pith. sign in

arxiv: 2605.28036 · v1 · pith:UIY6ATHEnew · submitted 2026-05-27 · 💻 cs.CV · cs.LG

Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales

Myeongsoo Kim , Eunji Kim , Minwoo Chae , Sangwoo Mo This is my paper

Pith reviewed 2026-06-29 13:08 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords group fairnessdiffusion modelsguidance scaleclassifier guidanceclassifier-free guidancedemographic paritybias decomposition
0
0 comments X

The pith

StayFair maintains group fairness in diffusion models at every guidance scale by correcting guidance bias separately.

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

Diffusion models adjust a guidance scale to trade prompt match against output variety. Existing fairness methods fix bias only at one fixed scale, so raising or lowering the scale reintroduces unfairness. The paper splits total bias into model bias and guidance bias, then shows the guidance component grows steadily with scale and dominates at the high values users often prefer. It derives a condition from extending demographic parity that keeps the target group ratio fixed across scales, and implements this by equalizing classifier outputs or shifting the null prompt embedding. The result is a method that works on top of any existing fair model and leaves image quality unchanged.

Core claim

Extending Strong Demographic Parity to the guidance step produces a condition under which the target distribution keeps its group ratio for any guidance scale; StayFair satisfies this condition by equalizing the classifier output distributions across groups in classifier guidance and by applying a prompt-dependent offset to the null embedding in classifier-free guidance.

What carries the argument

The condition derived from extending Strong Demographic Parity to guidance, which ensures the target distribution retains its group ratio across scales and is enforced by modifying only the guidance step.

If this is right

  • Fairness metrics become independent of the chosen guidance scale.
  • StayFair can be added to any previously debiased diffusion model without retraining.
  • Image quality and prompt alignment stay the same as in the original model.
  • The approach works for both class-conditional and text-to-image generation tasks.

Where Pith is reading between the lines

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

  • Users of image generators could change the guidance scale to control style or fidelity without fairness shifting as a side effect.
  • The same separation of guidance bias could be tested in other sampling methods that use classifier or null conditioning.
  • Combining StayFair with model-level debiasing might produce fairness guarantees that hold under distribution shift at inference time.

Load-bearing premise

Guidance bias grows monotonically with the guidance scale and eventually dominates high-guidance regimes.

What would settle it

Generate images at a range of guidance scales, compute group fairness metrics before and after applying StayFair, and check whether the metrics remain constant with scale.

Figures

Figures reproduced from arXiv: 2605.28036 by Eunji Kim, Minwoo Chae, Myeongsoo Kim, Sangwoo Mo.

Figure 1
Figure 1. Figure 1: StayFair preserves group fairness as users vary the guidance scale. Users choose the guidance scale to control how strongly generation follows the prompt. Thus, group fairness should remain stable regardless of users’ scale choice. However, prior work [4, 10] typically optimizes or evaluates fairness at a fixed scale, failing to maintain group fairness under different user selected scales. StayFair mitigat… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of StayFair. In (a) classifier guidance, StayFair debiases the classifier fϕ by adding the fairness constraint LWDP . It matches samples between groups using a joint cost on noise level and fϕ value, aligning the groupwise fϕ distributions. In (b) classifier-free guidance, StayFair replaces the fixed null prompt ∅ with an adaptive null prompt ∅y by steering its embedding along the attribute direct… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of guidance scale under classifier guidance. (a,b) Female ratio curves across guidance scales; dashed: unguided (w=0). (c) Qualitative comparison across guidance scales. As w increases, StayFair preserves minority groups while CG shifts toward the majority. Because the model bias under prompt y is correlated with the textual bias encoded in E(y), we use the latter as a proxy for α ⋆ (y) [55]. Implem… view at source ↗
Figure 4
Figure 4. Figure 4: Per-occupation attribute-ratio shifts across guidance scales. Each dot is an occupation– prompt pair, plotted by female ratio at low (x) and high (y) guidance scales and colored by optimal α. Gray (α=0) denotes CFG; distance from y=x reflects bias amplification (red) or mitigation (blue). Optimal α stays closer to the diagonal than CFG. uniformity, and report its range (max − min across guidance scales w).… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples for CFG and StayFair on SD3. StayFair retains the annotated minority attribute across guidance scales, whereas CFG drifts to the majority attribute. 2.5 5.0 7.5 10.0 12.5 Guidance Scale 0.0 0.2 0.4 0.6 0.8 1.0 Female Ratio -10 -5 -2.5 0 5 10 (a) Journalist (SD1.5) 2.5 5.0 7.5 10.0 12.5 Guidance Scale 0.0 0.2 0.4 0.6 0.8 1.0 Female Ratio -10 -5 -2.5 0 5 10 (b) Fashion Designer (FT) α = … view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study of StayFair. Female-ratio curves across guidance scales for different α values on (a) vanilla SD1.5 and (b) debiased SD1.5 by FT [10]. Black lines denote CFG, and yellow stars mark StayFair-selected α. (c) Samples at w=7.5, with selected α=5.0 highlighted in blue. range across guidance scales w, averaged over occupations and prompts. Image quality is measured by CLIP score [63], aesthetic sc… view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of bias amplification phenomenon. (a) For the doctor occupation, neutral prompts exhibit different amplification trends depending on their female ratio at low guidance scale. (b) The α = 0 plot shows CFG’s asymmetric amplification across gender groups. Neutralizing the female-biased unconditional term in guidance makes the trends more balanced (α = −5). the shift arises from visually plausible att… view at source ↗
Figure 8
Figure 8. Figure 8: Bias of FD and WG across the guidance scale. Female ratio curves across guidance scales and their deviation ∆ from the target ratio for (a) FD and (b) WG. Attribute-based methods exhibit weak attribute alignment [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Bias of FD+StayFair and WG+StayFair across the guidance scale. Female ratio curves across guidance scales on doctor (a, b) and librarian (c, d). Mitigation with StayFair. To apply StayFair in this setting, we select ∅y to improve alignment with the assigned attribute. Accordingly, we adjust the null prompt ∅y by shifting α in Eq. (10) toward the direction opposite to the assigned attribute with fixed scale… view at source ↗
Figure 10
Figure 10. Figure 10 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Asymmetric bias amplification under standard CFG on SD3 and SD1.5. Each dot is an occupation–prompt pair, plotted by its female ratio at low (x) and high (y) guidance scales. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Female ratio under increasing guidance scale w on the class-conditional model (Blond Hair). Increasing w raises the female ratio for CG and CG-RW, whereas StayFair keeps it close to the unguided baseline (w=0) across guidance scales. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Per-occupation bias range across vanilla and debiased models. Bias range (spread of the female ratio across guidance scales) for each of the 33 occupations, comparing CFG (blue) and StayFair (red), on vanilla models (a) SD1.5 and (b) SD3, and debiased models (c) FT and (d) UCE. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Per-occupation female ratio curves across guidance scales. Selected occupations per model showing CFG/vanilla (blue) and StayFair with optimal α ⋆ (red); the dashed line marks 0.5. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: and [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative examples showing the effect of varying α on SD3. Sweeping α shifts the null prompt ∅y from female-biased (α>0) to male-biased (α<0), and the generated samples shift in the opposite direction. The selected α used by StayFair is highlighted (red). Sample 1 Sample 2 w=0 w=1 w=3 w=5 w=9 w=0 w=1 w=3 w=5 w=9 CG CG-RW CG+StayFair Sample 3 Sample 4 w=0 w=1 w=3 w=5 w=9 w=0 w=1 w=3 w=5 w=9 CG CG-RW CG+S… view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative examples on class-conditional generation (condition: Blond Hair). Four samples on CelebA across guidance scales for CG, debiased classifier (CG-RW), and CG+StayFair. CG and CG-RW shift toward the majority gender as w grows, while CG+StayFair preserves the unguided attribute. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative examples on class-conditional generation (condition: Smiling). Four samples on CelebA across guidance scales for CG, debiased classifier (CG-RW), and CG+StayFair. CG and CG-RW shift toward the majority gender as w grows, while CG+StayFair preserves the unguided attribute. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗
read the original abstract

Diffusion models steer conditional generation with a tunable guidance scale to trade off prompt alignment and diversity. However, existing debiasing techniques are optimized for a single scale, degrading fairness when users adjust this parameter. We trace this behavior to a previously overlooked source by decomposing total bias into two components: a model bias and a guidance bias. While prior work primarily targets the former, we show that the guidance bias grows monotonically with the guidance scale, eventually dominating the high-guidance regimes users prefer. To address this, we extend Strong Demographic Parity to guidance and derive a condition under which the target distribution retains its group ratio across guidance scales. We propose StayFair, which leverages this condition to design fair guidance algorithms in both regimes. For classifier guidance, it equalizes the classifier's output distributions across groups; for classifier-free guidance, it shifts the null embedding by a prompt-dependent offset. Because StayFair modifies only the guidance step, it is orthogonal to model debiasing and can be layered onto existing fair diffusion models to extend their fairness across guidance scales. Across class-conditional and text-to-image generation, StayFair decouples fairness from the guidance scale without sacrificing image quality.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper decomposes bias in conditional diffusion models into model bias and guidance bias, shows that the latter grows monotonically with the guidance scale and dominates at high scales, extends Strong Demographic Parity to the guidance step, derives a condition under which group ratios are preserved across scales, and introduces StayFair. StayFair equalizes classifier output distributions for classifier guidance and applies a prompt-dependent offset to the null embedding for classifier-free guidance; the method modifies only the guidance step, is orthogonal to model-level debiasing, and is claimed to maintain image quality while decoupling fairness from the chosen guidance scale.

Significance. If the decomposition and derived condition hold, the result is significant because it removes a practical barrier to deploying fair diffusion models: existing debiasing methods are scale-specific and lose fairness when users increase the guidance scale for better prompt alignment. The orthogonality property allows StayFair to be layered on top of prior fair models without retraining, and the construction-by-equalization approach provides a clean, falsifiable mechanism that could be adopted in both class-conditional and text-to-image pipelines.

major comments (2)
  1. [Abstract / §3] The monotonic-growth claim for guidance bias (abstract) is load-bearing for the central argument that it eventually dominates high-guidance regimes. The manuscript must supply the explicit decomposition (presumably in §3) together with the proof or empirical verification that the guidance term increases with scale for arbitrary classifiers or null embeddings; without this, the necessity of the StayFair correction cannot be assessed.
  2. [§4] The derived condition for retaining group ratios across scales (abstract) is the foundation of both StayFair variants. The manuscript should state this condition as an explicit equation or theorem and show that the proposed equalisation of classifier outputs (classifier guidance) and null-embedding shift (CFG) satisfy it by construction; any hidden dependence on fitted parameters or additional distributional assumptions must be flagged.
minor comments (2)
  1. [Abstract] Notation for the two bias components and the extended SDP should be introduced once and used consistently; the abstract uses “total bias,” “model bias,” and “guidance bias” without symbols.
  2. [Experiments] The experimental section should report quantitative fairness metrics (e.g., demographic parity gap) at multiple guidance scales for both the baseline fair model and the StayFair-augmented version, together with standard image-quality metrics, to substantiate the “without sacrificing image quality” claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and the recommendation of minor revision. The comments help clarify the presentation of our theoretical contributions. We address each major comment below and will update the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract / §3] The monotonic-growth claim for guidance bias (abstract) is load-bearing for the central argument that it eventually dominates high-guidance regimes. The manuscript must supply the explicit decomposition (presumably in §3) together with the proof or empirical verification that the guidance term increases with scale for arbitrary classifiers or null embeddings; without this, the necessity of the StayFair correction cannot be assessed.

    Authors: We agree that an explicit decomposition and proof are necessary for rigor. Section 3 already contains the bias decomposition into model and guidance components. In the revision we will restate this decomposition as a formal equation, provide a proof that the guidance bias term is monotonically non-decreasing in the guidance scale (under the standard diffusion guidance assumptions), and add empirical plots verifying the growth for multiple classifiers and null embeddings. This will directly support the claim that guidance bias dominates at high scales. revision: yes

  2. Referee: [§4] The derived condition for retaining group ratios across scales (abstract) is the foundation of both StayFair variants. The manuscript should state this condition as an explicit equation or theorem and show that the proposed equalisation of classifier outputs (classifier guidance) and null-embedding shift (CFG) satisfy it by construction; any hidden dependence on fitted parameters or additional distributional assumptions must be flagged.

    Authors: We will revise Section 4 to present the group-ratio preservation condition as an explicit theorem. We will then show, by direct substitution, that both the classifier-output equalization (classifier guidance) and the prompt-dependent null-embedding offset (CFG) satisfy the theorem by construction. All distributional assumptions (e.g., on classifier outputs or embedding spaces) and any dependence on fitted parameters will be stated explicitly in the theorem statement and proof. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The abstract presents a decomposition of total bias into model bias plus guidance bias, asserts monotonic growth of the latter with scale, extends Strong Demographic Parity to derive a retention condition for group ratios, and constructs StayFair to satisfy the condition by equalizing classifier outputs or shifting null embeddings. This constitutes a standard design step that implements a mathematically derived requirement rather than renaming a fit, smuggling an ansatz via self-citation, or reducing a claimed prediction to its own inputs. No equations appear in the abstract that would allow a self-definitional or fitted-input reduction, and the orthogonality claim follows directly from the scope of the modification. The full manuscript would need to be inspected for any load-bearing self-citation chains, but the provided text shows an independent derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no free parameters, axioms, or invented entities can be identified from the provided text.

pith-pipeline@v0.9.1-grok · 5741 in / 892 out tokens · 26462 ms · 2026-06-29T13:08:03.602225+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

67 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    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. InCVPR, pages 10684–10695, 2022

    2022

  2. [2]

    Scaling rectified flow transform- ers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transform- ers for high-resolution image synthesis. InICML, pages 12606–12633, 2024

    2024

  3. [3]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, volume 34, pages 8780–8794, 2021

    2021

  4. [4]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  5. [5]

    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. InICCV, pages 3043–3054, 2023

    2023

  6. [6]

    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. InNAACL, pages 6367–6384, 2024

    2024

  7. [7]

    Racial biases in AIs and Gemini’s inability to write narratives about black people.Emerging Media, 2(2):277–287, 2024

    Julia Barroso da Silveira and Ellen Alves Lima. Racial biases in AIs and Gemini’s inability to write narratives about black people.Emerging Media, 2(2):277–287, 2024

    2024

  8. [8]

    Fair diffusion: Instructing text-to-image generation models on fairness.arXiv preprint at arXiv:2302.10893,

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

    work page arXiv 2023

  9. [9]

    Rethinking training for de-biasing text-to-image generation: Unlocking the potential of stable diffusion

    Eunji Kim, Siwon Kim, Minjun Park, Rahim Entezari, and Sungroh Yoon. Rethinking training for de-biasing text-to-image generation: Unlocking the potential of stable diffusion. InCVPR, pages 13361–13370, 2025

    2025

  10. [10]

    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. InICLR, 2024

    2024

  11. [11]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. InWACV, pages 5111–5120, 2024

    2024

  12. [12]

    Wasserstein fair classification, 2020

    Ray Jiang, Aldo Pacchiano, Tom Stepleton, Heinrich Jiang, and Silvia Chiappa. Wasserstein fair classification, 2020

    2020

  13. [13]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InICCV, pages 3730–3738, 2015

    2015

  14. [14]

    Stable bias: Evaluat- ing societal representations in diffusion models

    Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluat- ing societal representations in diffusion models. InNeurIPS, volume 36, pages 56338–56351, 2023

    2023

  15. [15]

    Typology of risks of generative text-to- image models

    Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. Typology of risks of generative text-to- image models. InAAAI/ACM AIES, pages 396–410, 2023

    2023

  16. [16]

    A large scale analysis of gender biases in text-to-image generative models.arXiv preprint arXiv:2503.23398, 2025

    Leander Girrbach, Stephan Alaniz, Genevieve Smith, and Zeynep Akata. A large scale analysis of gender biases in text-to-image generative models.arXiv preprint arXiv:2503.23398, 2025

    work page arXiv 2025

  17. [17]

    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. InAAAI/ACM AIES, 2024

    2024

  18. [18]

    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. InFAccT, pages 1493–1504, 2023

    2023

  19. [19]

    Venkatesh Babu

    Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. InCVPR, pages 6668–6678, 2024. 10

    2024

  20. [20]

    Fairqueue: Rethinking prompt learning for fair text-to-image generation.NeurIPS, 37:22878–22926, 2024

    Christopher T Teo, Milad Abdollahzadeh, Xinda Ma, and Ngai-Man Cheung. Fairqueue: Rethinking prompt learning for fair text-to-image generation.NeurIPS, 37:22878–22926, 2024

    2024

  21. [21]

    Self-discovering inter- pretable diffusion latent directions for responsible text-to-image generation

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

    2024

  22. [22]

    De-stereotyping text-to-image models through prompt tuning.arXiv preprint arXiv:2302.02369, 2024

    Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning.arXiv preprint arXiv:2302.02369, 2024

    work page arXiv 2024

  23. [23]

    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. InICLR, 2024

    2024

  24. [24]

    arXiv preprint arXiv:2302.00070 , year=

    Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts.arXiv preprint arXiv:2302.00070, 2023

    work page arXiv 2023

  25. [25]

    LightFair: Towards an efficient alternative for fair T2I diffusion via debiasing pre-trained text encoders

    Boyu Han, Qianqian Xu, Shilong Bao, Zhiyong Yang, Kangli Zi, and Qingming Huang. LightFair: Towards an efficient alternative for fair T2I diffusion via debiasing pre-trained text encoders. InNeurIPS, volume 38, pages 22671–22724, 2026

    2026

  26. [26]

    Mining GOLD samples for conditional GANs

    Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, and Jinwoo Shin. Mining GOLD samples for conditional GANs. InNeurIPS, 2019

    2019

  27. [27]

    Rethinking prompt design for inference-time scaling in text-to-visual generation

    Subin Kim, Sangwoo Mo, Mamshad Nayeem Rizve, Yiran Xu, Difan Liu, Jinwoo Shin, and Tobias Hinz. Rethinking prompt design for inference-time scaling in text-to-visual generation. InCVPR, 2026

    2026

  28. [28]

    Universal guidance for diffusion models

    Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. InCVPRW, pages 843–852, 2023

    2023

  29. [29]

    Tfg: Unified training-free guidance for diffusion models.NeurIPS, 37:22370–22417, 2024

    Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models.NeurIPS, 37:22370–22417, 2024

    2024

  30. [30]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models

    Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. InNeurIPS, 2024

    2024

  31. [31]

    Seyedmorteza Sadat, Otmar Hilliges, and Romann M. Weber. Eliminating oversaturation and artifacts of high guidance scales in diffusion models. InICLR, 2025

    2025

  32. [32]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. InNeurIPS, volume 37, pages 52996–53021, 2024

    2024

  33. [33]

    Improving sample quality of diffusion models using self-attention guidance

    Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungryong Kim. Improving sample quality of diffusion models using self-attention guidance. InICCV, pages 7462–7471, 2023

    2023

  34. [34]

    Self-rectifying diffusion sampling with perturbed-attention guidance

    Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. InECCV, pages 1–17, 2024

    2024

  35. [35]

    How I met your bias: Investigating bias amplification in diffusion models

    Nathan Roos, Ekaterina Iakovleva, Ani Gjergji, Vito Paolo Pastore, and Enzo Tartaglione. How I met your bias: Investigating bias amplification in diffusion models. InWACV, pages 5374–5383, 2026

    2026

  36. [36]

    A review on fairness in machine learning.ACM computing surveys (CSUR), 55(3):1–44, 2022

    Dana Pessach and Erez Shmueli. A review on fairness in machine learning.ACM computing surveys (CSUR), 55(3):1–44, 2022

    2022

  37. [37]

    Discovering and mitigating visual biases through keyword explanation

    Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. InCVPR, 2024. 11

    2024

  38. [38]

    Understanding the Effects of Distractors on Reasoning Vision-Language Models

    Jiyun Bae, Hyunjong Ok, Sangwoo Mo, and Jaeho Lee. Do reasoning vision-language models inversely scale in test-time compute? a distractor-centric empirical analysis.arXiv preprint arXiv:2511.21397, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Fairness through awareness

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. InITCS, 2012

    2012

  40. [40]

    Equality of opportunity in supervised learning

    Moritz Hardt, Eric Price, and Nathan Srebro. Equality of opportunity in supervised learning. In NeurIPS, 2016

    2016

  41. [41]

    Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi. Fairness constraints: Mechanisms for fair classification. InAISTATS, 2017

    2017

  42. [42]

    A reductions approach to fair classification

    Alekh Agarwal, Alina Beygelzimer, Miroslav Dudík, John Langford, and Hanna Wallach. A reductions approach to fair classification. InICML, 2018

    2018

  43. [43]

    Hashimoto, and Percy Liang

    Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. InICLR, 2020

    2020

  44. [44]

    Simple data balancing achieves competitive worst-group-accuracy

    Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. InCLeaR, 2022

    2022

  45. [45]

    Fair regression with Wasserstein barycenters

    Evgenii Chzhen, Christophe Denis, Mohamed Hebiri, Luca Oneto, and Massimiliano Pontil. Fair regression with Wasserstein barycenters. InNeurIPS, volume 33, pages 7321–7331, 2020

    2020

  46. [46]

    Optimal pre-processing to achieve fairness and its relationship with total variation barycenter.arXiv preprint arXiv:2101.06811, 2021

    Farhad Farokhi. Optimal pre-processing to achieve fairness and its relationship with total variation barycenter.arXiv preprint arXiv:2101.06811, 2021

    work page arXiv 2021

  47. [47]

    Fairness through matching.Transactions on Machine Learning Research, 2025

    Kunwoong Kim, Insung Kong, Jongjin Lee, Minwoo Chae, Sangchul Park, and Yongdai Kim. Fairness through matching.Transactions on Machine Learning Research, 2025

    2025

  48. [48]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, volume 33, pages 6840–6851, 2020

    2020

  49. [49]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InICLR, 2021

    2021

  50. [50]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InNeurIPS, volume 35, pages 26565–26577, 2022

    2022

  51. [51]

    Characteristic guidance: Non-linear correction for diffusion model at large guidance scale

    Candi Zheng and Yuan Lan. Characteristic guidance: Non-linear correction for diffusion model at large guidance scale. InICML, pages 61386–61412, 2024

    2024

  52. [52]

    Provable efficiency of guidance in diffusion models for general data distribution

    Gen Li and Yuchen Jiao. Provable efficiency of guidance in diffusion models for general data distribution. InICML, pages 35034–35046, 2025

    2025

  53. [53]

    Adaptive diffusion guidance via stochastic optimal control

    Iskander Azangulov, Peter Potaptchik, Qinyu Li, Eddie Aamari, George Deligiannidis, and Judith Rousseau. Adaptive diffusion guidance via stochastic optimal control. InAISTATS, 2026

    2026

  54. [54]

    Embedding arithmetic of multimodal queries for image retrieval

    Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. InCVPR, pages 4950–4958, 2022

    2022

  55. [55]

    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. InAAAI, volume 39, pages 26256–26264, 2025

    2025

  56. [56]

    Input perturbation reduces exposure bias in diffusion models

    Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Input perturbation reduces exposure bias in diffusion models. InICML, pages 26245–26265, 2023

    2023

  57. [57]

    Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation

    Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. InWACV, pages 1548–1558, 2021

    2021

  58. [58]

    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 Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. InNeurIPS, pages 6626–6637, 2017. 12

    2017

  59. [59]

    Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling.Transactions on Machine Learning Research, 2023

    Junhyun Nam, Sangwoo Mo, Jaeho Lee, and Jinwoo Shin. Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling.Transactions on Machine Learning Research, 2023

    2023

  60. [60]

    Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L

    George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. InNeurIPS, 2023

    2023

  61. [61]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InICML, pages 8748–8763, 2021

    2021

  62. [62]

    Yolov10: Real-time end-to-end object detection.NeurIPS, 37:107984–108011, 2024

    Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection.NeurIPS, 37:107984–108011, 2024

    2024

  63. [63]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. InEMNLP, pages 7514–7528, 2021

    2021

  64. [64]

    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 Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models....

    2022

  65. [65]

    Pick-a-Pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-Pic: An open dataset of user preferences for text-to-image generation. InNeurIPS, volume 36, pages 36652–36663, 2023

    2023

  66. [66]

    Perception prioritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. InCVPR, pages 11472–11481, 2022

    2022

  67. [67]

    % # +! $ # '&

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, volume 29, 2016. 13 Appendix A Additional Analysis 15 A.1 Combining StayFair with attribute-based methods . . . . . . . . . . . . . . . . . . 15 A.2 Analysis of guidance-induced ...

    2016