Pith. sign in

REVIEW 5 major objections 8 minor 59 references

Bias Analysis in Unconditional Image Generative Models

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Measured attribute bias shifts in unconditional image generation are governed by the density of the classifier's decision boundary, not by the attribute's true representational change.

desk verdict A careful but partly tautological study of classifier-based bias shift in unconditional generation; the boundary-density finding is real but needs a classifier validity check on generated images. read the letter →

arxiv 2506.09106 v1 pith:5URSDLLK submitted 2025-06-10 cs.CV cs.LG

classification cs.CVcs.LG
keywords unconditionalimagegenerationattributebiasshiftdecisionboundarydensityspectrum-basedattributesdiffusionmodelsGANsclassifier-predictedlabelsCelebA
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 investigates whether unconditional image generators shift attribute frequencies relative to the training distribution, and why some attributes appear more biased than others. Using diffusion models and a BigGAN trained on CelebA and DeepFashion, it labels training, validation, and generated images with the same attribute classifier and measures the frequency shift between generation and validation. The detected shifts are small on average, but spectrum-based attributes show much larger shifts than non-spectrum-based ones, and the paper traces this gap to the density of the classifier's pre-sigmoid logits at the decision boundary. When a boundary sits in a dense region, a small distribution shift moves many images across it; when it sits in a low-density gap, the same shift changes measured frequencies almost not at all. This matters because bias audits that rely on classifier labels can mistake the labeling geometry for the generator's behavior.

What carries the argument

The load-bearing object is the trained attribute classifier's pre-sigmoid logit distribution together with its decision boundary at zero. The paper proves in Appendix A that for a small translation $\delta$ of the distribution, the measured frequency shift is $\left|\int_{t-\delta}^{t} f(x)\,dx\right|$, so the density of $f$ near the threshold $t$ determines how much a fixed distribution shift changes the positive-class proportion. This single identity organizes the empirical results: it explains why the same generator can look nearly unbiased for Bangs and Eyeglasses but strongly shifted for Smiling and Young, and it supplies the 0.01 boundary-density cutoff used to partition attributes into spectrum-based and non-spectrum-based sets.

What would settle it

Human-label a sample of generated images for the same attributes and recompute the frequency shift against human-labeled validation data. The paper's account predicts that non-spectrum attributes stay near zero and spectrum-based attributes shift only if the underlying logit distribution genuinely moved; if human labels show a different pattern, the measured shifts are classifier artifacts.

Watch

Extended reading notes

Core claim

The paper's central claim is that attribute bias shift, defined as $B_{\mathrm{shift}}(C)=|P^{\mathrm{gen}}_{\mathrm{cls}}(C)-P^{\mathrm{val}}_{\mathrm{cls}}(C)|$, is controlled by the density of the classifier's pre-sigmoid logits at the decision threshold. For a distribution translated by $\delta$, the shift equals the integral of the density over the interval $[t-\delta,t]$, so a boundary in a dense region yields a large frequency change and a boundary in a sparse region yields a negligible one. Empirically, spectrum-based attributes such as Smiling, Young, No Dress, and Solid have boundary densities above 0.01 on the validation distribution, while non-spectrum-based attributes such as Bangs, Eyeglasses, and Pleated have boundaries in low-density gaps; the former show average bias shifts of 3.25–4.73%, while the latter stay below about 1%. The paper also shows that BigGAN and smaller diffusion models produce larger bias shifts than the large diffusion model despite comparable image-generation metrics, and that bias metrics do not track those metrics during training.

Load-bearing premise

The framework assumes the attribute classifier remains valid on generated images and that the training-to-generation change is a small approximately translative shift of its logits, so the Appendix A integral applies; if generated images are out of distribution for the classifier, the measured frequency shifts could be classifier artifacts rather than true attribute changes.

Editorial extensions

If this is right

  • Any classifier-based estimate of attribute bias shift should report where the decision boundary sits in the logit distribution; without that, differences across attributes and models cannot be interpreted.
  • Picking a generative model checkpoint by FID, KID, or FLD does not minimize bias shift, so bias must be tracked as its own objective during model development.
  • Attributes whose classifier boundary lies in a low-density region will almost always show small measured shifts, even if the generator's true representation of that attribute changes.
  • BigGAN and smaller diffusion models show larger attribute bias shifts than a large diffusion model at similar image quality, consistent with the paper's attribution to reduced variation and mode collapse.

Reading between the lines

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

  • If the boundary-density mechanism is the whole story, then 'spectrum-based' is not a fixed property of an attribute but a property of the classifier's confidence landscape; the same attribute could be categorized differently under a different classifier, which the paper's own appendix shows for Bangs, 5-o'clock Shadow, and Blond Hair.
  • A direct testable consequence is that deliberately relocating the decision boundary, for example by thresholding at a high-density point or adding an abstain option, should shrink or enlarge measured shifts without any change in the generator.
  • The same density analysis could be carried into text-to-image and guided generation, where prompt conditioning and classifier-free guidance may shift the logit distribution; the framework would isolate how much of reported bias amplification is due to guidance versus the unconditional generator.
  • Bias audits that report only aggregate frequency shifts may systematically understate representational issues for attributes with cleanly separable, low-density boundaries; the paper's focus on the classifier suggests combining frequency shifts with boundary-density diagnostics and human labels.
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 / 8 minor

Summary. This paper studies attribute bias shift in unconditional image generation. The authors train diffusion and GAN models on CelebA and DeepFashion, label the generated images with a fine-tuned attribute classifier, and define Bshift (Eq. 3) as the absolute difference between classifier-predicted attribute frequencies on the validation and generation sets. They report that overall shifts are small, that attributes whose classifier decision boundary falls in high-density regions (termed 'spectrum-based') show larger shifts, and that BigGAN and small diffusion models show larger shifts than large diffusion models. Appendix A proves that under a translative shift of the classifier logits, Bshift equals the probability mass of the density near the decision boundary, which the paper uses to explain the observed pattern.

Significance. If the findings hold, the paper makes a useful methodological contribution: classifier-based bias-shift measurements in unconditional generation can be strongly influenced by the location of the classifier's decision boundary, and attributes with ambiguous or spectrum-like labels may show larger measured shifts irrespective of the generator's true behavior. The paper's strengths include a clean from-scratch training setup on two datasets, two classifier backbones, three training seeds for selected analyses, and a detailed appendix with per-attribute logit distributions. The significance is currently conditional, however, because the headline numbers rely on an attribute classifier that is never validated on generated images, and the main explanatory claim reduces to a mathematical identity under an untested translative-shift assumption.

major comments (5)
  1. [Section 3.2, Eq. (3); Section 4.1] The Bshift values that drive all headline results are computed with an attribute classifier that is validated only on real validation images (Tables 6-8) and never on the 10,000 generated images per checkpoint. This matters because Eq. (3) defines Bshift directly as a difference of classifier-predicted frequencies, and because the paper itself describes generated images as distributionally different (Section 4.5: BigGAN and small-diffusion outputs are 'washed out' with 'fewer variations and less details'). Using the same classifier for both sets does not cancel classifier error unless the error is identical on both distributions, which is not established. The Appendix A translative-shift model assumes the classifier produces the same meaningful logits on generated images, which is exactly what is unverified. Please report classifier accuracy or calibration on a labeled subset of generated images (for example via human annotation) or otherwise show that the boundary-density pattern persists when classifier confidence on generated images is controlled.
  2. [Section 4.4; Appendix A] The central explanation—that boundary density determines bias shift—is a mathematical identity under the translative-shift model: Eq. (7) shows that Bshift is the integral of the density over the interval [t−δ, t], so a high-density boundary mechanically yields larger shifts for a given δ. The empirical content therefore rests entirely on whether the training-to-generation logit shifts are approximately translative and small. The paper only states that the shifts 'generally have low earth mover's distance' without reporting any EMD values or testing the translative assumption. Please report per-attribute EMD between validation and generation logit distributions, and compare the measured Bshift against the Eq. (7) prediction using the estimated δ, including a discussion of attributes where the prediction fails.
  3. [Section 4.4; Tables 1 and 5] The 0.01 decision-boundary density threshold is post hoc and unstable across classifier backbones: switching from ResNeXt to Swin Transformer moves 5-o-Clock Shadow, Bangs, and Blond Hair from the non-spectrum to the spectrum category, and the text notes that their densities are 'very close to the criteria boundary (0.01)'. Since the spectrum/non-spectrum distinction is the paper's main predictive claim, the categorization should be shown to be robust. Please report the actual boundary densities for all attributes, plot ABS_spectrum and ABS_non_spectrum as a function of the threshold, and quantify how many attributes change category over a reasonable threshold range. The label 'spectrum-based' also conflates a semantic property with a classifier property; the paper should state explicitly that the category is operational.
  4. [Section 4.3; Fig. 4] The abstract and Section 4.3 claim that the detected attribute shifts are 'small', but no baseline or confidence interval is provided. The stated values (0.71% and 0.98% for non-spectrum; 3.25% and 4.73% for spectrum) are compared only visually to the sampling-error curve in Fig. 4c, which itself has no error bars and is only for CelebA. Please report a null distribution (for example, ABS between two independent validation samples, or between training and validation classifier-predicted frequencies) and add seed-based confidence bands to the main ABS curves using the three seeds already present in Fig. 8.
  5. [Section 4.2, Tables 6-8; Section 4.4] Several attributes in the main analysis have very poor classifier performance even on the validation set: for DeepFashion, square neckline and leather have AUPR=1.03 and faux AUPR=17.70, and for CelebA, Big Lips AUPR=34.41, Pointy Nose AUPR=55.88, and Oval Face AUPR=50.47. For such attributes, the pre-sigmoid logit distribution and its density at the decision boundary largely reflect classifier uncertainty rather than the attribute's distribution in the data, so including them in ABS and in the spectrum/non-spectrum split may bias the conclusions. Please report the main results after excluding attributes below an AUPR threshold (for example 0.6) or otherwise demonstrate that the boundary-density pattern is not driven by low-confidence attributes.
minor comments (8)
  1. [Abstract; Section 1] The phrase 'remain disentangled' in the abstract and introduction appears to mean 'remain entangled' or 'are not yet disentangled'; please reword for clarity.
  2. [Table 1; Appendix B] The attributes in Table 1 do not sum to the 40 attributes listed in Table 2; the excluded attributes (Blurry, Wearing Earrings, Wearing Necklace, Wearing Necktie) should be explicitly listed in the main text or table caption.
  3. [Fig. 4] The main ABS curves in Figs. 4a and 4b have no error bars or seed variation, even though Section 4.6 reports three random seeds; please add confidence bands or state explicitly which seed is shown.
  4. [Appendix E] The terms 'subjective attributes', 'spectrum-based attributes', and 'decision-boundary density' are used almost interchangeably; please clarify that the operational categorization is based on the density at the classifier boundary, while the human-agreement argument in Appendix E is only a post hoc justification.
  5. [Section 4.5] The statement that BigGAN's larger ABS 'may be attributed to the well-known issue of mode collapse' is speculative; please either provide a diversity metric such as recall or reword the claim to be explicitly tentative.
  6. [Fig. 1] The caption labels 'in +ve samples' are unclear, and the red and blue areas mentioned in the text are not defined in the caption; please add a clear description of the shaded regions.
  7. [Section 3.2] In Eq. (3), P_gen_cls and P_val_cls are introduced only in the surrounding text; please define them explicitly in the equation or its immediate caption.
  8. [Appendix B] There is a typo: 'Validadtion set' should be 'Validation set'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the boundary-density sensitivity is a derived consequence of the bias-shift metric, not an assumed input or a fitted parameter renamed as a prediction.

full rationale

Eq. (3) defines Bshift as a difference of classifier-predicted attribute proportions, and Appendix A rewrites this quantity for a translative distribution shift as the probability mass in a small neighborhood of the decision boundary. This is a mathematical consequence of the definition plus an explicit translative-shift assumption, not a hidden restatement of the conclusion: the integral formula does not follow from Eq. (3) alone, and the paper does not use Bshift to define the density-based categorization. Section 4.4 categorizes attributes by validation-set boundary density and then measures ABS separately for the two categories, so the reported correlation is empirically testable and could have failed (e.g., if generator shifts were strongly non-translative with large EMD, the Appendix A formula would not apply). No parameter is fitted to bias-shift data and then presented as a prediction; the 0.01 density threshold is justified by an external human-agreement study (Wu et al., Appendix E), not by the measured shifts. There are no load-bearing self-citations: the methodological references (ADM training [39], FLD [45], Wu et al. [59]) are external to the author list. The concern that the attribute classifier is not validated on generated images is a validity/robustness issue, not circularity, because the paper explicitly defines its metric on classifier-predicted labels and acknowledges classifier error as a limitation. Overall, no step in the derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the classifier-based labeling protocol and the translation-shift approximation, both of which are assumptions rather than measured guarantees. The 0.01 threshold is the sole fitted free parameter. No new physical or conceptual entities are introduced beyond the spectrum/non-spectrum taxonomy.

free parameters (1)
  • Decision-boundary density threshold for spectrum/non-spectrum categorization = 0.01
    Attributes with validation-set logits density at the decision boundary greater than 0.01 are categorized as spectrum-based (Section 4.4). The threshold is chosen post hoc and justified by alignment with human labeling consistency (Appendix E), but it is not derived from first principles and shifts attribute categorizations across classifier backbones (Table 1 versus Table 5).
assumptions (3)
  • domain assumption The attribute classifier remains accurate on generated images, so classifier-predicted frequency shifts reflect true attribute shifts.
    The framework uses a classifier trained on real images to label generated images (Section 3.2). If generated images are out-of-distribution for the classifier, systematic mislabeling could create spurious shifts; the paper only checks classifier accuracy on the validation set, not on generated images.
  • domain assumption The distribution shift between training and generation in the classifier's logit space has low EMD and is approximately a translation, so the Appendix A integral formula applies.
    Appendix A proves Bshift equals the probability mass in the delta neighborhood of the decision boundary under a translation shift. Section 4.4 asserts empirically that EMD is low, but no EMD values are reported.
  • domain assumption Binary attribute annotations are adequate for measuring bias shifts; spectrum-based attributes can be binarized without loss.
    The paper uses binary labels for all attributes (Section 3.1) and acknowledges in Limitations that binary annotations are limiting for spectrum-based attributes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias Analysis in Unconditional Image Generative Models." pith.science (2026). https://pith.science/paper/5URSDLLK

@misc{pith2026250609106,
  author       = {Pith},
  title        = {Pith review of: Bias Analysis in Unconditional Image Generative Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5URSDLLK}},
  note         = {Machine review of arXiv:2506.09106}
}
read the original abstract

The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on this topic, the mechanisms by which bias emerges - especially in unconditional generation - remain disentangled. We define the bias of an attribute as the difference between the probability of its presence in the observed distribution and its expected proportion in an ideal reference distribution. In our analysis, we train a set of unconditional image generative models and adopt a commonly used bias evaluation framework to study bias shift between training and generated distributions. Our experiments reveal that the detected attribute shifts are small. We find that the attribute shifts are sensitive to the attribute classifier used to label generated images in the evaluation framework, particularly when its decision boundaries fall in high-density regions. Our empirical analysis indicates that this classifier sensitivity is often observed in attributes values that lie on a spectrum, as opposed to exhibiting a binary nature. This highlights the need for more representative labeling practices, understanding the shortcomings through greater scrutiny of evaluation frameworks, and recognizing the socially complex nature of attributes when evaluating bias.

Figures

Figures reproduced from arXiv: 2506.09106 by the authors.

Figure 1
Figure 1. Illustrations depicting bias shift. The plots represent the distributions of samples with respect to the likelihood of an attribute (solid for training data, dashed for generation). The decision boundary (brown) binarizes the likelihood into positive and negative classes. In each subfigure, the generation distribution is translated from the training. Bias shift is the difference between red and blue areas. When the … view at source ↗
Figure 2
Figure 2. Bias shift evaluation framework. Unconditional generative models are trained on the training set. The pre-trained classifier is fine-tuned on the training set and validated on the validation set using ground truth labels and is then used to classify training, validation, and generation sets. The bias shift evaluation metrics are calculated based on the classifier-predicted labels. Bias-related Attribute Label Predic… view at source ↗
Figure 3
Figure 3. Evaluation metrics for image generation throughout training. In 3a and 3b, FID, KID, and FLD values converge to small values showing the good quality of generated images and good coverage of modes of the training distribution. In 3c, the positive or slightly negative generalization gaps indicate that the trained models do not have severe memorization issues. 100 200 300 400 500 600 steps (K) 2.00 4.00 6.00 8.00 10.0… view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: Average attribute bias shift (ABS) for CelebA and DeepFashion. For both datasets, shown in Figs. 4a and 4b, ABS over spectrum-based attributes show a much larger attribute bias shift than non-spectrum-based ones. Fig. 4c presents that the error coming from sampling of …
Figure 5
Figure 5. Figure 5: CelebA classifier’s pre-sigmoid logits distributions of selected spectrum-based and non-spectrum-based attributes. The decision boundary for spectrum-based attributes (Fig. 5a, 5b) always falls in a high-density region, while that for non-spectrum-based attributes (Fig…
Figure 6
Figure 6. Figure 6: DeepFashion classifier’s pre-sigmoid logits distributions of selected spectrum-based and non-spectrum-based attribute. The decision boundary for spectrum-based attributes (Fig. 6a, 6b) falls in a high-density region, while that for non-spectrum-based attributes (Fig. 6…
Figure 7
Figure 7. Figure 7: FLD and ABS of different generative models on CelebA. The small diffusion model has slightly worse image generation quality but much larger ABS for both spectrum-based and non￾spectrum-based attributes compared to the large diffusion model. BigGAN has a similar FLD as …
Figure 8
Figure 8. Figure 8: Probabilities of selected spectrum-based and non-spectrum-based attributes during training on the CelebA dataset with three different random seeds. The probabilities of spectrum￾based attributes (Fig. 8a and 8b) present a gap between generation and validation data whil…
Figure 9
Figure 9. Figure 9: Probabilities of selected spectrum-based and non-spectrum-based attribute during training on the DeepFashion dataset. The probabilities of spectrum-based attributes (Fig. 9a, 9b) present a gap between generation and validation data while non-spectrum-based one (Fig. 9c…
Figure 10
Figure 10. Figure 10: ABS and image generation metrics using different inference methods and inference steps [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: ABS for CelebA dataset using different classifiers. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: The pre-sigmoid logits distribution of each attribute in CelebA (ResNeXt-based classi [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: The pre-sigmoid logits distribution of each attribute in CelebA (SwinTransformer-based [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: The pre-sigmoid logits distribution of each attribute in DeepFashion. [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Probabilities of attributes for DeepFashion dataset during training. Please note that it [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]
Figure 16
Figure 16. Figure 16: The probabilities of attributes in CelebA during training (ResNext-based classifier). [PITH_FULL_IMAGE:figures/full_fig_p025_16.png]
Figure 17
Figure 17. Figure 17: The probabilities of attributes in CelebA during training (SwinTransformer-based classi 26 [PITH_FULL_IMAGE:figures/full_fig_p026_17.png]
Figure 18
Figure 18. Figure 18: Image samples from large diffusion model generations on CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p027_18.png]
Figure 19
Figure 19. Figure 19: Image samples from the small diffusion model trained on CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p028_19.png]
Figure 20
Figure 20. Figure 20: Image samples from the BigGAN model trained on CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p028_20.png]
Figure 21
Figure 21. Figure 21: Image samples from the tiny diffusion model trained on CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p029_21.png]
Figure 22
Figure 22. Figure 22: Image samples from the large diffusion model trained on DeepFashion dataset. [PITH_FULL_IMAGE:figures/full_fig_p029_22.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 51 canonical work pages

  1. [1]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Tim- oth´ee Lacroix, Baptiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, Aur ´elien Ro- driguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023

  2. [2]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023

  3. [3]

    Hierarchical text-conditional image generation with CLIP latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022

  4. [4]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674–10685, 2022

  5. [5]

    Scaling rectified flow transformers for high-resolution im- age synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution im- age synthesis. In Forty-first International Conference on Machine Learning. OpenReview....

  6. [6]

    Audiogen: Textually guided audio generation

    Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre D ´efossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. In The Eleventh International Conference on Learning Representations. OpenReview.net, 2023

  7. [7]

    Gritsenko, William Chan, Mohammad Norouzi, and David J

    Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems, 2022

  8. [8]

    Make- a-video: Text-to-video generation without text-video data

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make- a-video: Text-to-video generation without text-video data. In The Eleventh International Con- ference on Learning Representations. OpenReview.net, 2023

Show all 59 references
  1. [9]

    Which ai image generator is the most biased?, 2023

    Matt Growcoot. Which ai image generator is the most biased?, 2023

  2. [10]

    These fake images reveal how ai amplifies our worst stereotypes, 2023

    Nitasha Tiku, Kevin Schaul, and Szu Yu Chen. These fake images reveal how ai amplifies our worst stereotypes, 2023

  3. [11]

    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 Proceedings of the 38th International Conference on Machine Learning, volume 139, pages 8821–8831. PMLR, 2021

  4. [12]

    DALL-EV AL: probing the reasoning skills and social biases of text-to-image generation models

    Jaemin Cho, Abhay Zala, and Mohit Bansal. DALL-EV AL: probing the reasoning skills and social biases of text-to-image generation models. In IEEE/CVF International Conference on Computer Vision, pages 3020–3031. IEEE, 2023

  5. [13]

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

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ...

  6. [14]

    Stable bias: Eval- uating societal representations in diffusion models

    Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Eval- uating societal representations in diffusion models. In Advances in Neural Information Pro- cessing Systems 36: Annual Conference on Neural Information Processing Systems, 2023

  7. [15]

    Auditing and instructing text-to-image generation mod- els on fairness

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Auditing and instructing text-to-image generation mod- els on fairness. AI and Ethics, pages 1–21, 2024. 10

  8. [16]

    The bias amplification paradox in text- to-image generation

    Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text- to-image generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pape...

  9. [17]

    Analyzing bias in diffusion-based face generation mod- els

    Malsha V Perera and Vishal M Patel. Analyzing bias in diffusion-based face generation mod- els. In 2023 IEEE International Joint Conference on Biometrics (IJCB) , pages 1–10. IEEE, 2023

  10. [18]

    LAION-5B: an open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmar- czyk, and Jenia Jitsev. LAI...

  11. [19]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Infor- mation Processing Systems, 2020

  12. [20]

    Generative modeling by estimating gradients of the data dis- tribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data dis- tribution. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems, pages 11895–11907, 2019

  13. [21]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Pro- cessing Sys...

  14. [22]

    Large scale GAN training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, 2019

  15. [23]

    Classifier-free diffusion guidance

    Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022

  16. [24]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynk ¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37:52996–53021, 2024

  17. [25]

    A brief review on algorithmic fairness

    Xiaomeng Wang, Yishi Zhang, and Ruilin Zhu. A brief review on algorithmic fairness. Man- agement System Engineering, 1, 2022

  18. [26]

    Finetuning text-to-image diffusion models for fairness

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Confer- ence on Learning Representations, 2024

  19. [27]

    On fairness of unified multimodal large language model for image generation

    Ming Liu, Hao Chen, Jindong Wang, Liwen Wang, Bhiksha Raj Ramakrishnan, and Wensheng Zhang. On fairness of unified multimodal large language model for image generation. ArXiv, abs/2502.03429, 2025

  20. [28]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015

  21. [29]

    Deepfashion: Powering robust clothes recognition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), June 2016

  22. [30]

    mindall-e on conceptual captions

    Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek. mindall-e on conceptual captions. https://github.com/kakaobrain/minDALL-E, 2021

  23. [31]

    Karlo-v1.0.alpha on coyo-100m and cc15m

    Donghoon Lee, Jiseob Kim, Jisu Choi, Jongmin Kim, Minwoo Byeon, Woonhyuk Baek, and Saehoon Kim. Karlo-v1.0.alpha on coyo-100m and cc15m. https://github.com/ kakaobrain/karlo, 2022. 11

  24. [32]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Pro- ceedings o...

  25. [33]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Inter- national Conference on Machine Learning, volume 202, pages 19730–19742. PMLR, 2023

  26. [34]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  27. [35]

    Hudson, Ehsan Adeli, Russ B

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Bryn- jolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen ...

  28. [36]

    CLIP the bias: How useful is balancing data in multimodal learning? In The Twelfth International Conference on Learning Representations

    Ibrahim Alabdulmohsin, Xiao Wang, Andreas Peter Steiner, Priya Goyal, Alexander D’Amour, and Xiaohua Zhai. CLIP the bias: How useful is balancing data in multimodal learning? In The Twelfth International Conference on Learning Representations. OpenReview.net, 2024

  29. [37]

    Fairness definitions explained

    Sahil Verma and Julia Sass Rubin. Fairness definitions explained. 2018 IEEE/ACM Interna- tional Workshop on Software Fairness (FairWare), pages 1–7, 2018

  30. [38]

    Tiago Palma Pagano, Rafael Bessa Loureiro, Fernanda Vitoria Nascimento Lisboa, Ro- drigo Matos Peixoto, Guilherme A. S. Guimar ˜aes, Gustavo Oliveira Ramos Cruz, Maira M. Araujo, Lucas Lisboa dos Santos, Marco A. S. Cruz, Ewerton Lopes Silva de Oliveira, Ingrid Winkler, and Er...

  31. [39]

    Diffusion models beat gans on image synthe- sis

    Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthe- sis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, pages 8780–8794, 2021

  32. [40]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi, editors, Medical Image Computing and Computer-Assisted Interven- tion - MICCA...

  33. [41]

    Girshick, Piotr Doll ´ar, Zhuowen Tu, and Kaiming He

    Saining Xie, Ross B. Girshick, Piotr Doll ´ar, Zhuowen Tu, and Kaiming He. Aggregated resid- ual transformations for deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition,, pages 5987–5995, 2017

  34. [42]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Bain- ing Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992–10002, 2021

  35. [43]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochre- iter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Infor- mation Pr...

  36. [44]

    Sutherland, Michael Arbel, and Arthur Gretton

    Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD gans. In 6th International Conference on Learning Representations . OpenReview.net, 2018

  37. [45]

    Feature likelihood score: Evaluating the generalization of generative models using samples

    Marco Jiralerspong, Avishek Joey Bose, Ian Gemp, Chongli Qin, Yoram Bachrach, and Gau- thier Gidel. Feature likelihood score: Evaluating the generalization of generative models using samples. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural ...

  38. [46]

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khali- dov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michae...

  39. [47]

    Yossi Rubner, Carlo Tomasi, and Leonidas J. Guibas. A metric for distributions with applica- tions to image databases. In Proceedings of the Sixth International Conference on Computer Vision, pages 59–66. IEEE Computer Society, 1998

  40. [48]

    Wasserstein generative adversarial net- works

    Mart ´ın Arjovsky, Soumith Chintala, and L´eon Bottou. Wasserstein generative adversarial net- works. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 214–223. PMLR, 2017

  41. [49]

    Big data’s disparate impact

    Solon Barocas and Andrew D Selbst. Big data’s disparate impact. California Law Review , 104(3):671–732, 2016

  42. [50]

    Gender shades: Intersectional accuracy disparities in commercial gender classification

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

  43. [51]

    Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial ai products.AAAI/ACM Conference on AI, Ethics, and Society, 2019

    Inioluwa Deborah Raji and Joy Buolamwini. Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial ai products.AAAI/ACM Conference on AI, Ethics, and Society, 2019

  44. [52]

    Auditing al- gorithms: Research methods for detecting discrimination on internet platforms

    Christian Sandvig, Kevin Hamilton, Karrie Karahalios, and Cedric Langbort. Auditing al- gorithms: Research methods for detecting discrimination on internet platforms. Data and discrimination: Collected essays, 2014

  45. [53]

    Paul, and Jed R

    Morgan Klaus Scheuerman, Jacob M. Paul, and Jed R. Brubaker. How computers see gender: An evaluation of gender classification in commercial facial analysis services. Proc. ACM Hum.-Comput. Interact., 3(CSCW), November 2019

  46. [54]

    Datasheets for datasets, 2021

    Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wal- lach, Hal Daum´e III, and Kate Crawford. Datasheets for datasets, 2021

  47. [55]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the 38th International Conference on Machine Learning, , volume 139 of Proceedings of Machine Learning Research, pages 8162–8171. PMLR, 2021

  48. [56]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations. OpenReview.net, 2021

  49. [57]

    Men also like shopping: Reducing gender bias amplification using corpus-level constraints

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2979– 2989, 2017

  50. [58]

    A systematic study of bias amplification

    Melissa Hall, Laurens van der Maaten, Laura Gustafson, and Aaron Adcock. A systematic study of bias amplification. CoRR, abs/2201.11706, 2022

  51. [59]

    Consistency and accuracy of celeba attribute values

    Haiyu Wu, Grace Bezold, Manuel G ¨unther, Terrance Boult, Michael C King, and Kevin W Bowyer. Consistency and accuracy of celeba attribute values. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3258–3266, 2023. 13 A Proof of bias shi...

Pith tools

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