Pith. sign in

REVIEW 4 major objections 5 minor 68 references

Multi-Group Proportional Representation for Text-to-Image Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read MPR measures worst-case representation gaps across intersectional groups in text-to-image models.

desk verdict A useful, honest application of a known metric to text-to-image fairness; the same noisy labeler feeds both measurement and training, and Proposition 2's proof is sketchy, but the package is worth refereeing. read the letter →

arxiv 2505.24023 v1 pith:LAYZFS44 submitted 2025-05-29 cs.CV cs.AI

classification cs.CVcs.AI
keywords text-to-imagegenerationrepresentationalharmintersectionalfairnessmulti-groupproportionalrepresentationdiffusionmodelsfine-tuningdemographicbiasmetric
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 paper introduces Multi-Group Proportional Representation (MPR) as a metric for measuring representational harm in text-to-image models. MPR is defined as the worst-case difference, over a chosen class of functions, between the average value of a function on generated images and on images from a reference population. The paper argues that this single definition covers both equal-representation goals and context-dependent historical-accuracy goals, and that tuning the function class controls how fine-grained the intersectional groups are. It also presents a fine-tuning algorithm that uses MPR as a training objective, and reports experiments showing reduced MPR for gender, age, and race while keeping CLIP image-quality scores stable.

What carries the argument

The metric itself: $\mathrm{MPR}(C,G_q,R)=\sup_{c\in C}|\mathbb{E}_{G_q}[c(X_g)]-\mathbb{E}_R[c(X_r)]|$, the worst-case gap between a generated distribution and a reference distribution over a bounded function class $C$. Choosing $C$ as bounded linear functions gives an explicit norm computation; choosing $C$ as depth-$\ell$ decision trees gives the maximum total variation between marginal distributions over any $\ell$-attribute intersection. These closed forms allow MPR to be computed and used in gradient-based fine-tuning, with buffers of past generated images and worst-case functions stabilizing the supremum, and CLIP/DINO regularization preserving image quality.

What would settle it

Generate the same image set for a fixed prompt, then label the images twice: once with the paper's linear classifiers on CLIP embeddings and once with human annotators for gender, age, and race. If MPR computed on human labels ranks the models differently from MPR on classifier labels, or if the fine-tuned model no longer beats the vanilla model on human-labeled MPR, the measurement and optimization claims collapse. A cheaper version: re-run the Table 4 comparison with race labels replaced by a second, independently trained race classifier and check whether the MPR ordering of methods is stable.

Watch

Extended reading notes

Core claim

The central claim is that MPR, originally proposed for retrieval, transfers to text-to-image generation as a theoretically grounded measure of proportional representation across arbitrary intersectional groups. With attribute vectors as inputs and decision trees or linear classifiers as the function class, the MPR supremum has closed forms tied to total variation and linear projections, making it computable and interpretable: the splits of the worst-case tree name the intersectional groups driving the bias. Used as a fine-tuning objective with buffered samples and regularization, MPR guides a diffusion model toward more balanced intersectional generation without a large drop in semantic fidelity.

Load-bearing premise

The load-bearing premise is that the automated classifiers used to label gender, age, and race on generated faces are accurate enough that MPR values and gradients reflect true demographic membership; the paper's own race classifier is 77% accurate, and the authors note these estimators may carry their own biases.

Editorial extensions

If this is right

  • MPR unifies earlier single-axis fairness metrics as special cases, so a single evaluation can replace separate gender, race, and age checks.
  • Decision-tree MPR identifies which intersectional group is most over- or under-represented, turning a harm score into an interpretable audit.
  • Contextual reference distributions let MPR reward historical accuracy, such as women-only ENIAC programmers, instead of always demanding uniform parity.
  • Fine-tuning with MPR reduces intersectional representational harm while maintaining CLIP scores, outperforming baselines that only address gender.
  • By trading function-class complexity for sample size, MPR scales to large attribute sets where enumerating all intersectional groups is impossible.

Reading between the lines

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

  • Our inference: because MPR with decision trees returns interpretable split attributes, the same measurement could double as an audit report that tells developers which intersectional group is most underrepresented, not just that bias exists.
  • Our inference: the buffer-based optimization recipe could generalize to other non-differentiable distributional metrics in generative models, since it converts the supremum into a finite set of cached worst-case functions.
  • Our inference: pairing MPR with official statistics from sources like Census or labor data could make 'proportional' representation operational for deployment, but it would also inherit whatever historical biases those references contain; the paper flags this tension but does not resolve it.
  • Our inference: Proposition 2 suggests MPR improvements should transfer to unseen prompts, so a cheap testable extension is to fine-tune on one set of occupations and evaluate on a held-out set; the paper's experiments average across professions but do not explicitly report held-out prompt generalization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Multi-Group Proportional Representation (MPR) for text-to-image models. MPR is defined in Definition 1 as the worst-case absolute difference, over a function class C, between the expectation of c under the generated distribution and under a reference distribution R. The authors give an empirical estimator (Eq. 2), generalization bounds (Propositions 1 and 2), closed-form expressions for bounded linear functions and decision trees (Propositions 3 and 4), and a fine-tuning algorithm (Algorithm 1) that optimizes the MPR objective with buffers and a CLIP/DINO regularization term. The experiments measure representational harm across professions, traits, and disabilities for several T2I models, and apply the fine-tuning method to Stable Diffusion v1.4, reporting improved intersectional representation relative to baseline debiasing methods while maintaining CLIP scores.

Significance. If the technical results are correct, MPR provides a flexible and principled framework that unifies existing fairness metrics for T2I models with IPM/MMD theory and multicalibration, while allowing context-dependent reference distributions and tunable function-class complexity. The closed-form computations for linear classifiers and decision trees, as well as the buffer-based fine-tuning scheme, are useful contributions, and the released code supports reproducibility. However, the central claims rest on the correctness of the generalization bounds and on the reliability of the automated attribute classifiers; both raise concerns that must be addressed before the measurement and mitigation conclusions can be accepted.

major comments (4)
  1. [Section 3.1, Eq. (4); Appendix B, Proof of Proposition 2] Proposition 2 as stated is not established by the supplied proof. The exponent in Eq. (4) is exp(-2(k+m)/B^2 * (epsilon/2 - 2*lambda)), which is linear in the margin, whereas the proof's own use of Proposition 1 would yield a squared term; Remark 2 in the appendix uses a squared term, confirming the inconsistency. Additionally, the proof applies a fixed-prompt deviation bound to each q_i and then asserts the same bound for the average without a union bound or uniform control of the Rademacher complexities, so the concentration over N prompts does not follow. The proof also contains indexing errors (e.g., 'n' in place of 'm' in the symmetrization step). This is load-bearing because Proposition 2 is cited as justification that MPR measured on finitely many prompts generalizes to the prompt distribution.
  2. [Section 4 and Appendix C; Algorithm 1] The empirical claims depend on automated attribute classifiers that are used both to measure MPR and as the training signal in Algorithm 1. The race classifier has only 77% test accuracy on FairFace, and the paper does not validate its behavior on synthetic images from Stable Diffusion, which are out-of-distribution for a CLIP linear probe. If the labeler makes systematic errors on generated images, the MPR reductions reported in Tables 1-4 and in the fine-tuning comparison could reflect exploitation of labeler bias rather than genuine demographic change. The authors acknowledge this risk in Appendix A, but no sensitivity analysis, confusion matrices, alternative labelers, or manually labeled subset validation is provided. I request such an analysis, or at minimum an explicit error model and a discussion of how label noise affects both the measurement and the optimization.
  3. [Table 4; Definition 1] The MPR values in Table 4 are inconsistent with the metric as defined. For decision-tree function classes, Proposition 4 implies MPR <= 1, yet Table 4 reports MPR(G)=1.18 and MPR(GAR)=1.89 for the Vanilla model. For a binary gender attribute with a balanced FairFace reference, even the linear function class of Proposition 3 (with one-hot encoding and ||w||<=1) would not allow MPR(G)=1.18: the maximum possible deviation with a balanced reference is at most sqrt(2)*0.5, about 0.71. The table does not specify which function class and which attribute encoding were used, nor whether the reported numbers are normalized. This makes the headline comparison in Table 4 difficult to interpret and potentially invalid. The authors must state the exact C used for each MPR column and reconcile the numerical range with Definition 1.
  4. [Appendix B, Proof of Proposition 1] The proof of Proposition 1 contains technical errors that prevent it from establishing the stated bound. The McDiarmid step bounds the martingale difference by B/min{m,k} in Eq. (11), but the exponent in the subsequent concentration inequality is written as exp(-2*epsilon^2*(m+k)/B^2), which would require a different range; the factor 1/min^2 is missing. The symmetrization argument also mixes the sample-size factors, with terms like 2*E sup (1/k) sum_{i=n+1}^{k+m} sigma*c(x_i) appearing instead of the reference-sample Rademacher average with 1/m. Since Proposition 1 is used to justify sample-size selection and to derive Proposition 2, the proof must be corrected before the theoretical claims can be accepted.
minor comments (5)
  1. [Section 5, Table 4] The caption states that results are averaged across seven professions, but only six are listed (chef, therapist, taxi driver, flight attendant, housekeeper, nurse); please correct the count or list all professions.
  2. [Section 3.2, Eq. (5)] The displayed formula for Proposition 3 is written as a fraction, but it should be read as the norm ||a^T X||; please clarify the notation and state explicitly that the supremum equals the Euclidean norm of X^T a.
  3. [Appendix D.2, Figure 4] The caption refers to the 'true MPR' as computed with 5,000 generated images; this is an approximation, and the caption should state that assumption explicitly to avoid overclaiming.
  4. [Appendix B, Eq. (20)] The index n is used without definition; the intended reference-sample size appears to be m, and the coefficient should be 1/m rather than 1/k for the second Rademacher term.
  5. [Appendix B, Remark 2, Eq. (24)] The Bernstein-style bound introduces sigma^2 and a second term exp(-2(k+m)(epsilon/4 - 2*lambda)^2/B^2), but the relationship between the two terms and the value of epsilon/4 versus epsilon/2 in Proposition 2 is not explained; please harmonize the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MPR metric is restated and its theoretical properties are proved in the paper; using the same attribute estimators in measurement and fine-tuning is a stated design choice, not a hidden reduction.

full rationale

The derivation chain is self-contained. Definition 1 states MPR as an IPM-type supremum over a function class, and the paper explicitly notes that MPR is also known as MMD and a special case of integral probability metrics. Propositions 1-4 are proved in the paper from Definition 1 using standard Rademacher/McDiarmid arguments and elementary linear-algebra and decision-tree optimizations; none of the proofs assumes the conclusion. For example, Proposition 3 reduces to the norm of the empirical mean difference, which follows directly from the supremum over a unit ball, and Proposition 4 follows from choosing optimal leaf signs in a decision tree. The citation to the earlier retrieval MPR paper is provenance, not load-bearing: the definition and proofs are reproduced here. The same reference distribution and the same CLIP-probe attribute classifiers feed both MPR evaluation and the fine-tuning objective in Algorithm 1; this is a deliberate optimization/evaluation loop, and the paper flags the classifier-bias risk in Appendix A: "These estimation methods may not only perpetuate harmful categorization practices but could also contain inherent biases themselves." That is a validity or robustness limitation, not a circular derivation: the claimed MPR reductions are reductions of the same quantity the method optimizes, which is what an optimization method should demonstrate, and the CLIP score check is an independent external benchmark. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the choice of C or R. Overall, no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim does not introduce new physical or conceptual entities. The load-bearing external inputs are the function class C, the reference distribution R, and the accuracy of the group classifiers, none of which are derived in the paper.

free parameters (2)
  • lambda (regularization strength) = 0.5
    Selected from grid [0.1, 0.5, 1, 5, 10, 50, 100] by preliminary 100-iteration runs; controls the trade-off between MPR reduction and CLIP/DINO alignment. Reported in Appendix C.
  • Buffer sizes B_MPR and B_C = 32
    Chosen to balance memory and optimization stability for Algorithm 1; no systematic tuning reported.
assumptions (3)
  • standard math Rademacher complexity and McDiarmid/Hoeffding inequalities give valid high-probability bounds for the MPR generalization gap.
    Invoked in Propositions 1 and 2 and their proofs in Appendix B.
  • ad hoc to paper Representational harm is adequately captured by the class of functions C (bounded linear or depth-limited decision trees) over the chosen demographic attributes.
    The paper assumes decision trees of depth up to 3 and linear functions are sufficient to identify the most over- or under-represented intersectional groups; this is a modeling choice made in Section 3.2 and used in all experiments.
  • domain assumption The reference distribution R (FairFace test set, or Google Images for the ENIAC case) is the appropriate target for proportional representation.
    Section 4 states that FairFace is assumed balanced for gender, age, and race; the ENIAC reference is constructed from Google Images. The paper acknowledges in Appendix A that determining what constitutes appropriate representation is inherently complex.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Group Proportional Representation for Text-to-Image Models." pith.science (2026). https://pith.science/paper/LAYZFS44

@misc{pith2026250524023,
  author       = {Pith},
  title        = {Pith review of: Multi-Group Proportional Representation for Text-to-Image Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAYZFS44}},
  note         = {Machine review of arXiv:2505.24023}
}
read the original abstract

Text-to-image (T2I) generative models can create vivid, realistic images from textual descriptions. As these models proliferate, they expose new concerns about their ability to represent diverse demographic groups, propagate stereotypes, and efface minority populations. Despite growing attention to the "safe" and "responsible" design of artificial intelligence (AI), there is no established methodology to systematically measure and control representational harms in image generation. This paper introduces a novel framework to measure the representation of intersectional groups in images generated by T2I models by applying the Multi-Group Proportional Representation (MPR) metric. MPR evaluates the worst-case deviation of representation statistics across given population groups in images produced by a generative model, allowing for flexible and context-specific measurements based on user requirements. We also develop an algorithm to optimize T2I models for this metric. Through experiments, we demonstrate that MPR can effectively measure representation statistics across multiple intersectional groups and, when used as a training objective, can guide models toward a more balanced generation across demographic groups while maintaining generation quality.

Figures

Figures reproduced from arXiv: 2505.24023 by the authors.

Figure 1
Figure 1. Overview of the Multi-Group Proportional Representation (MPR) framework for evaluating text-to-image generators. Given a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. MPRs according to C and group attributes when applying existing bsaeline methods to mitigate unfairness of SD v1.4 for “CEO”. DT represents decision trees, and the number following it indicates its depth. The text under DT refers to the group attributes used. harms related to traits and disabilities, which were qualita￾tively studied in the previous works [11, 42]. Experimental setup. For each concept under investig… view at source ↗
Figure 3
Figure 3. Left: A random sample of the top 100 Google Images [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Maximum deviation between the estimated MPRs and the true MPR depending on decision tree depth and the number of generated images. The number following “DT” indicates the depth of the decision tree, and the maximum deviation is calculated after 30 repetitions of the es…
Figure 5
Figure 5. Figure 5: Qualitative comparison of images generated for the prompt [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Continuation of Figure [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Heatmap of MPR standard deviations obtained by bootstrapping. Each value represents the MPR calculated using a linear classifier function class, considering gender, age, and race as group attributes. The observed decrease in standard deviation with increasing k and m a…
Figure 8
Figure 8. Figure 8: Images generated from the “CEO” prompt using the vanilla SD v1.4 model. The leftmost and second color bars for each image represent the estimated gender and race, respectively. Gender is indicated by blue for males and red for females. Race is represented by the follow…
Figure 9
Figure 9. Figure 9: Images generated from the “CEO” prompts using FairDiffusion. The color details follow the same scheme as [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Images generated from the “disability” prompts using the vanilla SD v1.4. The leftmost and second color bars for each image represent the estimated presence of a wheelchair and race, respectively. The presence of a wheelchair is indicated by blue for wheelchair users …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 56 canonical work pages

  1. [1]

    https://population.un.org/ wpp/

    United nations department of economic and social affairs: Population division. https://population.un.org/ wpp/. Accessed: 2024-11-14. 4

  2. [2]

    cdc.gov/SP

    Centers for disease control and prevention.https://dhds. cdc.gov/SP. Accessed Sep 11, 2024. 7

  3. [3]

    https://ec.europa.eu/eurostat/ web / population - demography / demography - population - stock - balance / database

    Eurostats data. https://ec.europa.eu/eurostat/ web / population - demography / demography - population - stock - balance / database. Ac- cessed: 2024-11-14. 4

  4. [4]

    https:// ilostat.ilo.org/

    International labour organization statistics. https:// ilostat.ilo.org/. Accessed: 2024-11-14. 4

  5. [5]

    https : / / www

    Oecd data. https : / / www . oecd . org / en / data . html. Accessed: 2024-11-14. 4

  6. [6]

    employed persons by detailed oc- cupation, sex, race, and hispanic or latino ethnicity

    Us bureau of labor statistics. employed persons by detailed oc- cupation, sex, race, and hispanic or latino ethnicity. https: //www.bls.gov/cps/cpsaat11.htm . Accessed: 2024-11-09. 4

  7. [7]

    https://data.worldbank

    World bank open data. https://data.worldbank. org/. Accessed: 2024-11-14. 4

  8. [8]

    Firefly, 2024

    Adobe. Firefly, 2024. Accessed Sep 7, 2024. https:// www.adobe.com/products/firefly/features/ text-to-image.html. 1

Show all 68 references
  1. [9]

    How well can text-to-image generative models under- stand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022

    Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models under- stand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022. 3, 5, 8

  2. [10]

    Pioneer programmer: Jean Jennings Bartik and the computer that changed the world

    Jean Jennings Bartik. Pioneer programmer: Jean Jennings Bartik and the computer that changed the world . Truman State University Press, 2013. 7

  3. [11]

    Easily acces- sible text-to-image generation amplifies demographic stereo- types 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 acces- sible text-to-image generation amplifies demographic stereo- types at large scale. In ACM Conference on Fa...

  4. [12]

    Inte- grating structured biological data by kernel maximum mean discrepancy

    Karsten M Borgwardt, Arthur Gretton, Malte J Rasch, Hans- Peter Kriegel, Bernhard Schölkopf, and Alex J Smola. Inte- grating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49–e57, 2006. 3

  5. [13]

    Con- centration inequalities: A nonasymptotic theory of indepen- dence

    Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Con- centration inequalities: A nonasymptotic theory of indepen- dence. 2013. 2

  6. [14]

    Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 7

  7. [15]

    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. In European Conference on Computer Vision, pages 429–446. Springer, 2024. 3

  8. [16]

    Dall-eval: Prob- ing the reasoning skills and social biases of text-to-image generation models

    Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Prob- ing the reasoning skills and social biases of text-to-image generation models. In European Conference on Computer Vision, pages 3043–3054, 2023. 1, 2, 3, 5

  9. [17]

    Fair sampling in diffusion models through switching mechanism

    Yujin Choi, Jinseong Park, Hoki Kim, Jaewook Lee, and Saerom Park. Fair sampling in diffusion models through switching mechanism. In AAAI Conference on Artificial In- telligence, pages 21995–22003, 2024. 3

  10. [18]

    Debiasing vision-language models via biased prompts

    Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070,

  11. [19]

    Openbias: Open-set bias detection in text-to-image generative models

    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. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 122...

  12. [20]

    Openbias: Open-set bias detection in text-to-image generative models

    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. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 122...

  13. [21]

    Diversity is not a one-way street: Pilot study on ethical inter- ventions for racial bias in text-to-image systems

    Kathleen C Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. Diversity is not a one-way street: Pilot study on ethical inter- ventions for racial bias in text-to-image systems. IEEE/CVF International Conference on Computer Vision, 2023. 1, 3

  14. [22]

    A friendly face: Do text-to-image systems rely on stereotypes when the input is under-specified? In The AAAI-23 Workshop on Creative AI Across Modalities, 2023

    Kathleen C Fraser, Isar Nejadgholi, and Svetlana Kiritchenko. A friendly face: Do text-to-image systems rely on stereotypes when the input is under-specified? In The AAAI-23 Workshop on Creative AI Across Modalities, 2023. 2

  15. [23]

    Fair diffusion: Instructing text-to-image generation models on fairness

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hin- tersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 1, 2, 3, 5, 7, 8

  16. [24]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. In IEEE/CVF Winter Conference on Ap- plications of Computer Vision, pages 5111–5120, 2024. 3, 5, 8

  17. [25]

    Introducing gemini: our largest and most capable ai model, 2023

    Google. Introducing gemini: our largest and most capable ai model, 2023. Accessed Sep 7, 2024. https://blog. google/technology/ai/google-gemini-ai/. 1

  18. [26]

    A kernel two-sample test

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bern- hard Schölkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723– 773, 2012. 3

  19. [27]

    Gender recognition or gender reductionism? the social implications of embedded gender recognition systems

    Foad Hamidi, Morgan Klaus Scheuerman, and Stacy M Bran- ham. Gender recognition or gender reductionism? the social implications of embedded gender recognition systems. In ACM Conference on Human Factors in Computing Systems, pages 1–13, 2018. 5

  20. [28]

    Multicalibration: Calibration for the (computationally-identifiable) masses

    Ursula Hébert-Johnson, Michael Kim, Omer Reingold, and Guy Rothblum. Multicalibration: Calibration for the (computationally-identifiable) masses. In International Con- ference on Machine Learning, pages 1939–1948, 2018. 3

  21. [29]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. 2021. 8

  22. [30]

    Fairface: Face at- tribute dataset for balanced race, gender, and age for bias measurement and mitigation

    Kimmo Karkkainen and Jungseock Joo. Fairface: Face at- tribute dataset for balanced race, gender, and age for bias measurement and mitigation. In IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 1548–1558,

  23. [31]

    De-stereotyping text-to-image models through prompt tuning

    Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. Workshop on Challenges in Deployable Generative AI at Inter- national Conference on Machine Learning, 2023. 2, 3

  24. [32]

    Multi- accuracy: Black-box post-processing for fairness in classifi- cation

    Michael P Kim, Amirata Ghorbani, and James Zou. Multi- accuracy: Black-box post-processing for fairness in classifi- cation. In AAAI/ACM Conference on AI, Ethics, and Society, pages 247–254, 2019. 3

  25. [33]

    Training unbiased diffusion models from biased dataset

    Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il-Chul Moon. Training unbiased diffusion models from biased dataset. International Conference on Learning Representations, 2024. 3

  26. [34]

    Davis E. King. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10:1755–1758, 2009. 6

  27. [35]

    Proving ground: The untold story of the six women who programmed the world’s first modern computer

    Kathy Kleiman. Proving ground: The untold story of the six women who programmed the world’s first modern computer. Hurst Publishers, 2022. 7

  28. [36]

    Playground v2

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights to- wards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 7

  29. [37]

    Fair text-to-image diffusion via fair mapping

    Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, and Di Wang. Fair text-to-image diffusion via fair mapping. arXiv preprint arXiv:2311.17695, 2023. 3

  30. [38]

    Controlling the fidelity and diversity of deep generative models via pseudo density

    Shuangqi Li, Chen Liu, Tong Zhang, Hieu Le, Sabine Süsstrunk, and Mathieu Salzmann. Controlling the fidelity and diversity of deep generative models via pseudo density. Transactions on Machine Learning Research, 2024. 3

  31. [39]

    Stable bias: Evaluating societal representa- tions in diffusion models

    Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representa- tions in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3

  32. [40]

    Bigbench: A unified bench- mark for social bias in text-to-image generative models based on multi-modal llm

    Hanjun Luo, Haoyu Huang, Ziye Deng, Xuecheng Liu, Ruizhe Chen, and Zuozhu Liu. Bigbench: A unified bench- mark for social bias in text-to-image generative models based on multi-modal llm. arXiv preprint arXiv:2407.15240, 2024. 1, 3

  33. [41]

    Latent consistency models: Synthesizing high- resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high- resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 7

  34. [42]

    they only care to show us the wheelchair

    Kelly Avery Mack, Rida Qadri, Remi Denton, Shaun K Kane, and Cynthia L Bennett. “they only care to show us the wheelchair”: disability representation in text-to-image ai mod- els. In CHI Conference on Human Factors in Computing Systems, pages 1–23, 2024. 1, 2, 4, 5, 7

  35. [43]

    On the method of bounded differences

    Colin McDiarmid et al. On the method of bounded differences. Surveys in combinatorics, 141(1):148–188, 1989. 2

  36. [44]

    Training diffusion models towards diverse image generation with reinforcement learning

    Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844–10853, 2024. 3

  37. [45]

    Midjourney.com, 2024

    Midjourney. Midjourney.com, 2024. Accessed Sep 7, 2024. https://www.midjourney.com/home. 1

  38. [46]

    On the epistemic limits of personalized prediction

    Lucas Monteiro Paes, Carol Long, Berk Ustun, and Flavio Calmon. On the epistemic limits of personalized prediction. Advances in Neural Information Processing Systems, 35:1979– 1991, 2022. 4

  39. [47]

    Integral probability metrics and their generat- ing classes of functions

    Alfred Müller. Integral probability metrics and their generat- ing classes of functions. Advances in applied probability, 29 (2):429–443, 1997. 3

  40. [48]

    Multi-group proportional representation in retrieval

    Alex Oesterling, Claudio Mayrink Verdun, Alexander Glynn, Carol Xuan Long, Lucas Monteiro Paes, Sajani Vithana, Mar- tina Cardone, and Flavio Calmon. Multi-group proportional representation in retrieval. In Advances in Neural Information Processing Systems, 2024. 2, 3

  41. [49]

    Editing implicit assumptions in text-to-image diffusion models

    Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision , pages 7053–7061, 2023. 2, 3

  42. [50]

    Factoring the matrix of domination: A critical review and reimagination of intersectionality in ai fairness

    Anaelia Ovalle, Arjun Subramonian, Vagrant Gautam, Gilbert Gee, and Kai-Wei Chang. Factoring the matrix of domination: A critical review and reimagination of intersectionality in ai fairness. In AAAI/ACM Conference on AI, Ethics, and Society, pages 496–511, 2023. 3

  43. [51]

    Bal- ancing act: Distribution-guided debiasing in diffusion models

    Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Bal- ancing act: Distribution-guided debiasing in diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6668–6678, 2024. 3

  44. [52]

    Würstchen: An efficient architecture for large-scale text-to-image diffusion models

    Pablo Pernias, Dominic Rampas, Mats L Richter, Christo- pher J Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637, 2023. 7

  45. [53]

    Computational optimal transport: With applications to data science

    Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355–607, 2019. 3

  46. [54]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    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. In International Conference on Learning Representations, 2024. 5

  47. [55]

    Gemini image generation got it wrong

    Prabhakar Raghavan. Gemini image generation got it wrong. we’ll do better. https : / / blog . google / products/gemini/gemini-image-generation- issue/. Accessed: 2024-11-09. 1

  48. [56]

    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. In International confer- ence on machine learning, pages 8821–8831. PMLR, 2021. 1

  49. [57]

    Google apologizes for ‘missing the mark’ after gemini generated racially diverse nazis

    Adi Robertson. Google apologizes for ‘missing the mark’ after gemini generated racially diverse nazis. https : / / www . theverge . com / 2024 / 2 / 21 / 24079371 / google- ai- gemini- generative- inaccurate- historical. Accessed: 2024-11-09. 1, 3

  50. [58]

    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. In IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 1, 5

  51. [59]

    Finetuning text-to-image dif- fusion models for fairness

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

  52. [60]

    Generalized people diversity: Learning a human perception-aligned diversity representa- tion for people images

    Hansa Srinivasan, Candice Schumann, Aradhana Sinha, David Madras, Gbolahan Oluwafemi Olanubi, Alex Beutel, Susanna Ricco, and Jilin Chen. Generalized people diversity: Learning a human perception-aligned diversity representa- tion for people images. In ACM Conference on Fairne...

  53. [61]

    On the empirical estimation of integral probability metrics

    Bharath K Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert RG Lanckriet. On the empirical estimation of integral probability metrics. Electronic Journal of Statistics, 6:1550–1599, 2012. 3

  54. [62]

    Stable diffusion v2.1 and dream- studio updates 7-dec 22, 2022

    Stabel Diffusion 2.1. Stable diffusion v2.1 and dream- studio updates 7-dec 22, 2022. Accessed Septem- ber 10, 2024. https : / / stability . ai / news / stablediffusion2-1-release7-dec-2022 . 5

  55. [63]

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

    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. 1

  56. [64]

    Concept algebra for (score-based) text-controlled generative models

    Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. Advances in Neural Information Processing Systems, 36, 2024. 1

  57. [65]

    Mist: Mitigating intersectional bias with disentangled cross- attention editing in text-to-image diffusion models

    Hidir Yesiltepe, Kiymet Akdemir, and Pinar Yanardag. Mist: Mitigating intersectional bias with disentangled cross- attention editing in text-to-image diffusion models. arXiv preprint arXiv:2403.19738, 2024. 1, 3

  58. [66]

    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. In IEEE/CVF Inter- national Conference on Computer Vision, pages 3969–3980,

  59. [67]

    a photo of concept

    Eric Zhao, De-An Huang, Hao Liu, Zhiding Yu, Anqi Liu, Olga Russakovsky, and Anima Anandkumar. Scaling fair learning to hundreds of intersectional groups. 2022. 3 Supplementary Material In this supplementary material, we provide detailed proofs, additional experimental results...

  60. [68]

    Male”, “Old

    to detect faces in the generated images and filtered to include only those images containing at least one detectable face. To estimate demographic attributes (gender, age, and race), we cropped the detected faces from the images. We applied our classifiers only to these croppe...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.