-to-image generative models, 2023. 3

    2023

  35. [43]

    Garrido-Merchán

    Adriana Fernández de Caleya Vázquez and Eduardo C. Garrido-Merchán. A taxonomy of the biases of the images created by generative artificial intelligence, 2024. version: 1. 2

    2024

  36. [44]

    The male CEO and the female assistant: Evaluation and mitigation of gender biasesintext-to-imagegenerationofdualsubjects, 2024

    Yixin Wan and Kai-Wei Chang. The male CEO and the female assistant: Evaluation and mitigation of gender biasesintext-to-imagegenerationofdualsubjects, 2024. 1, 2, 3

    2024

  37. [45]

    Survey of bias in text-to-image generation: Definition, evaluation, and mitigation, 2024

    Yixin Wan, Arjun Subramonian, Anaelia Ovalle, Zongyu Lin, Ashima Suvarna, Christina Chance, Hri- tik Bansal, Rebecca Pattichis, and Kai-Wei Chang. Survey of bias in text-to-image generation: Definition, evaluation, and mitigation, 2024. version: 3. 2

    2024

  38. [46]

    T2iat: Measuring valence and stereotypical biases in text-to-image generation, 2023

    Jialu Wang, Xinyue Gabby Liu, Zonglin Di, Yang Liu, and Xin Eric Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation, 2023. 3

    2023

  39. [47]

    New job, new gender? measuring the social bias in image generation models

    Wenxuan Wang, Haonan Bai, Jen-tse Huang, Yuxuan Wan, Youliang Yuan, Haoyi Qiu, Nanyun Peng, and Michael Lyu. New job, new gender? measuring the social bias in image generation models. InProceedings of the 32nd ACM International Conference on Multi- media, pages 3781–3789. Asso...

    2024

  40. [48]

    Happy young women, grumpy old men? emotion-driven demographic biases in synthetic face generation, 2026

    Mengting Wei, Aditya Gulati, Guoying Zhao, and Nuria Oliver. Happy young women, grumpy old men? emotion-driven demographic biases in synthetic face generation, 2026. 7

    2026

  41. [49]

    Understanding the im- pact of AI generated content on social media: The pixiv case, 2024

    Yiluo Wei and Gareth Tyson. Understanding the im- pact of AI generated content on social media: The pixiv case, 2024. 1

    2024

  42. [50]

    Exposure to negative stereotypes influences representa- tions of monetary incentives in the nucleus accumbens

    B Locke Welborn, Youngki Hong, and Kyle G Ratner. Exposure to negative stereotypes influences representa- tions of monetary incentives in the nucleus accumbens. 15(3):347–358, 2020. 1

    2020

  43. [51]

    Bias amplification in stable diffusion’s representation of stigma through skin tones and their homogeneity,

    Kyra Wilson, Sourojit Ghosh, and Aylin Caliskan. Bias amplification in stable diffusion’s representation of stigma through skin tones and their homogeneity,

  44. [52]

    ComfyUI-r1: Exploring reasoning models for workflow generation, 2025

    Zhenran Xu, Yiyu Wang, Xue Yang, Longyue Wang, Weihua Luo, Kaifu Zhang, Baotian Hu, and Min Zhang. ComfyUI-r1: Exploring reasoning models for workflow generation, 2025. 4 10

    2025

  45. [53]

    ComfyBench: Benchmark- ing LLM-based agents in ComfyUI for autonomously designing collaborative AI systems, 2026

    Xiangyuan Xue, Zeyu Lu, Di Huang, Zidong Wang, Wanli Ouyang, and Lei Bai. ComfyBench: Benchmark- ing LLM-based agents in ComfyUI for autonomously designing collaborative AI systems, 2026. 4

    2026

  46. [54]

    ComfyUI load prompts from text file workflow, 2024

    Andrés Zsögön. ComfyUI load prompts from text file workflow, 2024. Section: ComfyUI. 4 11 GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension Supplementary Material

    2024

  47. [55]

    a business executive

    GLEaN Methodology 1.1. Portrait filtering and alignment (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 i...

  48. [56]

    charitable

    Empirical Evaluation 2.1. Prompt classification by social identity group Table 1.Social identity classification Identity Included Prompts White-collar An architect, an astronaut, a banker, a business executive, a capitalist, a doctor, an elite, a judge, a lawyer, a leader, a n...

  49. [57]

    a photo of a man

    User research 3.1. Overview This user study investigates whether the GLEaN method proposed in the main text renders text-to-image (T2I) model biases more legible to the public. This supplement presents the full methodology, results, and analysis of the conducted study. (Please...

  50. [58]

    Overall, how confident are you in your assessment?

    (see section3.5 Empirical Evaluationin the main text) for each image contributing to a given prompt’s median portrait. We then average this value across the entire set of images. 3.4. Measures The survey comprised five measurement blocks administered in fixed order. (For the f...

  51. [59]

    Compensation Participants were paid $1.00 USD for an approximately 5-minute survey

    Full Survey Text 4.1. Compensation Participants were paid $1.00 USD for an approximately 5-minute survey. 4.2. Disclosure and Consent Study Title:Evaluation of AI-generated Images Principal Researcher:Bochu Ding (bochu.ding@duke.edu), Master’s of Engineering Candidate, Duke Un...

    2026

  52. [60]

    I generally trust the outputs of AI systems

  53. [61]

    AI companies generally develop their technology responsibly

  54. [62]

    I feel informed about how AI technologies work

  55. [63]

    AI systems treat all groups of people equally

  56. [64]

    a portrait of person

    Overall, I think AI will have a positive impact on society. •Strongly agree •Somewhat agree •Neither agree nor disagree •Somewhat disagree •Strongly disagree •Unsure 4.4. Introduction Instructions:Text-to-image AI models produce images based on text prompts. For example, this ...

  57. [65]

    What, if anything, did you notice from the information presented? 13

  58. [66]

    construction workers, security guards) (b) White-collar (e.g

    Based on your impression of the information presented, how would you describe AI depiction of the following groups, generally speaking: (a) Blue-collar (e.g. construction workers, security guards) (b) White-collar (e.g. business executive, lawyers) (c) Criminal-related (e.g. f...

  59. [67]

    Comprehension, Action, Evaluation

    Overall, how confident are you in your assessment? •Not confident at all •Not very confident •Neither confident nor unconfident •Somewhat confident •Very confident •Unsure 4.7. Comprehension, Action, Evaluation

  60. [68]

    (a) I understand the patterns in the AI’s outputs

    Please indicate the extent to which you agree with the following statements. (a) I understand the patterns in the AI’s outputs. (b) I could explain what I saw to someone else. (c) The patterns in these outputs are a serious concern. (d) The patterns could cause real-world harm...

  61. [69]

    (a) I would share what I saw with others

    Please indicate the extent to which you agree with the following statements. (a) I would share what I saw with others. (b) I would support policies requiring AI companies to audit their outputs. (c) Seeing this changed how I use or think about AI image tools. •Strongly agree •...

  62. [70]

    14 (a) The information presented clearly communicated the patterns in the AI’s outputs

    Please indicate the extent to which you agree with the following statements. 14 (a) The information presented clearly communicated the patterns in the AI’s outputs. (b) I had to work hard to make sense of what I was shown. (c) This question an attention check: please click str...

  63. [71]

    I feel like I have a good understanding of how AI technologies work

  64. [72]

    AI companies take adequate steps to develop their products

  65. [73]

    I am optimistic about the role AI will play in society

  66. [74]

    AI systems produce trustworthy outputs

  67. [75]

    •Strongly agree •Somewhat agree •Neither agree nor disagree •Somewhat disagree •Strongly disagree •Unsure 15

    AI technologies represent all groups of people equally. •Strongly agree •Somewhat agree •Neither agree nor disagree •Somewhat disagree •Strongly disagree •Unsure 15

Pith tools

Reviewed May 10, 2026 · model on record in the stance chip above.