Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Stereotype bias in modern text-to-image transformers is concentrated in a few internal layers that FairFlow steers to restore fairness without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 04:18 UTC pith:ZQC4QYQT

load-bearing objection Solid MM-DiT fairness paper: sparse hub localization plus early-window steering is real and useful, even if the causal map is still occupation-heavy. the 3 major comments →

arxiv 2607.03180 v1 pith:ZQC4QYQT submitted 2026-07-03 cs.CV

FairFlow: Demystifying and Mitigating Stereotype Bias in Text-to-Image Diffusion Transformers

Chen Chen , Yuanmin Huang , Zhenfei Zhang , Mi Zhang , Xiaohan Zhang , Yun Xiong , Xiaoyu You , Min Yang This is my paper
classification cs.CV
keywords fairnesstext-to-imagediffusion transformersbias mitigationsemantic binding hubsstereotype biasMM-DiTinference-time steering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Multimodal diffusion transformers that power current text-to-image systems systematically produce gender- and race-stereotyped images of occupations even from neutral prompts. This paper shows the bias is not spread evenly through the network; a sparse set of layers act as semantic binding hubs that first sketch structure, then inject stereotypical concepts from text, then broadcast them across the image. FairFlow learns small attribute-specific steering directions and injects them only at those hubs during early generation steps. On FLUX.1-dev and Stable Diffusion 3 the method balances demographic groups far better than prior fixes while keeping image quality and prompt fidelity high, at almost no extra compute. A sympathetic reader cares because deployed systems already face public backlash for these failures, and a lightweight internal fix is more practical than retraining or prompt rewriting.

Core claim

Bias representations in multimodal diffusion transformers are not uniformly distributed across depth; they are mediated by a sparse set of semantic binding hubs with stage-wise roles. Early hubs establish structural templates susceptible to bias, middle hubs extract core stereotypical concepts from textual conditioning, and late hubs solidify those biases through visual self-attention. Injecting optimized attribute-specific fair directions only at these hubs inside a constrained early inference window neutralizes gender, race, and intersectional occupational stereotypes while preserving fidelity.

What carries the argument

Semantic binding hubs — the sparse transformer layers located by residual causal tracing that stage-wise bind demographic attributes — plus FairFlow sparse steering, which learns regularized attribute-specific fair directions and injects them only at those hubs within the early denoising window.

Load-bearing premise

Same-seed residual patching of attribute-swapped occupation prompts over early timesteps correctly identifies the layers that causally mediate demographic binding.

What would settle it

If single-layer residual swaps at the claimed hubs fail to shift generated faces toward the target demographic while swaps at non-hub layers succeed, or if FairFlow steering at those hubs leaves fairness scores near the vanilla baseline, the sparse-hub claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Deployed MM-DiT services can reduce occupational gender and race stereotypes at inference time without separate fairness checkpoints or full retraining.
  • Text-side prompt rewriting and embedding neutralization become less necessary once the visual backbone’s binding points are controlled.
  • Cross-modal bias reinforcement can be disrupted by visual-only interventions, so text-side neutrality alone is incomplete.
  • The same hub-localization and sparse-steering recipe can be reapplied to other protected attributes once new fair directions are learned.

Where Pith is reading between the lines

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

  • Similar sparse hubs likely govern style, object, and other social-attribute binding, so the localization method could support broader controllable generation beyond fairness.
  • An automated monitor could apply FairFlow only on underspecified occupational prompts and leave explicit attribute requests untouched.
  • Hub indices may not transfer across model families; each new MM-DiT architecture would need its own short causal-tracing pass before deployment.

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

3 major / 5 minor

Summary. The paper argues that stereotype bias in multimodal diffusion transformers is not uniformly distributed across depth, but is mediated by a sparse set of semantic binding hubs with stage-wise roles: early hubs form structural templates, middle hubs inject stereotypical concepts from text, and late hubs consolidate them via visual self-attention. Using residual causal tracing on same-seed attribute-swapped occupation prompts (Eq. 6; Fig. 3), the authors identify hubs on FLUX.1-dev and Stable Diffusion 3, support the stage-wise picture with I2I/I2T attention metrics (Eqs. 7–11; Fig. 4), and document a cross-modal reinforcement loop (Fig. 5). They then propose FairFlow: optimize attribute-specific steering vectors with hub-matching, output-matching, and magnitude regularization (Eqs. 12–15), inject them only at hubs in an early window t∈[1.0,0.7] with norm rescaling (Eqs. 16–17), and evaluate fairness–fidelity tradeoffs against prompt-, embedding-, and parameter-updating baselines on gender, race, and intersectional occupational prompts, plus COCO utility, complex scenes, reinforcement disruption, ablations, and overhead.

Significance. If the hub localization and sparse-steering results hold beyond the occupation template, the paper would be a useful contribution for both mechanistic understanding of MM-DiTs and practical deployment-time fairness. The combination of residual causal tracing, attention-based functional roles, and a parameter-preserving intervention that reports strong fairness–fidelity tradeoffs on two modern backbones (Tables 1–2), near-zero overhead (Table 4), and disruption of the internal reinforcement loop (Fig. 9) is more architecture-aware than most prior T2I debiasing work, which was largely U-Net- or text-side-centric. The planned public release of code and datasets would further strengthen reproducibility. The main significance is therefore conditional on the hubs being genuine causal control points for demographic binding rather than artifacts of the occupation-prompt localization protocol.

major comments (3)
  1. [§4.1, Eq. (6), Fig. 3] §4.1 and Eq. (6): residual causal tracing is the load-bearing justification for sparse hub intervention, but it is performed only on same-seed occupation prompts that differ by an attribute token, patches only visual residuals in t∈[1.0,0.7], and ranks layers by 1−LPIPS. The paper’s own assumptions (stable non-attribute semantics under token swap; early high-level binding) are not stress-tested on multi-subject, non-occupation, or non-face prompts. Without transfer evidence that the selected hubs (FLUX 2/18/26/54; SD3 0/1/3/19/21/22) remain causal outside this template, the claim that bias is generally mediated by these hubs is under-supported, and the sparse design risks being occupation-specific.
  2. [Tables 1–2, §6 Metrics, Eq. (18)] Tables 1–2 and §6 Metrics: fairness is defined from DeepFace group proportions (Eq. 18) with no reported multi-seed variance, confidence intervals, or classifier-error analysis. Because DeepFace itself can be biased or brittle on generated faces, the large fairness gains (e.g., FLUX gender 0.084→0.600) may partly reflect classifier-sensitive face shifts rather than robust demographic balance. At minimum, multi-seed error bars and a second independent attribute estimator (or human audit on a subset) are needed before the ‘optimal fairness–fidelity’ claim can be treated as settled.
  3. [Table 3, §7.5] Table 3 ablations only compare DirectDiff, random layers, bottom layers, and α on the same occupation distribution used for hub discovery. They do not test whether hub selection transfers to the complex-scene set (RQ3) or to non-occupation prompts. A transfer ablation—re-localize hubs on a held-out prompt family, or apply occupation-derived hubs to non-occupation bias settings—would substantially strengthen (or falsify) the mechanistic story that underwrites FairFlow.
minor comments (5)
  1. [Figure 1] Figure 1 is described as showing the most diverse demographic distribution, but the main text does not quantify how the displayed samples were selected; a short caption note on sampling would help.
  2. [Fig. 2, §5.1] Notation for residual outputs r_l^s / r_l^t in Fig. 2 is not fully aligned with the h_l / ô notation used in §5.1; a brief glossary would reduce friction.
  3. [§5.2, §6 Implementation] The intervention window [1.0,0.7] and α values are stated as fixed; a short sensitivity plot (beyond the low/high α rows in Table 3) would make the free-parameter surface clearer.
  4. [§2.3] Related-work coverage of recent MM-DiT layer analyses is present but brief; a clearer contrast with Stable Flow and other vital-layer editing papers would better position the fairness-specific contribution.
  5. [Appendix B, §6] Appendix ethical note correctly flags DeepFace and simplified gender/race categories as proxies; a one-sentence pointer to that limitation in the main evaluation section would be appropriate.

Circularity Check

0 steps flagged

No equation-level circularity: hubs from causal patching, vectors from feature matching, fairness from an external classifier.

full rationale

FairFlow’s chain is localization → optimized sparse steering → external fairness evaluation, and none of these steps reduces to its inputs by construction. Residual causal tracing (Eq. 6) ranks layers by 1−LPIPS after same-seed visual residual patching; that is an empirical causal probe, not a definition of fairness. Steering vectors are then fit with L_hub + L_out + L_reg to match target hub features and denoising outputs on held-out development occupations, not to maximize the DeepFace fairness score F. Final fairness, CLIP, and MUSIQ are measured on a separate 100-occupation test set, complex-scene prompts, and COCO utility prompts. There is no self-definitional loop, no fitted parameter renamed as a prediction of the same quantity, no load-bearing uniqueness theorem from the authors, and no ansatz smuggled in via self-citation. Mild shared use of occupation-style prompts for hub finding and evaluation is domain overlap, not circular derivation. The paper is self-contained against external baselines and metrics; score 0 is the honest finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central story rests on standard generative-model machinery plus several domain assumptions about early semantic binding and LPIPS-based causal scores, plus hand-chosen intervention hyperparameters and the conceptual construct of semantic binding hubs. No new physical entity is postulated; the main invented construct is the hub abstraction used to justify sparse steering.

free parameters (5)
  • steering strength α
    Chosen per model and attribute (e.g., FLUX gender 7.0, race 5.0; SD3 3.0; intersectional 2.0/3.0) and shown in ablation to trade fairness vs quality; not derived from first principles.
  • loss weights λ_hub, λ_out, λ_reg
    Set to 1.0, 0.2, 0.01 for steering-vector optimization; control how closely hubs/outputs match and how large vectors grow.
  • intervention time window [1.0, 0.7]
    Early-window restriction used for both causal tracing and injection; justified by prior high-level-semantics timing claims but selected as a fixed hyperparameter.
  • hub layer sets
    FLUX layers 2,18,26,54 and SD3 layers 0,1,3,19,21,22 selected from causal-score peaks; peak-picking and count of hubs involve judgment beyond a unique automatic rule.
  • optimizer schedule (lr, epochs)
    Adam 5e-4; 30 epochs on FLUX, 100 on SD3—training choices that affect learned fair directions.
axioms (5)
  • domain assumption Same seed with only the demographic attribute token changed keeps non-attribute semantics relatively stable in high-fidelity MM-DiTs.
    Stated as a tracing assumption in §4.1; required for residual patching to isolate attribute causality.
  • domain assumption Demographic attributes are high-level factors that bind mainly in an early portion of the rectified-flow trajectory.
    Used to restrict patching and FairFlow injection to t∈[1.0,0.7] (§4.1, §5.2).
  • domain assumption 1−LPIPS between patched and target images is a valid scalar for layer-wise causal influence of demographic attributes.
    Defines C(l) in Eq. (6); perceptual distance stands in for attribute-specific causal effect.
  • domain assumption DeepFace facial-attribute classifications are adequate proxies for demographic group proportions in fairness scoring.
    Fairness metric F in §6 depends on this automated labeler; paper notes proxy limitations in ethics appendix.
  • standard math Rectified-flow / flow-matching generation and joint-attention MM-DiT blocks as described for FLUX and SD3.
    Background generative model equations (1)–(3) and architecture facts from cited model papers.
invented entities (2)
  • semantic binding hubs no independent evidence
    purpose: Name the sparse layers that causally mediate demographic attribute binding and justify restricting intervention to those layers.
    Defined operationally via residual causal tracing peaks and attention roles; not an independently measured physical object outside this analysis pipeline.
  • attribute-specific fair directions (steering vectors c_l^a) no independent evidence
    purpose: Provide compact additive interventions in image hidden space at hub layers.
    Learned from source–target pairs with a regularized objective; standard steering-vector style construct specialized to hubs.

pith-pipeline@v1.1.0-grok45 · 26973 in / 3440 out tokens · 39544 ms · 2026-07-12T04:18:05.837815+00:00 · methodology

0 comments
read the original abstract

Multimodal diffusion transformers (MM-DiTs) have emerged as the prevalent backbone for modern text-to-image generation systems. However, they exhibit critical alignment vulnerabilities, systematically manifesting severe stereotype biases even under benign prompts. This poses a significant risk of algorithmic discrimination in deployed systems. Since most existing mitigation strategies were tailored for legacy U-Net architectures, the precise remediation of these vulnerabilities in MM-DiTs remains a critical open challenge. In this work, we first investigate the root cause of this vulnerability via mechanistic analysis. We reveal that bias representations in MM-DiTs are not uniformly distributed across depth, but are mediated by a sparse set of layers functioning as internal semantic binding hubs. These hubs exhibit a stage-wise propagation driving bias manifestation: early hubs establish the structural templates susceptible to bias, middle hubs actively extract core stereotypical concepts from textual conditioning, and late hubs globally solidify these biases through visual self-attention. Leveraging these architectural insights, we propose FairFlow, an intrinsic, mechanism-guided mitigation framework. FairFlow acts as an internal regulator by employing sparse steering: it learns attribute-specific fair directions and injects them exclusively at the identified semantic hubs within a constrained inference window. Evaluations on FLUX.1-dev and Stable Diffusion~3 demonstrate that FairFlow effectively neutralizes these stereotypical vulnerabilities across gender, race, and intersectional settings, achieving an optimal fairness-fidelity balance. With near-zero inference overhead and robustness to complex prompts, FairFlow provides a lightweight and practical bias mitigation for large-scale deployed MM-DiT systems. Code and datasets will be publicly released upon acceptance.

Figures

Figures reproduced from arXiv: 2607.03180 by Chen Chen, Min Yang, Mi Zhang, Xiaohan Zhang, Xiaoyu You, Yuanmin Huang, Yun Xiong, Zhenfei Zhang.

Figure 1
Figure 1. Figure 1: Comparison of vanilla FLUX, two baseline debias [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Layer-wise causal scores from residual causal trac [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Residual causal tracing for locating hub layers. For [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: I2I and I2T attention analysis on FLUX.1-dev and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-modal bias reinforcement on Stable Diffu [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of FairFlow. FairFlow precomputes attribute-specific fair directions, injects them only at the semantic binding hubs identified by causal tracing, restricts intervention to an early denoising window, and evaluates the resulting generations with fairness and quality metrics. The design is parameter-free at deployment time, efficient, and lightweight. This term directly fits the causal layers identi… view at source ↗
Figure 7
Figure 7. Figure 7: Fidelity scores on bias-irrelevant COCO prompts. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fairness–fidelity tradeoff on complex-scene [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of FairFlow on cross-modal bias reinforce￾ment. We compare the source and steered forward passes by measuring (a) their cosine similarity to the bias and anti-bias trajectories and (b) their positions on the bias axis during generation [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison on bias-irrelevant COCO prompts. Left: FLUX.1-dev. Right: Stable Diffusion 3. In both models, [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison on complex-scene prompts. Left: FLUX.1-dev. Right: Stable Diffusion 3. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗

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

57 extracted references · 10 linked inside Pith

  1. [1]

    2024.AI Image Generators Often Give Racist and Sexist Results: Can They Be Fixed?https://www.nature.com/articles/d41586-024-00674-9 Nature News Feature

    Ananya. 2024.AI Image Generators Often Give Racist and Sexist Results: Can They Be Fixed?https://www.nature.com/articles/d41586-024-00674-9 Nature News Feature. Accessed April 17, 2026

    2024

  2. [2]

    Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. 2025. Stable flow: Vital layers for training-free image editing. InProceedings of the Computer Vision and Pattern Recognition Conference. 7877–7888

    2025

  3. [3]

    Basim Azam and Naveed Akhtar. 2025. Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control. InPro- ceedings of the Computer Vision and Pattern Recognition Conference. 2976–2985

    2025

  4. [4]

    Samyadeep Basu, Keivan Rezaei, Priyatham Kattakinda, Vlad I Morariu, Nanxuan Zhao, Ryan A Rossi, Varun Manjunatha, and Soheil Feizi. 2024. On mechanis- tic knowledge localization in text-to-image generative models. InForty-first International Conference on Machine Learning

    2024

  5. [5]

    Samyadeep Basu, Nanxuan Zhao, Vlad I Morariu, Soheil Feizi, and Varun Man- junatha. 2023. Localizing and editing knowledge in text-to-image generative models. InThe Twelfth International Conference on Learning Representations

    2023

  6. [6]

    Black Forest Labs. 2024. Announcing Black Forest Labs. https://blackforestlabs. ai/announcing-black-forest-labs/

    2024

  7. [7]

    Sepehr Dehdashtian, Gautam Sreekumar, and Vishnu Naresh Boddeti. 2025. Oasis uncovers: High-quality t2i models, same old stereotypes.arXiv preprint arXiv:2501.00962(2025)

    Pith/arXiv arXiv 2025

  8. [8]

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

    2025

  9. [9]

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

    2024

  10. [10]

    fal.ai. 2026. Generative media platform for developers. https://fal.ai/

    2026

  11. [11]

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. 2023. Fair diffusion: Instruct- ing text-to-image generation models on fairness.arXiv preprint arXiv:2302.10893 (2023)

    Pith/arXiv arXiv 2023

  12. [12]

    Zihao Fu, Ryan Brown, Shun Shao, Kai Rawal, Eoin Delaney, and Chris Russell

  13. [13]

    arXiv preprint arXiv:2510.21363(2025)

    FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models. arXiv preprint arXiv:2510.21363(2025)

    arXiv 2025

  14. [14]

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. 2024. Unified concept editing in diffusion models. InProceedings of the IEEE/CVF winter conference on applications of computer vision. 5111–5120

    2024

  15. [15]

    Boyu Han, Qianqian Xu, Shilong Bao, Zhiyong Yang, Kangli Zi, and Qingming Huang. 2025. LightFair: Towards an Efficient Alternative for Fair T2I Diffusion via Debiasing Pre-trained Text Encoders.arXiv preprint arXiv:2509.23639(2025)

    arXiv 2025

  16. [16]

    Lin-Chun Huang, Ching Chieh Tsao, Fang-Yi Su, and Jung-Hsien Chiang. 2025. Debiasing diffusion model: Enhancing fairness through latent representation learning in stable diffusion model.arXiv preprint arXiv:2503.12536(2025)

    Pith/arXiv arXiv 2025

  17. [17]

    Sangwon Jung, Alex Oesterling, Claudio Mayrink Verdun, Sajani Vithana, Taesup Moon, and Flavio P Calmon. 2025. Multi-Group Proportional Representations for Text-to-Image Models. InProceedings of the Computer Vision and Pattern Recognition Conference. 23744–23754

    2025

  18. [18]

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. 2021. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF interna- tional conference on computer vision. 5148–5157

    2021

  19. [19]

    Eunji Kim, Siwon Kim, Minjun Park, Rahim Entezari, and Sungroh Yoon. 2025. Rethinking training for de-biasing text-to-image generation: Unlocking the po- tential of stable diffusion. InProceedings of the Computer Vision and Pattern Recognition Conference. 13361–13370

    2025

  20. [20]

    Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. 2023. De-stereotyping text-to-image models through prompt tuning. (2023)

    2023

  21. [21]

    Binglei Li, Mengping Yang, Zhiyu Tan, Junping Zhang, and Hao Li. [n. d.]. Revis- iting Block-wise Interactions of MMDiT for Training-free Improved Synthesis. ([n. d.])

  22. [22]

    Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. 2024. Self- discovering interpretable diffusion latent directions for responsible text-to-image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12006–12016

    2024

  23. [23]

    Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, and Di Wang

  24. [24]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Fair text-to-image diffusion via fair mapping. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 26256–26264

  25. [25]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. InEuropean conference on computer vision. Springer, 740–755

    2014

  26. [26]

    Zihao Lin, Samyadeep Basu, Mohammad Beigi, Varun Manjunatha, Ryan A Rossi, Zichao Wang, Yufan Zhou, Sriram Balasubramanian, Arman Zarei, Keivan Rezaei, et al. 2025. A survey on mechanistic interpretability for multi-modal foundation models.arXiv preprint arXiv:2502.17516(2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le

  28. [28]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747 (2022)

    Pith/arXiv arXiv 2022

  29. [29]

    Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003(2022)

    Pith/arXiv arXiv 2022

  30. [30]

    Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. 2023. Stable bias: Evaluating societal representations in diffusion models.Advances in Neural Information Processing Systems36 (2023), 56338–56351

    2023

  31. [31]

    Yunbo Lyu, Zhou Yang, Yuqing Niu, Jing Jiang, and David Lo. 2025. Do existing testing tools really uncover gender bias in text-to-image models?. InProceedings of the 33rd ACM International Conference on Multimedia. 11687–11696

    2025

  32. [32]

    Ranjita Naik and Besmira Nushi. 2023. Social biases through the text-to-image generation lens. InProceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. 786–808

    2023

  33. [33]

    Jeonghoon Park, Juyoung Lee, Chaeyeon Chung, Jaeseong Lee, Jaegul Choo, and Jindong Gu. 2025. Fair Generation without Unfair Distortions: Debiasing Text-to-Image Generation with Entanglement-Free Attention. InProceedings of the IEEE/CVF International Conference on Computer Vision. 17567–17576

    2025

  34. [34]

    Fan Qi, Zhan Wang, Changsheng Xu, and Huaiwen Zhang. 2025. Fine-tuning Bias Neurons for Fair Text-to-Image Generation. InProceedings of the 33rd ACM International Conference on Multimedia. 9657–9666

    2025

  35. [35]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  36. [36]

    In International conference on machine learning

    Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 8748–8763

  37. [37]

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

    2021

  38. [38]

    Replicate. 2026. Run AI with an API. https://replicate.com/

    2026

  39. [39]

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

    2022

  40. [40]

    Jinya Sakurai, Yuki Koyama, and Issei Sato. 2025. Fairt2i: Mitigating social bias in text-to-image generation via large language model-assisted detection and attribute rebalancing.arXiv preprint arXiv:2502.03826(2025)

    arXiv 2025

  41. [41]

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. 2023. Finetuning text-to-image diffusion models for fairness.arXiv preprint arXiv:2311.07604(2023)

    Pith/arXiv arXiv 2023

  42. [42]

    Yingdong Shi, Changming Li, Yifan Wang, Yongxiang Zhao, Anqi Pang, Sibei Yang, Jingyi Yu, and Kan Ren. 2025. Dissecting and mitigating diffusion bias via mechanistic interpretability. InProceedings of the Computer Vision and Pattern Recognition Conference. 8192–8202

    2025

  43. [43]

    Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. 2014. Deepface: Closing the gap to human-level performance in face verification. InProceedings of the IEEE conference on computer vision and pattern recognition. 1701–1708. 13 Chen Chen, Yuanmin Huang, Zhenfei Zhang, Mi Zhang, Xiaohan Zhang, Yun Xiong, Xiaoyu You, Min Yang

    2014

  44. [44]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805(2023)

    Pith/arXiv arXiv 2023

  45. [45]

    The Guardian. 2024. ‘We definitely messed up’: why did Google’s AI tool make offensive historical images? https://www.theguardian.com/technology/2024/ mar/08/we-definitely-messed-up-why-did-google-ai-tool-make-offensive- historical-images

    2024

  46. [46]

    Soobin Um and Jong Chul Ye. 2025. Minority-focused text-to-image generation via prompt optimization. InProceedings of the Computer Vision and Pattern Recognition Conference. 20926–20936

    2025

  47. [47]

    Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. 2025. Quantifying bias in text-to-image generative models.IEEE Transactions on Dependable and Secure Computing(2025)

    2025

  48. [48]

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

    Pith/arXiv arXiv 2024

  49. [49]

    Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022. Interpretability in the wild: a circuit for indirect object identifi- cation in gpt-2 small.arXiv preprint arXiv:2211.00593(2022)

    Pith/arXiv arXiv 2022

  50. [50]

    Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. 2024. Detecting, explain- ing, and mitigating memorization in diffusion models. InThe Twelfth International Conference on Learning Representations

    2024

  51. [51]

    Yunzhi Yao, Ningyu Zhang, Zekun Xi, Mengru Wang, Ziwen Xu, Shumin Deng, and Huajun Chen. 2024. Knowledge circuits in pretrained transformers.Advances in neural information processing systems37 (2024), 118571–118602

    2024

  52. [52]

    Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. 2024. Towards understanding the working mechanism of text-to-image diffusion model.Advances in Neural Information Processing Systems37 (2024), 55342–55369

    2024

  53. [53]

    Liu Yu, Jiajun Sun, Ping Kuang, Rui Zhou, Fan Zhou, and Zhikun Feng. 2025. Bimodal Debiasing for Text-to-Image Diffusion: Adaptive Guidance in Textual and Visual Spaces. InProceedings of the 33rd ACM International Conference on Multimedia. 11249–11258

    2025

  54. [54]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  55. [55]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  56. [56]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang

  57. [57]

    In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)

    Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers). 15–20. A Open Science We will publicly release the code and artifacts needed to evaluate the core contributions ...

    2018