pith. sign in

arxiv: 2606.05829 · v1 · pith:LHN3FNALnew · submitted 2026-06-04 · 💻 cs.CV

Gender Artifacts from Art History to Text-to-Image Generation

Piera Riccio , Miriam Doh , Benedikt H\"oltgen , Noa Garcia , Nanne van Noord This is my paper

Pith reviewed 2026-06-28 02:14 UTC · model grok-4.3

classification 💻 cs.CV
keywords gender artifactsartistic stylestext-to-image generationStyleGender datasetSGA metricsart historybias amplificationvisual features
0
0 comments X

The pith

Text-to-image models amplify gender patterns embedded in historical artistic styles.

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

Artistic styles carry historical constructions of gender that shape visual features such as composition and pixel-level signals. The paper builds the StyleGender dataset of 74k images across 19 styles, pairing historical artworks with images generated from matching style and gender prompts, plus a semantically aligned comparison set. Two new Set Gender Artifact metrics track these signals at the pixel level and in object masks. Measurements show that gender representation already varies systematically with style in the historical collection, that style keywords transfer those patterns into generated outputs, and that the generative step increases the strength of the signals relative to the source images. A sympathetic reader would care because the work isolates style as a carrier of social information rather than a neutral visual filter.

Core claim

The central claim is that gender representation shapes visual features across artistic styles, that these patterns are carried into text-to-image outputs by style keywords, and that generative models amplify the resulting gender artifacts beyond the levels observed in the historical sources.

What carries the argument

The StyleGender dataset together with the PixelSGA and MaskSGA metrics, which quantify gender signals respectively at the pixel level and through compositional masks while holding style and gender prompts constant.

If this is right

  • Gender representation already influences measurable visual features inside each artistic style in the historical record.
  • Style keywords alone are sufficient to reproduce and transmit those gender-linked visual patterns into generated images.
  • The generative process increases the strength of gender artifacts relative to the historical baseline under the same style and gender specifications.
  • Direct art-history-to-generation comparisons become possible once a semantically aligned image set is constructed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the amplification holds, then interventions that target only content prompts may leave residual bias carried by style descriptors.
  • The same measurement approach could be applied to other social categories such as age or ethnicity to test whether amplification is specific to gender.
  • Repeated use of historical style prompts in training or inference may create a feedback loop that further entrenches the amplified patterns.
  • The dataset enables future work to test whether removing or reweighting particular style keywords reduces the observed amplification.

Load-bearing premise

The controlled style and gender prompts produce generated images whose only systematic difference from the historical set is the generative process itself.

What would settle it

A controlled experiment in which the same prompt set yields generated images whose measured gender signals match or fall below the historical baseline after accounting for any measurable differences in content distribution or annotation quality.

Figures

Figures reproduced from arXiv: 2606.05829 by Benedikt H\"oltgen, Miriam Doh, Nanne van Noord, Noa Garcia, Piera Riccio.

Figure 1
Figure 1. Figure 1: Sample Neoclassicism artworks, illustrating systematic visual differences in setting, color [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of male and female artworks across the 19 artistic styles in our curated sample [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample images generated with Stable Diffusion 3.5 Medium (SD) and Flux.1-dev (Flux) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The “Immaculate Conception” by Diego Velázquez is first captioned by Qwen2-VL-2B [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: On the left-hand side, "The broken column" by Frida Khalo. In the center, exemplary ma [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: PIXELSGA and MASKSGA scores for historic paintings of different art styles. Symbolism (bottom left) shows the least encoding of gender, while Neoclassicism (top right) shows the most. Right: Scores for images generated with SD and Flux for the same styles. All sets of images show moderate correlation between PIXELSGA and MASKSGA. Because kNN on CLIP embeddings may be unstable under image transformati… view at source ↗
Figure 7
Figure 7. Figure 7: PIXELSGA and MASKSGA scores on generated images based on prompts that are semantically aligned with our art history dataset, to enable a cross-domain comparison [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Robustness analysis of our metrics (PixelSGA and MaskSGA) to the choice of embedding [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of gender correction applied to Qwen2-VL-generated captions. For each painting, [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of manual subject correction. The raw Qwen2-VL caption (top) describes the [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative examples of the captioning and generation pipeline. Each row shows a [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Classifier performance on Art Nouveau art historical images at different resolutions and [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Classifier performance on Orientalism art historical images at different resolutions and [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: SGA scores per art style for historic paintings (x-axis) and images generated with [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Robustness analysis of our metrics (PixelSGA and MaskSGA) to the choice of embedding [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
read the original abstract

Artistic styles are rooted in specific socio-historical contexts that encode social hierarchies, including distinct constructions of gender. Yet in AI research, style has long been treated as a surface-level visual property: a filter of color, brushstroke, and texture applied to otherwise content-neutral scenes. We introduce the first dataset to investigate the interplay between gender representation and style in both historical and generated images. StyleGender comprises 74k images spanning 19 artistic styles, comprising art historical images with style and gender annotations, T2I-generated images under controlled style and gender prompts, and a semantically aligned set enabling direct art history-to-generation comparison. By proposing two Set Gender Artifact (SGA) metrics (PixelSGA and MaskSGA), capturing gender signals at the pixel level and in compositional structure, we show that (1) gender representation shapes visual features across artistic styles, (2) style keywords carry these patterns into T2I generation, and (3) generative models tend to amplify gender artifacts beyond what is observed in historical sources.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces the StyleGender dataset of 74k images across 19 artistic styles, consisting of annotated historical art images, T2I-generated images under controlled style+gender prompts, and a semantically aligned set for direct comparison. It defines PixelSGA and MaskSGA metrics to capture gender signals at pixel and compositional levels, and reports three findings: gender representation shapes visual features across styles, style keywords transmit these patterns into T2I outputs, and generative models amplify gender artifacts beyond historical baselines.

Significance. If the central comparison holds, the work supplies a large-scale resource and two quantitative metrics for measuring how socio-historical gender patterns in art are reproduced and intensified by text-to-image models. The dataset construction and the explicit historical-to-generated alignment are strengths that could support follow-on studies on bias propagation in generative vision systems.

major comments (2)
  1. [Dataset construction and abstract claim (3)] Abstract and § on dataset construction: the amplification claim (generated images exceed historical gender signals on PixelSGA/MaskSGA) requires that the only systematic difference between the historical set and the T2I set is the generative process. The semantically aligned set is built via prompt-based matching, but the manuscript reports no verification of content overlap (object categories, spatial layout, or embedding distances). Uncontrolled differences in how models render scenes or add elements could inflate SGA scores independently of amplification.
  2. [PixelSGA and MaskSGA definitions] Metrics section: no inter-annotator agreement, label validation, or statistical controls for prompt content are described for the gender annotations underlying PixelSGA and MaskSGA. Without these, the empirical comparisons rest on unexamined measurement assumptions and cannot reliably support the amplification result.
minor comments (1)
  1. [Abstract] The abstract states 74k images but does not break down the split between historical, generated, and aligned subsets; a table with exact counts per style would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback. We address each major comment below with clarifications on our methodology and indicate planned revisions where appropriate.

read point-by-point responses

  1. Referee: Abstract and § on dataset construction: the amplification claim (generated images exceed historical gender signals on PixelSGA/MaskSGA) requires that the only systematic difference between the historical set and the T2I set is the generative process. The semantically aligned set is built via prompt-based matching, but the manuscript reports no verification of content overlap (object categories, spatial layout, or embedding distances). Uncontrolled differences in how models render scenes or add elements could inflate SGA scores independently of amplification.

    Authors: We agree that explicit verification of semantic overlap would strengthen the isolation of generative effects. The aligned set was constructed by extracting style and gender descriptors from historical annotations to form matched prompts, with the goal of holding semantic content constant. In revision we will add quantitative alignment checks, including CLIP embedding cosine similarities between historical and generated pairs plus a sampled review of object categories and layouts, to confirm that SGA differences reflect amplification rather than content drift. revision: yes

  2. Referee: Metrics section: no inter-annotator agreement, label validation, or statistical controls for prompt content are described for the gender annotations underlying PixelSGA and MaskSGA. Without these, the empirical comparisons rest on unexamined measurement assumptions and cannot reliably support the amplification result.

    Authors: Gender labels on historical images followed standard art-historical visual criteria (e.g., attire, posture, facial features) applied by domain experts; generated-image labels derive directly from the controlled gender prompts. Inter-annotator agreement was not computed originally. We will expand the metrics section to detail the annotation protocol and prompt-template controls (fixed structure varying only style/gender tokens). Because IAA data were not collected, we cannot retroactively report it but will note this limitation and the reliance on established conventions. revision: partial

Circularity Check

0 steps flagged

No circularity in empirical dataset comparisons and metric definitions

full rationale

The paper introduces a new dataset (StyleGender) and two metrics (PixelSGA, MaskSGA) to perform direct empirical comparisons of gender signals between historical art images and T2I-generated images under controlled prompts. The central claim of amplification is obtained by computing these metrics on independent collections and observing higher values in the generated set; no equations, fitted parameters, or self-citations reduce this comparison to a tautology or to the inputs by construction. The construction of the semantically aligned set via prompts is a methodological step whose validity can be debated on external grounds, but it does not create a self-definitional or fitted-input loop within the reported results.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that gender can be reliably annotated as a binary or categorical attribute in both historical and generated images and that style keywords function as the primary causal carrier of gender signals. No free parameters or invented physical entities are introduced.

axioms (2)
  • domain assumption Gender labels assigned to images are stable and comparable across historical and generated domains
    Required for both the dataset construction and the SGA metric comparisons; location implied in the description of annotations and controlled prompts.
  • domain assumption Style keywords in prompts isolate the transfer of gender artifacts without confounding changes in subject matter or composition
    Underpins the claim that observed amplification is due to style transfer rather than prompt content variation.

pith-pipeline@v0.9.1-grok · 5720 in / 1326 out tokens · 23836 ms · 2026-06-28T02:14:31.402474+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

74 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Stable diffusion 3.5 medium

    Stability AI. Stable diffusion 3.5 medium. https://huggingface.co/stabilityai/ stable-diffusion-3.5-medium, 2024. Accessed: 2026-05-06

    2024

  2. [2]

    Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, 2022

    2022

  3. [3]

    Pinar Barlas, Kyriakos Kyriakou, Olivia Guest, Styliani Kleanthous, and Jahna Otterbacher. To" see" is to stereotype: Image tagging algorithms, gender recognition, and the accuracy-fairness trade-off.Proceedings of the ACM on Human-Computer Interaction, 4(CSCW3):1–31, 2021

    2021

  4. [4]

    An anthropology of images: Picture, medium, body

    Hans Belting. An anthropology of images: Picture, medium, body. 2022

    2022

  5. [5]

    A comprehensive survey on object detection in visual art: taxonomy and challenge.Multimedia Tools and Applications, 83(5):14637–14670, 2024

    Siwar Bengamra, Olfa Mzoughi, André Bigand, and Ezzeddine Zagrouba. A comprehensive survey on object detection in visual art: taxonomy and challenge.Multimedia Tools and Applications, 83(5):14637–14670, 2024

    2024

  6. [6]

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

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to- image generation amplifies demographic stereotypes at large scale. InProceedings of the 2023 ACM conference on fairness, accountability, and transparency, pages 1493–1504, 2023

    2023

  7. [7]

    Into the laion’s den: Investigating hate in multimodal datasets.Advances in neural information processing systems, 36:21268–21284, 2023

    Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion’s den: Investigating hate in multimodal datasets.Advances in neural information processing systems, 36:21268–21284, 2023

    2023

  8. [8]

    Large image datasets: A pyrrhic win for computer vision? In2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536–1546

    Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536–1546. IEEE, 2021

    2021

  9. [9]

    Multimodal datasets: misog- yny, pornography, and malignant stereotypes, 2021

    Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misog- yny, pornography, and malignant stereotypes, 2021

    2021

  10. [10]

    Gender shades: Intersectional accuracy disparities in commercial gender classification

    Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. InConference on fairness, accountability and transparency, pages 77–91. PMLR, 2018

    2018

  11. [11]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergiy Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision (ECCV), pages 213–229. Springer, 2020

    2020

  12. [12]

    Controlstyle: Text-driven stylized image generation using diffusion priors

    Jingwen Chen, Yingwei Pan, Ting Yao, and Tao Mei. Controlstyle: Text-driven stylized image generation using diffusion priors. InProceedings of the 31st ACM international conference on multimedia, pages 7540–7548, 2023

    2023

  13. [13]

    Tibet: Identifying and evaluating biases in text-to-image generative models

    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. InEuropean Conference on Computer Vision, pages 429–446. Springer, 2024

    2024

  14. [14]

    Laurence King Publishing, 2004

    Thomas E Crow.The rise of the sixties: American and European art in the era of dissent. Laurence King Publishing, 2004. 10

    2004

  15. [15]

    Oasis uncovers: High-quality t2i models, same old stereotypes

    Sepehr Dehdashtian, Gautam Sreekumar, and Vishnu Boddeti. Oasis uncovers: High-quality t2i models, same old stereotypes. InThe Thirteenth International Conference on Learning Representations

  16. [16]

    Aesthetics as structural harm: Algorithmic lookism across text-to-image generation and classification, 2026

    Miriam Doh, Aditya Gulati, Corinna Canali, and Nuria Oliver. Aesthetics as structural harm: Algorithmic lookism across text-to-image generation and classification, 2026

    2026

  17. [17]

    Position: The categorization of race in ml is a flawed premise

    Miriam Doh, Benedikt Höltgen, Piera Riccio, and Nuria M Oliver. Position: The categorization of race in ml is a flawed premise. InForty-second International Conference on Machine Learning Position Paper Track, 2025

    2025

  18. [18]

    CAN: Creative Adversarial Networks, Generating "Art" by Learning About Styles and Deviating from Style Norms

    Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone. Can: Creative adversarial networks, generating" art" by learning about styles and deviating from style norms. arXiv preprint arXiv:1706.07068, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  19. [19]

    Cultural bias in text-to-image models: a systematic review of bias identification, evaluation and mitigation strategies.IEEE Access, 2025

    Wala Elsharif, Mahmood Alzubaidi, and Marco Agus. Cultural bias in text-to-image models: a systematic review of bias identification, evaluation and mitigation strategies.IEEE Access, 2025

    2025

  20. [20]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024

    2024

  21. [21]

    Art nouveau: A universal style

    Europeana. Art nouveau: A universal style. https://www.europeana.eu/en/ exhibitions/art-nouveau-a-universal-style, 2017. Last Access: 04.03.2026

    2017

  22. [22]

    A survey on bias in visual datasets.Computer Vision and Image Understanding, 223:103552, 2022

    Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets.Computer Vision and Image Understanding, 223:103552, 2022

    2022

  23. [23]

    Fraser, Svetlana Kiritchenko, and Isar Nejadgholi

    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. A friendly face: Do text-to-image systems rely on stereotypes when the input is under-specified?, 2023

    2023

  24. [24]

    Uncurated image-text datasets: Shedding light on demographic bias

    Noa Garcia, Yusuke Hirota, Yankun Wu, and Yuta Nakashima. Uncurated image-text datasets: Shedding light on demographic bias. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6957–6966, 2023

    2023

  25. [25]

    Harm amplification in text-to-image models

    Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Ryan Poplin, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024

    work page arXiv 2024

  26. [26]

    Bias in gender bias benchmarks: How spurious features distort evaluation

    Yusuke Hirota, Ryo Hachiuma, Boyi Li, Ximing Lu, Michael Ross Boone, Boris Ivanovic, Yejin Choi, Marco Pavone, Yu-Chiang Frank Wang, Noa Garcia, et al. Bias in gender bias benchmarks: How spurious features distort evaluation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8634–8644, 2025

    2025

  27. [27]

    Quantifying societal bias amplification in image captioning

    Yusuke Hirota, Yuta Nakashima, and Noa Garcia. Quantifying societal bias amplification in image captioning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13450–13459, 2022

    2022

  28. [28]

    Midjourney.https://www.midjourney.com, 2022

    David Holz. Midjourney.https://www.midjourney.com, 2022

    2022

  29. [29]

    Neu- ral style transfer: A review.IEEE transactions on visualization and computer graphics, 26(11):3365–3385, 2019

    Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neu- ral style transfer: A review.IEEE transactions on visualization and computer graphics, 26(11):3365–3385, 2019

    2019

  30. [30]

    Ultralytics YOLO, 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO, 2023

    2023

  31. [31]

    Gender representations in Neoclassical art

    Knowledge.Deck.no. Gender representations in Neoclassical art. https://knowledge.deck. no/art-and-literature/art-history/neoclassicism/gender-representations ,

  32. [32]

    Last Access: 04.03.2026

    2026

  33. [33]

    Content and style disentanglement for artistic style transfer

    Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 11

    2019

  34. [34]

    Flux.https://github.com/black-forest-labs/flux, 2024

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

    2024

  35. [35]

    Flux.1-dev

    Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/FLUX. 1-dev, 2024. Accessed: 2026-05-06

    2024

  36. [36]

    Stable bias: Evaluat- ing societal representations in diffusion models.Advances in Neural Information Processing Systems, 36:56338–56351, 2023

    Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluat- ing societal representations in diffusion models.Advances in Neural Information Processing Systems, 36:56338–56351, 2023

    2023

  37. [37]

    Bigbench: A unified benchmark for evaluating multi-dimensional social biases in text-to-image models.arXiv preprint arXiv:2407.15240, 2024

    Hanjun Luo, Haoyu Huang, Ziye Deng, Xinfeng Li, Hewei Wang, Yingbin Jin, Yang Liu, Wenyuan Xu, and Zuozhu Liu. Bigbench: A unified benchmark for evaluating multi-dimensional social biases in text-to-image models.arXiv preprint arXiv:2407.15240, 2024

    work page arXiv 2024

  38. [38]

    Representing medieval genders and sexualities in europe: Construction, transformation, and subversion, 600–1530

    Elizabeth L’Estrange and Alison More. Representing medieval genders and sexualities in europe: Construction, transformation, and subversion, 600–1530. InRepresenting Medieval Genders and Sexualities in Europe, pages 1–13. Routledge, 2016

    2016

  39. [39]

    Multimodal composite association score: Measuring gender bias in generative multimodal models.arXiv preprint arXiv:2304.13855, 2023

    Abhishek Mandal, Susan Leavy, and Suzanne Little. Multimodal composite association score: Measuring gender bias in generative multimodal models.arXiv preprint arXiv:2304.13855, 2023

    work page arXiv 2023

  40. [40]

    Analysing gender bias in text-to-image models using object detection

    Harvey Mannering. Analysing gender bias in text-to-image models using object detection. arXiv preprint arXiv:2307.08025, 2023

    work page arXiv 2023

  41. [41]

    Gender artifacts in visual datasets

    Nicole Meister, Dora Zhao, Angelina Wang, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Gender artifacts in visual datasets. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4837–4848, 2023

    2023

  42. [42]

    Civiverse: A dataset for analyz- ing user engagement with open-source text-to-image models.arXiv preprint arXiv:2408.15261, 2024

    Maria-Teresa De Rosa Palmini, Laura Wagner, and Eva Cetinic. Civiverse: A dataset for analyz- ing user engagement with open-source text-to-image models.arXiv preprint arXiv:2408.15261, 2024

    work page arXiv 2024

  43. [43]

    Styleforge: Enhancing text-to- image synthesis for any artistic styles with dual binding.Applied Sciences, 15(19), 2025

    Junseo Park, Beomseok Ko, Minji Kang, and Hyeryung Jang. Styleforge: Enhancing text-to- image synthesis for any artistic styles with dual binding.Applied Sciences, 15(19), 2025

    2025

  44. [44]

    Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models

    Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. InProceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3403–3417, 2023

    2023

  45. [45]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

  46. [46]

    No annota- tions for object detection in art through stable diffusion

    Patrick Ramos, Nicolas Gonthier, Selina Khan, Yuta Nakashima, and Noa Garcia. No annota- tions for object detection in art through stable diffusion. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6228–6237. IEEE, 2025

    2025

  47. [47]

    High- resolution image synthesis with latent diffusion models

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

    2022

  48. [48]

    Brill, 2015

    Marice E Rose and Alison C Poe.Receptions of antiquity, constructions of gender in European art, 1300-1600. Brill, 2015

    2015

  49. [49]

    Training-free style and content transfer by leveraging u-net skip connections in stable diffusion.arXiv preprint arXiv:2501.14524, 2025

    Ludovica Schaerf, Andrea Alfarano, Fabrizio Silvestri, and Leonardo Impett. Training-free style and content transfer by leveraging u-net skip connections in stable diffusion.arXiv preprint arXiv:2501.14524, 2025

    work page arXiv 2025

  50. [50]

    Theory and philosophy of art: Style, artist, and society

    Meyer Schapiro. Theory and philosophy of art: Style, artist, and society. 1994

    1994

  51. [51]

    Smiling women pitching down: auditing representational and presentational gender biases in image-generative ai.Journal of Computer-Mediated Communication, 29(1):zmad045, 2024

    Luhang Sun, Mian Wei, Yibing Sun, Yoo Ji Suh, Liwei Shen, and Sijia Yang. Smiling women pitching down: auditing representational and presentational gender biases in image-generative ai.Journal of Computer-Mediated Communication, 29(1):zmad045, 2024. 12

    2024

  52. [52]

    PhD thesis, 2016

    Laura Elizabeth Thiel.The Gentleman-Scholar at Home: Domesticity, Masculinity, and Civility in Dutch Seventeenth-Century Genre Painting. PhD thesis, 2016

    2016

  53. [53]

    How many pixels make an image?Visual neuroscience, 26(1):123–131, 2009

    Antonio Torralba. How many pixels make an image?Visual neuroscience, 26(1):123–131, 2009

    2009

  54. [54]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. InCVPR 2011, pages 1521–1528. IEEE, 2011

    2011

  55. [55]

    Stereotypes and smut: The (mis) representa- tion of non-cisgender identities by text-to-image models

    Eddie Ungless, Björn Ross, and Anne Lauscher. Stereotypes and smut: The (mis) representa- tion of non-cisgender identities by text-to-image models. InFindings of the Association for Computational Linguistics: ACL 2023, pages 7919–7942, 2023

    2023

  56. [56]

    Survey of bias in text-to-image generation: Definition, evaluation, and mitigation.arXiv preprint arXiv:2404.01030, 2024

    Yixin Wan, Arjun Subramonian, Anaelia Ovalle, Zongyu Lin, Ashima Suvarna, Christina Chance, Hritik Bansal, Rebecca Pattichis, and Kai-Wei Chang. Survey of bias in text-to-image generation: Definition, evaluation, and mitigation.arXiv preprint arXiv:2404.01030, 2024

    work page arXiv 2024

  57. [57]

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

    Jialu Wang, Xinyue Liu, Zonglin Di, Yang Liu, and Xin Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation. InFindings of the Association for Computa- tional Linguistics: ACL 2023, pages 2560–2574, 2023

    2023

  58. [58]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision- language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  59. [59]

    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 Multimedia, pages 3781–3789, 2024

    2024

  60. [60]

    Diffusiondb: A large-scale prompt gallery dataset for text-to-image genera- tive models.arXiv preprint arXiv:2210.14896, 2022

    Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image genera- tive models.arXiv preprint arXiv:2210.14896, 2022

    work page arXiv 2022

  61. [61]

    Beyond the universal art dataset: Issues and mitigations of western bias in computational art analysis

    Amanda Wasielewski. Beyond the universal art dataset: Issues and mitigations of western bias in computational art analysis. InProceedings of the ECCV 2022 Workshop on Computer VISion on ART analysis, number 4, pages 1–6, 2022

    2022

  62. [62]

    The reification of style in ai image generation

    Amanda Wasielewski. The reification of style in ai image generation. 2024

    2024

  63. [63]

    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. InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society, volume 8, pages 2705–2717, 2025

    2025

  64. [64]

    Artfid: Quantitative evaluation of neural style transfer

    Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560–576. Springer, 2022

    2022

  65. [65]

    Goya: leveraging generative art for content-style disentanglement.Journal of Imaging, 10(7):156, 2024

    Yankun Wu, Yuta Nakashima, and Noa Garcia. Goya: leveraging generative art for content-style disentanglement.Journal of Imaging, 10(7):156, 2024

    2024

  66. [66]

    Stable diffusion exposed: Gender bias from prompt to image

    Yankun Wu, Yuta Nakashima, and Noa Garcia. Stable diffusion exposed: Gender bias from prompt to image. InProceedings of the AAAI/ACM conference on AI, ethics, and society, volume 7, pages 1648–1659, 2024

    2024

  67. [67]

    Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy

    Kaiyu Yang, Klint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. InProceedings of the 2020 conference on fairness, accountability, and transparency, pages 547–558, 2020

    2020

  68. [68]

    Enhancing fairness in face detection in computer vision systems by demographic bias mitigation

    Yu Yang, Aayush Gupta, Jianwei Feng, Prateek Singhal, Vivek Yadav, Yue Wu, Pradeep Natara- jan, Varsha Hedau, and Jungseock Joo. Enhancing fairness in face detection in computer vision systems by demographic bias mitigation. InProceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society, pages 813–822, 2022. 13

    2022

  69. [69]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

    2023

  70. [70]

    Iti-gen: Inclusive text-to-image generation

    Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-gen: Inclusive text-to-image generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969–3980, 2023

    2023

  71. [71]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023

    2023

  72. [72]

    Bias in generative ai, 2024

    Mi Zhou, Vibhanshu Abhishek, Timothy Derdenger, Jaymo Kim, and Kannan Srinivasan. Bias in generative ai, 2024

    2024

  73. [73]

    A painting of 〈gender〉, 〈viewpoint〉, 〈pose〉[〈modifier〉], in the artistic style known as〈style〉

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. InProceedings of the IEEE international conference on computer vision, pages 2223–2232, 2017. 14 A The STYLEGENDERDataset In this section we provide additional information on how the dataset STYLEGENDERis construct...

    2017

  74. [74]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...