Pith. sign in

REVIEW 5 major objections 4 minor 44 references

EmergencyBias: Bias in Text-to-Image Models under Emergency Scenarios

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read EmergencyBias claims that text-to-image models skew both who appears in crisis scenes and who is shown helping, and that controlling the prompt does not erase the behavioral gap.

desk verdict Valuable new framing for T2I bias evaluation, but the headline numbers are not yet established because the pipeline silently drops images and lacks uncertainty quantification. read the letter →

arxiv 2608.00598 v1 pith:U2TXKEGI submitted 2026-08-01 cs.MM cs.CY

classification cs.MMcs.CY
keywords text-to-imagegenerationemergencyscenariosdemographicbiasbehavioralgenderdisparityevaluationmitigationsoft-tokencalibration
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 EmergencyBias, a two-part bias construct for text-to-image generation in emergency scenarios: demographic bias (who appears) and behavioral bias (who acts, helps, or stays passive). It evaluates seven leading models on six emergency scenarios with blank and demographically controlled prompts, totaling 4,200 emergency images plus matched non-emergency controls. The results show that models over-represent men and middle-aged, medium-skinned people in both the person-in-crisis and nearest-bystander roles, and that emergency scenes produce stronger demographic skew than matched ordinary scenes. Under controlled prompts, helping behavior remains tied to demographics, with gender showing the largest disparities, especially in physical intervention. The paper also contributes ActionAlign, a soft-token prompt-embedding calibration that reduces these behavioral gaps more than an explicit ethical-prompt baseline while preserving image quality.

What carries the argument

The central object is the EmergencyBias evaluation pipeline, built on two scores and one intervention. Demographic Bias Score (DBS) measures deviation from a uniform demographic distribution for each attribute; Help Bias Score (HBS) measures imbalance in helping propensity; Jensen-Shannon divergence (JSD) measures divergence between action distributions. The person-in-crisis and nearest-bystander roles are annotated along gender, age, and skin tone, and bystander behavior is decomposed into five channels based on atomic-action principles. ActionAlign is the mechanism that carries the mitigation claim: a single learnable soft token appended to the scenario prompt, trained on a behavior-balanc

What would settle it

Count, for each model and prompt, how many of the 10 generated images are excluded because no nearest bystander can be identified, then check whether the exclusion rate varies by the bystander's gender, age, or skin tone. If excluded images skew toward a particular group, recomputing DBS, HBS, and JSD while including those cases with an 'unidentifiable' category should shrink or invert the reported disparities.

Watch

Extended reading notes

Core claim

The central claim is that text-to-image models do not merely mirror representational stereotypes in static portraits; in high-stakes emergency scenarios they also assign behavioral roles unequally. The paper argues that emergency scenarios are a revealing testbed because they force the model to decide who is at risk, who intervenes, and how. Using blank prompts, the authors find strong demographic skew in crisis-person and bystander roles. Using controlled prompts that pin down the bystander's gender, age, or appearance, they find that helping propensity and fine-grained action distributions still vary by group, with the widest and most consistent gap between male and female bystanders. The

Load-bearing premise

The analysis assumes that every generated image has an identifiable person-in-crisis and nearest bystander, or that any images failing this test are dropped at random; if the dropped images are not random, the reported bias scores do not describe the models' true output distributions.

Editorial extensions

If this is right

  • If EmergencyBias is correct, bias audits that only count who appears in generated images are incomplete; they must also measure who is shown acting, helping, and remaining passive.
  • Because emergency scenes show larger DBS than matched non-emergency scenes, high-stakes contexts should be included in standard T2I fairness evaluations.
  • Controlled prompting improves demographic coverage but does not remove behavioral bias; explicit demographic specification still leaves group-based differences in helping propensity and action composition.
  • The largest and most model-consistent gender gap appears in physical intervention, so mitigating bias requires targeting action-level scripts, not just balancing character counts.
  • ActionAlign's larger HBS and JSD reductions than EntiGen, with less image-quality loss, indicate that embedding-level soft-token calibration is a practical alternative to explicit ethical prompting.

Reading between the lines

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

  • The paper's uniform-reference definition treats a perfectly balanced demographic distribution as the neutral baseline, but real emergency populations are not uniform; if actual victim and bystander demographics are skewed, DBS conflates model bias with real-world base rates.
  • The ActionAlign recipe—learn a soft token from behavior-balanced reference data—could apply beyond emergencies to any domain where generated images script social roles, such as medical, legal, or security imagery.
  • Since HBS and JSD do not move together across models, a single aggregate bias metric would be misleading; future audits should report both helping propensity and action-composition divergence separately.
  • If images without an identifiable bystander are dropped non-randomly—for example, if darker-skinned bystanders are harder to annotate—the reported skin-tone disparities could be an artifact; a re-analysis that tracks exclusion rates per demographic group would settle this.
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

5 major / 4 minor

Summary. The paper introduces EmergencyBias, a two-part notion of bias in text-to-image generation under emergency scenarios: demographic bias (who is depicted as victim/bystander across gender, age, skin tone) and behavioral bias (how the nearest bystander is depicted as helping or not). The authors construct a unified prompt template, generate images with seven T2I models across six emergency scenarios, annotate the person in crisis and nearest bystander via a multimodal-LLM pipeline with human verification, and quantify bias with DBS, HBS, and JSD. They report that, under blank prompts, demographic distributions are heavily skewed (e.g., 85.67% male crisis persons, 7.86% older crisis persons); under controlled prompts, behavioral differences persist, especially by gender; and they propose ActionAlign, a learned soft-token method that reduces HBS and JSD more than an EntiGen prompt baseline on two open-source models.

Significance. If the measurement were sound, the paper would make a useful contribution: it broadens T2I bias evaluation from static representation to action-level role allocation in a high-stakes domain, and it proposes a lightweight mitigation that preserves image quality. The scenario framework, the annotation schema, and the comparison across seven contemporary models are valuable assets. However, the empirical claims currently rest on three load-bearing problems—unaccounted exclusion of images without an annotatable bystander, lack of any uncertainty quantification or significance testing on tiny per-prompt samples, and a confounded skin-tone manipulation—so the reported magnitudes cannot yet be taken as reliable properties of the models.

major comments (5)
  1. [§3.4 and Appendix B] The manuscript never specifies how generated images that lack a clearly identifiable nearest bystander are treated. Appendix B explicitly states that 'some outputs do not contain a clearly identifiable nearest bystander, while others provide a candidate bystander whose facial or bodily cues are too ambiguous for reliable demographic annotation.' Since every aggregate in Tables 2–5 requires both a person in crisis and a nearest bystander, silently dropping such images would make DBS/HBS/JSD computations conditional on a non-random subset of each model's output. Exclusion rates plausibly vary by model, scenario, and demographic attribute, so the reported 85.67% male share and all model-wise scores would not describe the models' unconditional output. The paper must state the treatment (e.g., exclusion counts, reweighting, or imputation) and show that results are robust to the missing-bystan
  2. [§3.3 vs. Appendix A] There is a direct contradiction about sample sizes. Section 3.3 says 'For each prompt, we generate 10 images with each of the seven models, resulting in 4,200 emergency images.' Appendix A says 'Unless otherwise noted, each prompt generated one image in the remote setup and 10 images in the local setup.' If remote models generated only one image per prompt, the total emergency image count would be 5×60 + 2×600 = 1,500, not 4,200. This discrepancy affects every denominator and percentage in Tables 2–5. Please reconcile the two statements and report per-model and per-condition counts explicitly.
  3. [§4.2, Tables 2–5] All bias metrics are point estimates from small samples—roughly 10 images per prompt per model, and only six blank prompts can plausibly be used for the RQ1 results—yet the text describes results as 'clear' and 'significant' with no confidence intervals, standard errors, or significance tests. For example, Table 3 reports model-wise DBS values based on blank prompts only; a single gender DBS of 44.74 vs. 9.32 could easily arise from sampling noise with n≈60 per model. The same holds for Table 4 help rates and JSD values. The paper should provide CIs or bootstrap intervals and, where claims of group differences are made, appropriate tests (or temper the language to 'observed differences' without 'significant').
  4. [§4.2.2 and Appendix C.3] The skin-tone behavioral analysis is confounded. Controlled prompts for skin tone use region-based appearance descriptors such as 'East Asian nearby person' (Appendix C.3), not skin-tone descriptors. The final evaluation maps these to annotated skin-tone labels, but the annotations cannot separate skin tone from region/cultural appearance cues. Therefore Table 4's skin-tone HBS and JSD values (e.g., dark skin tone 66.85% help rate) may reflect regional or ethnic presentation rather than skin tone per se. This should be acknowledged as a confound, or the prompts should vary skin-tone descriptors directly, or the claim should be restricted to 'region-based appearance cues' rather than 'skin tone.'
  5. [§5.1.1 and Table 6] ActionAlign inherits the missing-bystander selection problem: the training set retains only samples with a nearest bystander present and identified as male or female, and the evaluation pipeline uses the same annotation schema. The reported HBS/JSD reductions (e.g., 78.14% on FLUX.1-dev) are point estimates with no CIs, based on 50 generated images per setting. Since the EntiGen comparison and the ActionAlign condition could have different exclusion rates, the comparison could be biased. Please report exclusion counts for each condition and provide uncertainty estimates for the improvements.
minor comments (4)
  1. [Abstract / §1] The abstract states that T2I models 'exhibit clear demographic bias' and that behavioral bias is 'systematically associated' with demographics. Given the statistical issues listed above, the claims should be phrased as observed associations in the studied sample, not as established population-level facts.
  2. [Table 1] The annotation agreement table reports averages over behavioral channels but does not state how the 400-image human verification set was sampled, whether it covers all models/scenarios, or how disagreements were resolved. Please add a short description.
  3. [Appendix A] The statement that 'remote setup' generated one image per prompt while 'local setup' generated ten is surprising and should be cross-checked with the main text; if it is a typo, it should be corrected.
  4. [§5.2] Table 6 reports ΔHBS and ΔJSD in percentages but does not show the absolute HBS/JSD values or raw counts. Absolute values would help interpret whether a 15.90% reduction on Qwen-Image-2512 is meaningful in magnitude.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DBS/HBS are explicit measurement definitions, and ActionAlign is a trained intervention evaluated on fresh generations.

full rationale

The paper's central results are empirical measurements made with explicitly defined metrics. DBS and HBS are introduced as operational definitions of distributional deviation from a uniform reference (Section 3.1.1: 'Following this definition, we adopt the same perspective in this work'; Section 3.5.1: DBS = 1/n Σ |p_i − 1/n|). The finding that T2I models exhibit demographic bias is not circular: it asserts that measured generation distributions deviate from uniform, which is a contingent empirical fact. If models produced uniform distributions, DBS would be zero; the reported nonzero values come from annotated outputs, not from the definition itself. Similarly, HBS is a descriptive inequality index over observed help rates, not a fitted parameter renamed as a prediction. ActionAlign is the only step that could resemble a fitted-input-called-prediction pattern: it learns a soft token on a behavior-balanced reference set constructed from annotated data (Section 5.1.1). However, the reported mitigation results (Table 6) are computed on freshly generated images ('we generate 50 images under each setting using the same generation parameters and annotate all generated images with the same pipeline'), so the HBS/JSD reductions are empirical transfer outcomes, not algebraic consequences of the training set construction. If evaluation had been performed on the training reference set itself, that would be circular; the paper does not do that. No load-bearing self-citations are present: references to prior work are external benchmarks or standard methodology (e.g., skin-tone scales, AVA action principles) and no uniqueness theorem or author-imported ansatz is invoked. Appendix B's acknowledged failure cases (images lacking an identifiable nearest bystander) raise a legitimate selection-threat concern for interpreting Tables 2–5, but they do not make any reported quantity equal to its own input by construction. The derivation chain is therefore self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim that emergency scenes exhibit demographic and behavioral bias depends on several choices the paper makes upstream: uniform as the neutral reference, scenario selection as representative, MLLM annotation as accurate and group-consistent, region descriptors as clean skin-tone proxies, and the assumption that all images yield identifiable roles. The most fragile of these are the skin-tone proxy (which can confound behavior with the region cue) and the uniform reference (which is a fairness prior, not an empirical baseline).

free parameters (4)
  • Skin tone binning thresholds = 1-3 light, 4-6 medium, 7-10 dark
    Hand-chosen grouping of the 10-point scale (Section 3.4) directly determines the skin-tone DBS and behavioral comparisons.
  • Help indicator threshold = at least one explicit intervention
    Binary helping label defined in Section 3.4; changing the threshold changes HBS values.
  • Number of images per prompt = 10
    Sampling design in Section 3.3; the per-group sample sizes affect the stability of all DBS/HBS estimates.
  • ActionAlign soft token tau = learned
    A learnable embedding trained on the authors' annotated dataset (Section 5.1.2); the reduction in HBS/JSD depends on this fitted token.
assumptions (6)
  • domain assumption Uniform demographic distribution is the neutral reference for bias (DBS/HBS compare empirical distributions to 1/n).
    Section 3.1.1 adopts the common assumption that unspecified demographic attributes should be uniformly distributed, which is a fairness prior, not a fact about the world.
  • domain assumption The six selected scenarios (drowning, subway fall, street fall, apartment fire, car accident, flood) are representative of emergencies and preserve clear person-in-crisis/bystander roles.
    Section 3.2 justifies scenario selection via WHO injury categories, but the representativeness claim is assumed.
  • domain assumption MLLM majority-vote annotations are accurate and demographically consistent proxies for human judgment.
    Section 3.4 relies on GPT-5.5, Qwen3.6-Plus, and Gemini 2.5 Flash with a 400-image human check; agreement is 88.2% on average but posture has a 9.19% gender gap.
  • ad hoc to paper Region-based appearance prompts (e.g., 'East Asian') elicit skin-tone variation without independently altering behavior.
    Appendix C.3 uses region descriptors as proxies for skin tone. If the descriptor itself affects the generated action distribution, the skin-tone behavioral results are confounded.
  • domain assumption Every generated image contains an identifiable person in crisis and nearest bystander, or exclusions are random.
    Section 3.4 assumes these roles are identifiable, but Appendix B shows failure cases without a clear bystander, and the main text does not state how such images are treated.
  • domain assumption CLIP-IQA is a valid measure of image quality for comparing debiasing methods.
    Section 5.2 uses relative CLIP-IQA changes as the quality metric.
invented entities (1)
  • ActionAlign soft token
    purpose: A learnable prompt embedding appended to scenario prompts to steer behavioral action distributions toward parity across gender groups.
    The token is trained on the authors' own annotated dataset and evaluated on the same two models; it has no falsifiable handle outside the paper, such as a predicted effect in independent data or fixed deployment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EmergencyBias: Bias in Text-to-Image Models under Emergency Scenarios." pith.science (2026). https://pith.science/paper/U2TXKEGI

@misc{pith2026260800598,
  author       = {Pith},
  title        = {Pith review of: EmergencyBias: Bias in Text-to-Image Models under Emergency Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2TXKEGI}},
  note         = {Machine review of arXiv:2608.00598}
}
read the original abstract

Bias in Text-to-Image (T2I) generation has become an important problem in multimedia content creation and communication. However, existing studies have primarily focused on relatively static and explicit forms of bias, such as disparities in the representation of gender, race, and geo-cultural attributes. Less attention has been paid to behavioral bias in how different groups are portrayed acting, reacting, and occupying social roles. Emergency scenarios provide a revealing setting for studying such bias because they require models to depict not only who is present, but also who is at risk, who intervenes, and how responsibility is allocated. In this paper, we define EmergencyBias, a form of bias in T2I generation under emergency scenarios that includes both demographic bias and behavioral bias. We construct an evaluation framework to systematically study EmergencyBias across seven leading T2I models, six representative emergency scenarios, and three demographic dimensions. Our experimental results reveal three main findings. First, under blank prompts without demographic specification, T2I models exhibit clear demographic bias in emergency scenarios, reflected in the distributions of portrayed individuals across gender, age, and skin tone. Second, under controlled prompts, behavioral bias in emergency responses remains systematically associated with demographic variation, with particularly pronounced disparities along gender and substantial differences across models. Third, we introduce ActionAlign, a lightweight prompt-embedding calibration method that outperforms a representative prompt-based baseline in reducing behavioral disparities while largely preserving image quality. Overall, our work identifies emergency scenarios as an important setting for bias evaluation in T2I models and offers a practical direction toward fairer visual generation in socially consequential contexts.

Figures

Figures reproduced from arXiv: 2608.00598 by the authors.

Figure 1
Figure 1. Illustration of the EmergencyBias studied in this work. Second, such scenarios are especially likely to expose implicit bias, because the model must not only depict people but also implicitly assign behavioral roles, such as who helps, who hesitates, and who becomes the victim. This shifts the focus from static representation to action-level social portrayal. Third, these biases may have broader downstream consequen… view at source ↗
Figure 2
Figure 2. Visualization of the evaluation framework used in EmergencyBias. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Representative examples of the fine-grained behavior channels in our hierarchical bystander behavior [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of the ActionAlign pipeline. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: A few failure cases in emergency generations. Some outputs lack a clear nearest bystander or provide a nearest bystander without discernible facial details. In some cases, the model also misinterprets the scene, for example by placing a person on the railway tracks or …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages

  1. [1]

    Reem Abbas and Todd Miller. Exploring communication inefficiencies in disaster response: Perspectives of emergency managers and health professionals.International Journal of Disaster Risk Reduction, 120:105393, 2025

  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? In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, Abu Dhabi, United Arab Emirates,...

  3. [3]

    Inspecting the geographical representativeness of images from text-to-image models

    Abhipsa Basu, R Venkatesh Babu, and Danish Pruthi. Inspecting the geographical representativeness of images from text-to-image models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5136–5147, 2023

  4. [4]

    Renaissance: A survey into ai text-to-image generation in the era of large model.IEEE transactions on pattern analysis and machine intelligence, 47(3):2212–2231, 2024

    Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. Renaissance: A survey into ai text-to-image generation in the era of large model.IEEE transactions on pattern analysis and machine intelligence, 47(3):2212–2231, 2024

  5. [5]

    Black Forest Labs. Flux 2. https://bfl.ai/models/flux-2, 2026. Official model page. Accessed: 2026-03- 22

  6. [6]

    Improving subject-driven image synthesis with subject-agnostic guidance

    Kelvin CK Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733–6742, 2024

  7. [7]

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

    Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to- image generation models. InProceedings of the IEEE/CVF international conference on computer vision, pages 3043–3054, 2023

  8. [8]

    Bystander intervention in emergencies: diffusion of responsibility.Journal of personality and social psychology, 8(4p1):377, 1968

    John M Darley and Bibb Latané. Bystander intervention in emergencies: diffusion of responsibility.Journal of personality and social psychology, 8(4p1):377, 1968

Show all 44 references
  1. [9]

    Oasis uncovers: High-quality t2i models, same old stereotypes.arXiv preprint arXiv:2501.00962, 2025

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

  2. [10]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

  3. [11]

    Ever Josue Figueroa. Casting heroes and victims of disaster events: Representations of race and gender in hurricane harvey front page news images.Critical Studies in Media Communication, 39(5):455–471, 2022

  4. [12]

    The bystander-effect: a meta-analytic review on bystander intervention in dangerous and non-dangerous emergencies.Psychological bulletin, 137(4):517, 2011

    Peter Fischer, Joachim I Krueger, Tobias Greitemeyer, Claudia V ogrincic, Andreas Kastenmüller, Dieter Frey, Moritz Heene, Magdalena Wicher, and Martina Kainbacher. The bystander-effect: a meta-analytic review on bystander intervention in dangerous and non-dangerous emergencie...

  5. [13]

    An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

  6. [14]

    Effects of visual framing in multimodal media environments: A systematic review of studies between 1979 and 2023.Journalism & Mass Communication Quarterly, 102(3):796–823, 2025

    Stephanie Geise and Yi Xu. Effects of visual framing in multimodal media environments: A systematic review of studies between 1979 and 2023.Journalism & Mass Communication Quarterly, 102(3):796–823, 2025

  7. [15]

    Why are emergency responses ineffective and inefficient? lessons learnt from past events.Journal of Loss Prevention in the Process Industries, 90:105335, 2024

    Fakhradin Ghasemi, Omid Kalatpour, Vahid Salehi, and Leila Omidi. Why are emergency responses ineffective and inefficient? lessons learnt from past events.Journal of Loss Prevention in the Process Industries, 90:105335, 2024

  8. [16]

    Gemini image – nano banana

    Google DeepMind. Gemini image – nano banana. https://deepmind.google/models/gemini-image/,

  9. [17]

    Ava: A video dataset of spatio-temporally localized atomic visual actions

    Chunhui Gu, Chen Sun, David A Ross, Carl V ondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijaya- narasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. InProceedings of the IEEE confe...

  10. [18]

    Facet: Fairness in computer vision evaluation benchmark

    Laura Gustafson, Chloe Rolland, Nikhila Ravi, Quentin Duval, Aaron Adcock, Cheng-Yang Fu, Melissa Hall, and Candace Ross. Facet: Fairness in computer vision evaluation benchmark. InProceedings of the IEEE/CVF international conference on computer vision, pages 20370–20382, 2023...

  11. [19]

    Towards geographic inclusion in the evaluation of text-to-image models

    Melissa Hall, Samuel J Bell, Candace Ross, Adina Williams, Michal Drozdzal, and Adriana Romero Soriano. Towards geographic inclusion in the evaluation of text-to-image models. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 585–601, 2024

  12. [20]

    Aitti: Learning adaptive inclusive token for text-to-image generation.International Journal of Computer Vision, 134(3):108, 2026

    Xinyu Hou, Xiaoming Li, and Chen Change Loy. Aitti: Learning adaptive inclusive token for text-to-image generation.International Journal of Computer Vision, 134(3):108, 2026

  13. [21]

    Implicit bias injection attacks against text-to-image diffusion models

    Huayang Huang, Xiangye Jin, Jiaxu Miao, and Yu Wu. Implicit bias injection attacks against text-to-image diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 28779–28789, 2025

  14. [22]

    Bafis: Dataset+ framework to assess occupational bias and human preference in modern text-to-image models

    Thomas Klassert, Adrian Ulges, and Biying Fu. Bafis: Dataset+ framework to assess occupational bias and human preference in modern text-to-image models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2168–2177, 2026

  15. [23]

    Divergence measures based on the shannon entropy.IEEE Transactions on Information theory, 37(1):145–151, 2002

    Jianhua Lin. Divergence measures based on the shannon entropy.IEEE Transactions on Information theory, 37(1):145–151, 2002

  16. [24]

    Do existing testing tools really uncover gender bias in text-to-image models? InProceedings of the 33rd ACM International Conference on Multimedia, pages 11687–11696, 2025

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

  17. [25]

    Social biases through the text-to-image generation lens

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

  18. [26]

    Gpt image 1.5

    OpenAI. Gpt image 1.5. https://openai.com/index/new-chatgpt-images-is-here/ , 2025. Official model page. Accessed: 2026-03-28

  19. [27]

    A clearer picture: The contribution of visuals and text to framing effects.Journal of communication, 65(6):997–1017, 2015

    Thomas E Powell, Hajo G Boomgaarden, Knut De Swert, and Claes H De Vreese. A clearer picture: The contribution of visuals and text to framing effects.Journal of communication, 65(6):997–1017, 2015

  20. [28]

    Emergency care research as a global health priority: key scientific opportunities and challenges.BMJ global health, 4(Suppl 6), 2019

    Junaid Razzak, Blythe Beecroft, Jeremy Brown, Stephen Hargarten, and Nalini Anand. Emergency care research as a global health priority: key scientific opportunities and challenges.BMJ global health, 4(Suppl 6), 2019

  21. [29]

    Generative adversarial text to image synthesis

    Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. InInternational conference on machine learning, pages 1060–1069. Pmlr, 2016

  22. [30]

    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

  23. [31]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information ...

  24. [32]

    Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

    Team Seedream, Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xiaoxia Hou, Weilin Huang, Yixuan Huang, et al. Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

  25. [33]

    Exposing blindspots: Cultural bias evaluation in generative image models.arXiv preprint arXiv:2510.20042, 2025

    Huichan Seo, Sieun Choi, Minki Hong, Yi Zhou, Junseo Kim, Lukman Ismaila, Naome Etori, Mehul Agarwal, Zhixuan Liu, Jihie Kim, et al. Exposing blindspots: Cultural bias evaluation in generative image models.arXiv preprint arXiv:2510.20042, 2025

  26. [34]

    Fairrag: Fair human generation via fair retrieval augmentation

    Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11996–12005, 2024

  27. [35]

    Decision support for emergency situations.Information Systems and E-Business Management, 6(3):295–316, 2008

    Bartel Van de Walle and Murray Turoff. Decision support for emergency situations.Information Systems and E-Business Management, 6(3):295–316, 2008

  28. [36]

    The male ceo and the female assistant: Evaluation and mitigation of gender biases in text-to-image generation of dual subjects

    Yixin Wan and Kai-Wei Chang. The male ceo and the female assistant: Evaluation and mitigation of gender biases in text-to-image generation of dual subjects. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages...

  29. [37]

    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

  30. [38]

    Moesd: Mixture of experts stable diffusion to mitigate gender bias.arXiv preprint arXiv:2407.11002, 2024

    Guorun Wang and Lucia Specia. Moesd: Mixture of experts stable diffusion to mitigate gender bias.arXiv preprint arXiv:2407.11002, 2024. 12 EmergencyBias: Bias in Text-to-Image Models under Emergency ScenariosA PREPRINT

  31. [39]

    Injuries and violence

    World Health Organization. Injuries and violence. https://www.who.int/news-room/fact-sheets/ detail/injuries-and-violence, 2024. Accessed: 2026-04-01

  32. [40]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

  33. [41]

    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. In Proceedings of the AAAI/ACM conference on AI, ethics, and society, volume 7, pages 1648–1659, 2024

  34. [42]

    Attngan: Fine-grained text to image generation with attentional generative adversarial networks

    Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1...

  35. [43]

    Text-to-image diffusion models in generative ai: A survey.arXiv preprint arXiv:2303.07909, 2023

    Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, In So Kweon, and Junmo Kim. Text-to-image diffusion models in generative ai: A survey.arXiv preprint arXiv:2303.07909, 2023. A Model Access and Inference Setup For reproducibility, we summarize the access type, deployment mode,...

  36. [2025]

    Accessed: 2026-03-22

    Official model page. Accessed: 2026-03-22

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.