pith. sign in

arxiv: 2607.01147 · v1 · pith:HL3PDMLUnew · submitted 2026-07-01 · 💻 cs.CV

EquiSteer: Cross-Attention Steering Towards a Fairer Text-Guided Image Generation

Tatiana Gaintseva , Akshit Achara , Gregory Slabaugh , Jiankang Deng , Ismail Elezi This is my paper

Pith reviewed 2026-07-02 13:28 UTC · model grok-4.3

classification 💻 cs.CV
keywords debiasingtext-to-image generationdiffusion modelscross-attentioninference-time interventionfairnessgenerative models
0
0 comments X

The pith

EquiSteer reduces demographic bias in text-to-image diffusion models by steering cross-attention activations at inference time using precomputed vectors and a prompt-aware gate.

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

The paper establishes that a training-free technique can correct skewed outputs on neutral prompts such as photos of nurses or CEOs by clearing unwanted attribute signals in cross-attention and injecting balanced ones instead. It precomputes steering vectors once from contrastive prompt pairs and applies them selectively during generation so that only neutral prompts receive the correction. This matters to a sympathetic reader because current debiasing approaches often demand retraining, batch processing, or per-prompt tuning, while the new method operates per sample on already-deployed models. If correct, it shows that inference-time intervention at the cross-attention layer is sufficient to shrink parity gaps substantially without retraining the underlying diffusion model.

Core claim

EquiSteer is a training-free method that works per sample by steering cross-attention activations at inference time. For each target attribute it precomputes steering vectors from contrastive prompts. At generation time a prompt-aware gate leaves attribute-specific prompts untouched while for neutral ones it clears existing attribute signals from the CA activations and injects a target attribute. Across SD-1.5, SD-2.1, SDXL, and SANA the method reduces the average parity gap by up to 87 percent with minimal effect on image quality and text-image alignment.

What carries the argument

The prompt-aware gate that selectively clears existing attribute signals from cross-attention activations and injects precomputed steering vectors derived from contrastive prompts.

If this is right

  • The method applies without retraining to SD-1.5, SD-2.1, SDXL, and SANA.
  • Average parity gap on common demographic attributes drops by as much as 87 percent.
  • Image quality and text-image alignment remain essentially unchanged under the metrics used.
  • No batch-level control or prompt-specific tuning is required at generation time.

Where Pith is reading between the lines

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

  • The same cross-attention steering pattern could be tested on attributes beyond the gender and race pairs examined here.
  • Because the gate operates per prompt, the technique might be combined with user-provided safety filters in deployed systems.
  • If the steering vectors prove stable across prompt styles, they could be shared as lightweight bias-correction modules.

Load-bearing premise

Steering vectors derived from contrastive prompts can be selectively injected via a prompt-aware gate without introducing new unintended biases or degrading the generative distribution in ways not captured by the reported metrics.

What would settle it

A test set of neutral prompts where applying EquiSteer either widens the measured parity gap on the target attribute or lowers text-image alignment scores below the baseline model on the same prompts.

Figures

Figures reproduced from arXiv: 2607.01147 by Akshit Achara, Gregory Slabaugh, Ismail Elezi, Jiankang Deng, Tatiana Gaintseva.

Figure 1
Figure 1. Figure 1: Examples of EquiSteer for debiasing the gender concept. Top block: generations for the prompt “A photo of a nurse” with SANA, bottom block: “A photo of a CEO” with SDXL. In both cases, the top row corresponds to the vanilla model and the bottom row to EquiSteer, with ten generation seeds shown for each prompt. Abstract. Text-to-image diffusion models power everyday creative tasks, but they still reproduce … view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the proposed method, EquiSteer. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Failure cases of the basic steering mechanism for gender (left) and eyeglasses (right). We first introduce a gating mecha￾nism that determines whether debias￾ing applies to the current prompt. CA activations already encode whether an attribute is explicitly specified in the prompt, and we exploit this to build an attribute-specificity detector. Attribute expression as a dot-product signal. For a target att… view at source ↗
Figure 4
Figure 4. Figure 4: Dot-product statistics between cross-attention outputs and the steering vector of the male attribute. (a) For SANA, attribute-specific prompts induce substantially stronger dot-product responses than attribute-neutral prompts, as seen in the heatmaps of layer l = 5 at denoising step t = 0. (b) For SD-1.5, the maximal dot-product statistic separates attribute-specific from attribute-neutral prompts in inter… view at source ↗
Figure 5
Figure 5. Figure 5: Attribute-preservation results on SD-1.5 for attribute-specific prompts. The target value is 1.0, meaning that the gender explicitly stated in the prompt should be preserved in the gen￾erated image. Aggregated over 8 professions. Model Method CLIP ↑ CMMD ↓ SD 1.5 Vanilla SD 26.42 0.532 FairDiff 26.03 0.586 UCE 18.10 1.240 FTDiff 25.61 0.783 SelfDisc 24.97 0.900 TEI 26.56 0.509 EquiSteer 26.63 0.519 SD 2 Va… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative examples of EquiSteer on the race concept. Top row in each panel: vanilla model; bottom row: EquiSteer. Overall, across all four backbones, EquiSteer consistently improves debiasing performance on the gender concept while preserving fidelity to attribute-specific prompts. As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Debiasing of the age attribute on SDXL. For each profession, top row: vanilla SDXL; bottom row: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Debiasing of the age attribute on SANA-1.5. For each profession, top row: vanilla SANA-1.5; bottom row: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Debiasing of the body type attribute on SDXL. For each profession, top row: vanilla SDXL; bottom row: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p033_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Debiasing of the body type attribute on SANA-1.5. For each profession, top row: vanilla SANA-1.5; bottom row: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p034_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Debiasing of eyeglasses concept on SD-1.5. Top: vanilla SD-1.5, bottom: Eq￾uiSteer model and with EquiSteer enabled. The results are shown in [PITH_FULL_IMAGE:figures/full_fig_p036_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Debiasing of eyeglasses concept on SDXL. Top: vanilla SDXL, bottom: Equi￾Steer (a) CEO (b) Doctor (c) Pilot (d) Technician (e) Teacher (f ) Librarian (g) Nurse (h) Fashion Designer [PITH_FULL_IMAGE:figures/full_fig_p037_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Debiasing of eyeglasses concept on SANA-1.5. Top: vanilla SANA-1.5, bottom: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p037_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Joint debiasing of gender and race. Each profession block shows 20 EquiSteer￾debiased generations, illustrating the diversity achieved by simultaneously steering both attributes [PITH_FULL_IMAGE:figures/full_fig_p044_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Joint debiasing of gender, race, age, and body type. Each profession block shows 20 EquiSteer-debiased generations, illustrating the diversity achieved by simultaneously steering all four attributes [PITH_FULL_IMAGE:figures/full_fig_p045_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Layer-wise gate AUROC for the gender attribute on all three backbones used in the main paper. Each cell of the heatmap is the AUROC of the maximal token response statistic (Eq. 4) for the male-to-female steering direction at the corresponding (denoising step, CA block index), computed on the same npos=nneg calibration prompts used to fit the per-direction threshold thra . The red vertical line marks l gat… view at source ↗
Figure 17
Figure 17. Figure 17: Debiasing of race concept on SD-1.5. Top: vanilla SD-1.5, bottom: EquiSteer (a) CEO (b) Doctor (c) Pilot (d) Technician (e) Teacher (f ) Librarian (g) Nurse (h) Fashion Designer [PITH_FULL_IMAGE:figures/full_fig_p060_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Debiasing of race concept on SDXL. Top: vanilla SDXL, bottom: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p060_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Debiasing of race concept on SANA-1.5. Top: vanilla SANA-1.5, bottom: EquiSteer [PITH_FULL_IMAGE:figures/full_fig_p061_19.png] view at source ↗
read the original abstract

Text-to-image diffusion models power everyday creative tasks, but they still reproduce the demographic biases in their training data. On common prompts such as ``a photo of a nurse,'' ``a photo of a CEO'', they skew their outputs toward one gender, driven by the statistics of training data rather than anything in the text. Existing debiasing methods show promise in narrow settings but require retraining, batch-level control, or prompt-specific tuning, limiting their scalability. We propose \emph{EquiSteer}, a training-free method that works per sample by steering cross-attention (CA) activations at inference time. For each target attribute, EquiSteer precomputes steering vectors from contrastive prompts. Then at generation time, a prompt-aware gate leaves attribute-specific prompts untouched, while for neutral ones it clears existing attribute signals from the CA activations and injects a target attribute. Across SD-1.5, SD-2.1, SDXL, and SANA, EquiSteer reduces the average parity gap by up to $87\%$, with minimal effect on image quality and text-image alignment. Code is available at \href{https://github.com/Atmyre/EquiSteer}{https://github.com/Atmyre/EquiSteer}.%

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 introduces EquiSteer, a training-free inference-time method for mitigating demographic biases in text-to-image diffusion models. It precomputes steering vectors from contrastive prompts for target attributes and uses a prompt-aware gate to selectively clear and inject attribute signals in cross-attention activations during generation, leaving attribute-specific prompts untouched. Experiments across SD-1.5, SD-2.1, SDXL, and SANA report up to 87% reduction in average parity gap with minimal degradation in image quality and text-image alignment; code is released.

Significance. If the empirical results hold under rigorous validation, EquiSteer would represent a practical advance in scalable bias mitigation for generative models, avoiding retraining or batch-level constraints of prior work. The training-free design and public code release strengthen the contribution by enabling direct reproducibility and extension.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (experimental protocol): the reported parity-gap reductions (up to 87%) and claims of minimal quality/alignment impact lack explicit definitions of the parity-gap metric, details on baseline implementations, statistical significance testing, or controls for post-hoc prompt/seed selection; without these, the central quantitative claim cannot be fully assessed from the provided information.
  2. [§3.2] §3.2 (prompt-aware gate): the assumption that selective injection of steering vectors via the gate avoids introducing new unintended biases or altering the generative distribution in unmeasured ways is load-bearing for the fairness claim, yet the manuscript provides no ablation isolating gate-induced side effects on non-target attributes.
minor comments (2)
  1. [§2-3] Notation for cross-attention activations and steering vectors should be defined once in §2 or §3 with consistent symbols across equations and text.
  2. [Figures] Figure captions for qualitative examples should explicitly state the prompt, model, and whether the image is before/after steering.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment, recognition of the practical contribution, and recommendation for minor revision. We address each major comment below and will update the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (experimental protocol): the reported parity-gap reductions (up to 87%) and claims of minimal quality/alignment impact lack explicit definitions of the parity-gap metric, details on baseline implementations, statistical significance testing, or controls for post-hoc prompt/seed selection; without these, the central quantitative claim cannot be fully assessed from the provided information.

    Authors: We agree that greater explicitness will improve clarity. The parity-gap metric is defined in §4 as the absolute difference between the empirical frequencies of each demographic category across generated images for a fixed prompt, then averaged over the prompt set and attribute pairs. Baseline methods are re-implemented from their original papers using the authors’ recommended hyperparameters and the same prompt/seed pool; we will state these settings verbatim in the revised §4. Statistical significance testing was omitted because of the prohibitive cost of repeated full-generation runs across four models, but we will report per-seed variance and add a limitations note. All quantitative results use a fixed set of 10 seeds and the complete prompt list with no post-hoc filtering or selection; we will add an explicit statement of this protocol in both the abstract and §4. revision: yes

  2. Referee: [§3.2] §3.2 (prompt-aware gate): the assumption that selective injection of steering vectors via the gate avoids introducing new unintended biases or altering the generative distribution in unmeasured ways is load-bearing for the fairness claim, yet the manuscript provides no ablation isolating gate-induced side effects on non-target attributes.

    Authors: We acknowledge that an explicit ablation isolating gate effects on non-target attributes is absent. The gate is intended to remain inactive for attribute-specific prompts, thereby preserving the original distribution for those cases. To directly address the concern, the revised manuscript will include a new ablation (added to §4) that measures parity gaps on unrelated attributes when steering is applied only to a target attribute (e.g., gender parity when steering for race). This will quantify any measurable side effects and will be reported alongside the existing results. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an empirical, training-free inference-time steering method that constructs contrastive steering vectors from external prompts and applies them via a prompt-aware gate. No equations, derivations, or first-principles predictions are presented that reduce the reported parity-gap reductions to fitted parameters, self-definitions, or self-citation chains. The central claims are falsifiable empirical measurements across multiple diffusion models and are not equivalent to their inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method relies on precomputed steering vectors whose construction details are not specified here.

pith-pipeline@v0.9.1-grok · 5779 in / 1036 out tokens · 23917 ms · 2026-07-02T13:28:12.989059+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

32 extracted references · 12 canonical work pages · 4 internal anchors

  1. [1]

    arXiv preprint arXiv:2509.13496 (2025)

    Chakraborty, R., Che, X., Xu, D., Faklaris, C., Niu, X., Yuan, S.: Biasmap: Lever- agingcross-attentionstodiscoverandmitigatehiddensocialbiasesintext-to-image generation. arXiv preprint arXiv:2509.13496 (2025)

    work page arXiv 2025

  2. [2]

    ACM trans- actions on Graphics (TOG)42(4), 1–10 (2023)

    Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM trans- actions on Graphics (TOG)42(4), 1–10 (2023)

    2023

  3. [3]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 5343–5353 (2024)

    2024

  4. [4]

    In: European Conference on Computer Vision

    Chinchure, A., Shukla, P., Bhatt, G., Salij, K., Hosanagar, K., Sigal, L., Turk, M.: Tibet: Identifying and evaluating biases in text-to-image generative models. In: European Conference on Computer Vision. pp. 429–446. Springer (2024)

    2024

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    D’Incà, M., Peruzzo, E., Mancini, M., Xu, D., Goel, V., Xu, X., Wang, Z., Shi, H., Sebe, N.: Openbias: Open-set bias detection in text-to-image generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12225–12235 (2024)

    2024

  6. [6]

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

    Friedrich, F., Brack, M., Struppek, L., Hintersdorf, D., Schramowski, P., Luccioni, S., Kersting, K.: Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893 (2023)

    work page arXiv 2023

  7. [7]

    arXiv preprint arXiv:2510.21363 (2025)

    Fu, Z., Brown, R., Shao, S., Rawal, K., Delaney, E., Russell, C.: Fairima- gen: Post-processing for bias mitigation in text-to-image models. arXiv preprint arXiv:2510.21363 (2025)

    work page arXiv 2025

  8. [8]

    CASteer: Cross-Attention Steering for Controllable Concept Erasure

    Gaintseva, T., Benning, M., Slabaugh, G., Elezi, I.: Casteer: Steering diffusion models for controllable generation. arXiv preprint arXiv:2503.09630 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111–5120 (2024)

    2024

  10. [10]

    arXiv preprint arXiv:2503.23398 (2025)

    Girrbach, L., Alaniz, S., Smith, G., Akata, Z.: A large scale analysis of gender biases in text-to-image generative models. arXiv preprint arXiv:2503.23398 (2025)

    work page arXiv 2025

  11. [11]

    arXiv preprint arXiv:2509.23639 (2025)

    Han, B., Xu, Q., Bao, S., Yang, Z., Zi, K., Huang, Q.: Lightfair: Towards an efficient alternative for fair t2i diffusion via debiasing pre-trained text encoders. arXiv preprint arXiv:2509.23639 (2025)

    work page arXiv 2025

  12. [12]

    In: Proceedings of the 1st ACM Multimedia Workshop on Multi-modal Misinformation Governance in the Era of Foundation Models

    He, R., Xue, C., Tan, H., Zhang, W., Yu, Y., Bai, S., Qi, X.: Debiasing text-to- image diffusion models. In: Proceedings of the 1st ACM Multimedia Workshop on Multi-modal Misinformation Governance in the Era of Foundation Models. pp. 29–36 (2024)

    2024

  13. [13]

    Prompt-to-Prompt Image Editing with Cross Attention Control

    Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  14. [14]

    In: Proceedings of the 2021 conference on empirical methods in natural language processing

    Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference- free evaluation metric for image captioning. In: Proceedings of the 2021 conference on empirical methods in natural language processing. pp. 7514–7528 (2021)

    2021

  15. [15]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 9307–9315 (2024)

    2024

  16. [16]

    In: EquiSteer: Cross-Attention Steering for Fairer T2I Generation 17 Proceedings of the Computer Vision and Pattern Recognition Conference

    Kim, E., Kim, S., Park, M., Entezari, R., Yoon, S.: Rethinking training for de- biasing text-to-image generation: Unlocking the potential of stable diffusion. In: EquiSteer: Cross-Attention Steering for Fairer T2I Generation 17 Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 13361–13370 (2025)

    2025

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, H., Shen, C., Torr, P., Tresp, V., Gu, J.: Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12006– 12016 (2024)

    2024

  18. [18]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Li, J., Hu, L., Zhang, J., Zheng, T., Zhang, H., Wang, D.: Fair text-to-image diffusion via fair mapping. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 26256–26264 (2025)

    2025

  19. [19]

    Proceedings of Machine Learning Research, vol

    Li, J., Li, D., Xiong, C., Hoi, S.C.H.: BLIP: bootstrapping language-image pre- trainingforunifiedvision-languageunderstandingandgeneration.In:International Conference on Machine Learning, ICML 2022, 17–23 July 2022, Baltimore, Mary- land, USA. Proceedings of Machine Learning Research, vol. 162, pp. 12888–12900. PMLR (2022)

    2022

  20. [20]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014)

    2014

  21. [21]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, B., Wang, C., Cao, T., Jia, K., Huang, J.: Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7817– 7826 (2024)

    2024

  22. [22]

    Stable Bias: Analyz- ing Societal Representations in Diffusion Models, 2023

    Luccioni, A.S., Akiki, C., Mitchell, M., Jernite, Y.: Stable bias: Analyzing societal representations in diffusion models. arXiv preprint arXiv:2303.11408 (2023)

    work page arXiv 2023

  23. [23]

    Yidong Ouyang, Liyan Xie, Hongyuan Zha, and Guang Cheng

    Na, B., Park, M., Sim, G., Shin, D., Bae, H., Kang, M., Kwon, S.J., Kang, W., Moon, I.C.: Diffusion adaptive text embedding for text-to-image diffusion models. arXiv preprint arXiv:2510.23974 (2025)

    work page arXiv 2025

  24. [24]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Parihar, R., Bhat, A., Basu, A., Mallick, S., Kundu, J.N., Babu, R.V.: Balanc- ing act: distribution-guided debiasing in diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6668–6678 (2024)

    2024

  25. [25]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684– 10695 (June 2022)

    2022

  28. [28]

    Scientific Reports15(1), 31401 (2025)

    Schneider, M., Hagendorff, T.: Investigating toxicity and bias in stable diffusion text-to-image models. Scientific Reports15(1), 31401 (2025)

    2025

  29. [29]

    In: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)

    Seshadri, P., Singh, S., Elazar, Y.: 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 Papers). pp. 6367–6384 (2024)

    2024

  30. [30]

    Finetuning text-to-image diffusion models for fairness

    Shen, X., Du, C., Pang, T., Lin, M., Wong, Y., Kankanhalli, M.: Finetuning text- to-image diffusion models for fairness. arXiv preprint arXiv:2311.07604 (2023)

    work page arXiv 2023

  31. [31]

    Journal of Imaging11(2), 35 (2025) 18 T

    Wu, Y., Nakashima, Y., Garcia, N.: Revealing gender bias from prompt to image in stable diffusion. Journal of Imaging11(2), 35 (2025) 18 T. Gaintseva et al

    2025

  32. [32]

    SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers

    Xie, E., Chen, J., Chen, J., Cai, H., Tang, H., Lin, Y., Zhang, Z., Li, M., Zhu, L., Lu, Y., et al.: Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629 (2024) Supplementary Material – Contents S1 Limitations ................................................... P1 S2 Future Work .................

    work page internal anchor Pith review Pith/arXiv arXiv 2024