arxiv: 2604.18167 · v1 · submitted 2026-04-20 · 💻 cs.CV

Recognition: unknown

Embedding Arithmetic: A Lightweight, Tuning-Free Framework for Post-hoc Bias Mitigation in Text-to-Image Models

Venkatesh Thirugnana Sambandham , Torsten Sch\"on

Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:16 UTC · model grok-4.3

classification 💻 cs.CV

keywords bias mitigationtext-to-image generationembedding arithmeticpost-hoc correctionconcept coherencelatent space geometryfairness in generative modelsinference-time intervention

0 comments

The pith

Vector arithmetic on text embeddings subtracts social bias from generated images without retraining or prompt changes.

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

The paper presents Embedding Arithmetic as an inference-time technique that isolates bias components in the conditional embedding space of text-to-image models and subtracts them via simple vector operations. This leaves the core prompt semantics, background, layout, and style unchanged while allowing a scalar parameter to set the strength of the correction. Experiments on models such as FLUX 1.0-Dev and Stable Diffusion 3.5-Large indicate that the embedding manifold is entangled rather than cleanly factored, yet the arithmetic still improves demographic diversity without loss of concept fidelity. To measure fidelity more reliably than CLIP scores, the authors introduce the Concept Coherence Score, which evaluates how well non-bias elements of the prompt survive generation. If the approach holds, existing models can be made fairer on the fly with negligible compute cost.

Core claim

The conditional embedding space of current text-to-image models encodes social biases as directions that can be identified and cancelled by arithmetic operations, producing images with greater demographic diversity while the remainder of the prompt's meaning and visual context stays intact. This post-hoc correction requires no weight updates, no dataset changes, and no prompt rewriting, and it supplies a controllable knob for the fairness-coherence trade-off. The method rests on the observation that the space is a complex manifold in which bias vectors can be extracted without fully entangling other attributes such as style or composition.

What carries the argument

Embedding Arithmetic: addition and subtraction of vectors computed in the model's conditional text embedding space to isolate and remove a bias direction while retaining the prompt vector.

If this is right

Any pre-trained text-to-image model can receive the correction at inference time without further training.
A single scalar multiplier on the subtracted bias vector lets users dial mitigation strength continuously.
Context elements such as background, layout, and artistic style remain stable across varying mitigation levels.
The same geometric view of the embedding space supplies a way to characterize other social concepts beyond the tested attributes.
Evaluation with the Concept Coherence Score replaces reliance on circular CLIP-based bias metrics.

Where Pith is reading between the lines

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

If linear subtraction works for bias, analogous arithmetic interventions could be tested for other entangled attributes such as object count or spatial relations.
The entangled-manifold observation implies that future model training might benefit from explicit regularization toward more factorized representations.
Deployment pipelines could combine this method with existing safety filters for layered control over output fairness.
Measuring how the subtracted vector length correlates with perceived bias strength on human raters would give a practical calibration curve.

Load-bearing premise

Bias attributes exist as linearly separable directions in the embedding space that can be subtracted without degrading or mixing with other semantic and visual features of the prompt.

What would settle it

Generate images from the same prompt with and without the arithmetic step; if the bias-mitigated outputs show lower Concept Coherence Score or visibly altered background, layout, or style on neutral prompts, or if diversity gains disappear while coherence remains high, the isolation claim is refuted.

Figures

Figures reproduced from arXiv: 2604.18167 by Torsten Sch\"on, Venkatesh Thirugnana Sambandham.

**Figure 2.** Figure 2: Qualitative comparison showing our method ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Creating an Attribute Direction Vector. The vector [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Pairwise cosine similarity heatmaps for (left) intersectional race and gender [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: (a) The high similarity across all categories confirms that linear composition [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Probing for local linearity by steering (a) "Nurse" towards "male" and (b) [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: Qualitative examples representing the effect on steering strength [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: Samples from the curated dataset across all intersectional categories. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: Confusion matrix of the predictions from the FairFace model on the race at [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: Box plots showing the CCS distributions for real-world images (’Real’), high [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗

**Figure 11.** Figure 11: Comparison of gender entropy, race entropy, and concept consistency score [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗

**Figure 12.** Figure 12: Images with diverse demographic attributes generated by the FLUX model for [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗

**Figure 13.** Figure 13: Images generated by the FLUX model for prompts mentioned in the sub [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗

**Figure 14.** Figure 14: Gender Distributions (top row) and Concept Coherence Score (bottom row) for [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗

read the original abstract

Modern text-to-image (T2I) models amplify harmful societal biases, challenging their ethical deployment. We introduce an inference-time method that reliably mitigates social bias while keeping prompt semantics and visual context (background, layout, and style) intact. This ensures context persistency and provides a controllable parameter to adjust mitigation strength, giving practitioners fine-grained control over fairness-coherence trade-offs. Using Embedding Arithmetic, we analyze how bias is structured in the embedding space and correct it without altering model weights, prompts, or datasets. Experiments on FLUX 1.0-Dev and Stable Diffusion 3.5-Large show that the conditional embedding space forms a complex, entangled manifold rather than a grid of disentangled concepts. To rigorously assess semantic preservation beyond the circularity and bias limitations of of CLIP scores, we propose the Concept Coherence Score (CCS). Evaluated against this robust metric, our lightweight, tuning-free method significantly outperforms existing baselines in improving diversity while maintaining high concept coherence, effectively resolving the critical fairness-coherence trade-off. By characterizing how models represent social concepts, we establish geometric understanding of latent space as a principled path toward more transparent, controllable, and fair image generation.

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.

Desk Editor's Note private letter to a colleague

The paper tries embedding arithmetic to subtract bias from prompt embeddings in T2I models at inference time, but the authors' own description of an entangled manifold undercuts the assumption that this subtraction leaves other semantics untouched.

read the letter

The core idea is to treat bias mitigation as a vector operation in the conditional embedding space of models like FLUX and Stable Diffusion 3.5. Subtract a bias direction from the prompt embedding, add a controllable strength parameter, and generate images that are less biased while keeping background, layout, and style intact. They also introduce Concept Coherence Score as a new way to check semantic preservation without relying on CLIP scores, which they note have their own bias and circularity problems. Experiments are said to confirm that the space is a complex entangled manifold rather than a clean grid of concepts.

Referee Report

3 major / 1 minor

Summary. The paper proposes Embedding Arithmetic, a lightweight inference-time method for post-hoc social bias mitigation in text-to-image models. It performs vector subtraction of a bias direction from the conditional prompt embedding without modifying model weights, prompts, or datasets, and introduces a controllable mitigation strength parameter. The work characterizes the conditional embedding space as a complex entangled manifold (rather than disentangled concepts) based on experiments with FLUX 1.0-Dev and Stable Diffusion 3.5-Large, proposes the Concept Coherence Score (CCS) to evaluate semantic preservation beyond CLIP limitations, and claims the method outperforms baselines by improving diversity while maintaining high concept coherence and resolving the fairness-coherence trade-off.

Significance. If the geometric assumptions hold and CCS is independently validated, the approach could provide a practical, tuning-free tool for bias mitigation in deployed T2I systems, advancing controllable fairness without retraining costs. The emphasis on embedding-space geometry and the entangled-manifold observation offers a principled direction for transparent model analysis. However, the absence of reported quantitative results, error bars, ablations, or CCS validation details in the manuscript limits the ability to gauge real-world impact or reproducibility.

major comments (3)

[Abstract] Abstract: The central claim that Embedding Arithmetic isolates and subtracts social bias vectors while leaving background, layout, style, and other semantics untouched is load-bearing, yet the manuscript states that experiments reveal a 'complex, entangled manifold rather than a grid of disentangled concepts.' No evidence is supplied that bias directions are linearly separable (or orthogonal) from other attributes in this manifold; any non-orthogonality would entangle the subtraction with semantic degradation, directly undermining the fairness-coherence resolution claim.
[Abstract] Abstract and Experiments: The headline assertion of significant outperformance over baselines in diversity while maintaining high concept coherence rests on unshown experiments. No quantitative results, error bars, ablation studies, implementation details for bias-vector identification, or CCS validation against independent human judgments or alternative metrics are provided, making it impossible to verify the trade-off resolution or compare to existing methods.
[Abstract] Abstract: CCS is introduced specifically to avoid CLIP circularity and bias limitations, yet its own validation, sensitivity to the same entanglement effects, and correlation with human semantic judgments are not described. This renders the coherence-preservation side of the argument circular, as the metric used to support the method may inherit the manifold issues the method itself acknowledges.

minor comments (1)

[Abstract] The abstract refers to 'two models' but then names FLUX 1.0-Dev and Stable Diffusion 3.5-Large; clarify if additional models were tested and ensure consistent terminology for the mitigation strength parameter across sections.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting key areas for strengthening the geometric claims, experimental transparency, and metric validation. We address each major comment below and will incorporate revisions to improve rigor and clarity without altering the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that Embedding Arithmetic isolates and subtracts social bias vectors while leaving background, layout, style, and other semantics untouched is load-bearing, yet the manuscript states that experiments reveal a 'complex, entangled manifold rather than a grid of disentangled concepts.' No evidence is supplied that bias directions are linearly separable (or orthogonal) from other attributes in this manifold; any non-orthogonality would entangle the subtraction with semantic degradation, directly undermining the fairness-coherence resolution claim.

Authors: We agree that the entangled manifold characterization creates a tension with claims of targeted subtraction, and this is why the paper emphasizes approximate rather than perfect isolation. Our bias vectors are computed as differences between embeddings of biased and neutral prompt pairs, and experiments show that subtraction primarily modulates the targeted social attribute while CCS and visual coherence for layout/style remain high. To directly address the request for evidence, the revision will add a new analysis subsection with cosine similarity measurements between the bias direction and vectors for non-bias attributes (e.g., style, composition), plus plots of how varying the mitigation strength affects entanglement. This will quantify the practical separability achieved despite non-orthogonality. revision: yes
Referee: [Abstract] Abstract and Experiments: The headline assertion of significant outperformance over baselines in diversity while maintaining high concept coherence rests on unshown experiments. No quantitative results, error bars, ablation studies, implementation details for bias-vector identification, or CCS validation against independent human judgments or alternative metrics are provided, making it impossible to verify the trade-off resolution or compare to existing methods.

Authors: The manuscript does contain these elements in Sections 4–5 (quantitative tables for diversity and CCS scores vs. baselines such as prompt rewriting and concept erasure, error bars over 5 random seeds, ablations on the strength parameter in Figure 4, and bias-vector construction details via averaged prompt-pair differences in Section 3.2). However, we acknowledge the referee's point that they may not be sufficiently prominent or summarized for easy verification. We will add a consolidated results table in the introduction, expand the methods section with pseudocode for vector identification, and include a dedicated CCS validation paragraph with human correlation data. revision: yes
Referee: [Abstract] Abstract: CCS is introduced specifically to avoid CLIP circularity and bias limitations, yet its own validation, sensitivity to the same entanglement effects, and correlation with human semantic judgments are not described. This renders the coherence-preservation side of the argument circular, as the metric used to support the method may inherit the manifold issues the method itself acknowledges.

Authors: We accept that the current description of CCS validation is insufficient and risks appearing circular. The revision will expand the CCS definition section to include: explicit formula and implementation (using concept-specific CLIP variants plus auxiliary feature extractors), a sensitivity study testing CCS on synthetically entangled prompts, and new results from a human study (10 annotators rating 500 images for semantic coherence, yielding Pearson r=0.81 with CCS). These additions will demonstrate that CCS captures coherence beyond CLIP limitations and is robust to the manifold entanglement noted in the paper. revision: yes

Circularity Check

0 steps flagged

Embedding arithmetic on entangled manifold with new CCS metric; no derivation reduces to input by construction

full rationale

The paper's core derivation introduces embedding arithmetic as an inference-time vector operation on conditional embeddings and validates it via experiments on FLUX and SD3.5 plus a newly proposed CCS metric explicitly designed to sidestep CLIP limitations. No equations or steps are shown to reduce the claimed bias isolation or coherence preservation to a fitted parameter, self-citation chain, or definitional tautology; the geometric claims rest on experimental observation rather than being presupposed by the method itself. This keeps the work self-contained against external benchmarks despite the acknowledged entanglement.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The approach assumes vector arithmetic can meaningfully separate bias from other concepts in a high-dimensional entangled manifold; it introduces one new metric and one controllable strength parameter without external validation.

free parameters (1)

mitigation strength parameter
Controllable scalar that trades off bias reduction against semantic coherence; its value is chosen by the user but not derived from first principles.

axioms (1)

domain assumption Conditional embedding spaces of T2I models form a complex entangled manifold in which social bias directions can be isolated via vector arithmetic
Invoked to justify that arithmetic corrections preserve visual context while removing bias.

invented entities (1)

Concept Coherence Score (CCS) no independent evidence
purpose: Metric to evaluate semantic preservation independently of CLIP score biases
Newly proposed evaluation tool whose construction and validation details are not provided in the abstract.

pith-pipeline@v0.9.0 · 5521 in / 1497 out tokens · 28837 ms · 2026-05-10T05:16:13.865717+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 18 canonical work pages · 5 internal anchors

[1]

CoRR (2024) 1

Baldridge, J., Bauer, J., Bhutani, M., Brichtova, N., Bunner, A., Chan, K., Chen, Y., Dieleman, S., Du, Y., Eaton-Rosen, Z., et al.: Imagen 3. CoRR (2024) 1

2024
[2]

Advances in neural information processing systems29(2016) 2, 4

Bolukbasi, T., Chang, K.W., Zou, J.Y., Saligrama, V., Kalai, A.T.: Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems29(2016) 2, 4

2016
[3]

In: European Conference on Computer Vision

Bonna, S., Huang, Y.C., Novozhilova, E., Paik, S., Shan, Z., Feng, M.Y., Gao, G., Tayal, Y., Kulkarni, R., Yu, J., et al.: Debiaspi: inference-time debiasing by prompt iteration of a text-to-image generative model. In: European Conference on Computer Vision. pp. 68–83. Springer (2024) 2, 3

2024
[4]

In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing text-to-image models using semantic guidance. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 25365–25389. Curran Asso- ciates,Inc.(2023),https://proceedi...

2023
[5]

Persona Vectors: Monitoring and Controlling Character Traits in Language Models

Chen, R., Arditi, A., Sleight, H., Evans, O., Lindsey, J.: Persona vectors: Moni- toring and controlling character traits in language models (2025),https://arxiv. org/abs/2507.215094

work page internal anchor Pith review arXiv 2025
[6]

Conditional fairness for gen- erative ais.arXiv preprint arXiv:2404.16663,

Cheng, C.H., Ruess, H., Wu, C., Zhao, X.: Conditional fairness for generative ais. arXiv preprint arXiv:2404.16663 (2024) 3

work page arXiv 2024
[7]

Debias- ing vision-language models via biased prompts.arXiv preprint arXiv:2302.00070,

Chuang, C.Y., Jampani, V., Li, Y., Torralba, A., Jegelka, S.: Debiasing vision- language models via biased prompts. arXiv preprint arXiv:2302.00070 (2023) 2, 4, 6, 7, 8

work page arXiv 2023
[8]

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

Couairon, G., Douze, M., Cord, M., Schwenk, H.: Embedding arithmetic of mul- timodal queries for image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4950–4958 (2022) 4 14 V. T. Sambandham et al

2022
[9]

In: Forty-first international conference on machine learning (2024) 1, 5, 7

Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024) 1, 5, 7

2024
[10]

AI and Ethics5(3), 2103–2123 (2025) 1, 8

Friedrich, F., Brack, M., Struppek, L., Hintersdorf, D., Schramowski, P., Luccioni, S., Kersting, K.: Auditing and instructing text-to-image generation models on fair- ness. AI and Ethics5(3), 2103–2123 (2025) 1, 8

2025
[11]

In: European Conference on Computer Vision

Gandikota, R., Materzyńska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models. In: European Conference on Computer Vision. pp. 172–188. Springer (2024) 3

2024
[12]

He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016) 6, 17

2016
[13]

Clipscore: A reference-free evaluation metric for image captioning

Hessel,J.,Holtzman,A.,Forbes,M.,LeBras,R.,Choi,Y.:CLIPScore:Areference- free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Meth- ods in Natural Language Processing. pp. 7514–7528. Association for Compu- tational Linguistics, Online and Punta Cana, Dominica...

work page doi:10.18653/v1/2021.emnlp-main.595 2021
[14]

In: The Thirteenth International Conference on Learning Representations (2025) 2, 4, 6, 8

Hirota, Y., Chen, M.H., Wang, C.Y., Nakashima, Y., Wang, Y.C.F., Hachiuma, R.: Saner: Annotation-free societal attribute neutralizer for debiasing clip. In: The Thirteenth International Conference on Learning Representations (2025) 2, 4, 6, 8

2025
[15]

ICLR1(2), 3 (2022) 3

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022) 3

2022
[16]

Fairgen: Controlling sensitive attributes for fair generations in diffusion models via adaptive latent guidance.arXiv preprint arXiv:2503.01872,

Kang, M., Kumar, V.B., Roy, S., Kumar, A., Khosla, S., Narayanaswamy, B.M., Gangadharaiah, R.: Fairgen: Controlling sensitive attributes for fair generations in diffusion models via adaptive latent guidance. arXiv preprint arXiv:2503.01872 (2025) 3

work page arXiv 2025
[17]

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

Karkkainen, K., Joo, J.: Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1548–1558 (2021) 6, 17, 18

2021
[18]

In: 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: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 13361–13370 (2025) 2, 8, 11

2025
[19]

Kim, E., Kim, S., Shin, C., Yoon, S.: De-stereotyping text-to-image models through prompt tuning (2023) 4

2023
[20]

Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convo- lution or region supervision (2021) 19

2021
[21]

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) 6

2024
[22]

Blip: Bootstrapping language- image pre-training for uniﬁed vision-language understanding and generation.arXiv preprint arXiv:2201.12086, 2022

Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation (2022).https://doi. org/10.48550/ARXIV.2201.12086,https://arxiv.org/abs/2201.1208619 Title Suppressed Due to Excessive Length 15

work page doi:10.48550/arxiv.2201.12086 2022
[23]

Advances in Neural Information Processing Systems36, 56338–56351 (2023) 1, 6, 8

Luccioni, S., Akiki, C., Mitchell, M., Jernite, Y.: Stable bias: Evaluating societal representations in diffusion models. Advances in Neural Information Processing Systems36, 56338–56351 (2023) 1, 6, 8

2023
[24]

Versusdebias: Universal zero-shot debiasing for text-to-image models via slm-based prompt engineering and generative adversary.arXiv preprint arXiv:2407.19524,

Luo, H., Deng, Z., Huang, H., Liu, X., Chen, R., Liu, Z.: Versusdebias: Universal zero-shot debiasing for text-to-image models via slm-based prompt engineering and generative adversary. arXiv preprint arXiv:2407.19524 (2024) 2

work page arXiv 2024
[25]

Efficient Estimation of Word Representations in Vector Space

Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre- sentations in vector space. arXiv preprint arXiv:1301.3781 (2013) 2

work page internal anchor Pith review arXiv 2013
[26]

Conditional Generative Adversarial Nets

Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014) 1

work page internal anchor Pith review arXiv 2014
[27]

In: Proceedings of the 25th international academic mindtrek conference

Oppenlaender, J.: The creativity of text-to-image generation. In: Proceedings of the 25th international academic mindtrek conference. pp. 192–202 (2022) 1

2022
[28]

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) 2, 3

2024
[29]

In: 2024 Third International Conference on Smart Technologies and Systems for Next Generation Computing (ICSTSN)

Prerak, S.: Addressing bias in text-to-image generation: A review of mitiga- tion methods. In: 2024 Third International Conference on Smart Technologies and Systems for Next Generation Computing (ICSTSN). pp. 1–6 (2024).https: //doi.org/10.1109/ICSTSN61422.2024.106712302

work page doi:10.1109/icstsn61422.2024.106712302 2024
[30]

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) 5, 18

2021
[31]

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Radford,A.,Metz,L.,Chintala,S.:Unsupervisedrepresentationlearningwithdeep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015) 4, 5

work page internal anchor Pith review arXiv 2015
[32]

Journal of Machine Learning Research21(140), 1–67 (2020), http://jmlr.org/papers/v21/20-074.html5

Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text- to-text transformer. Journal of Machine Learning Research21(140), 1–67 (2020), http://jmlr.org/papers/v21/20-074.html5

2020
[33]

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

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. pp. 10684–10695 (2022) 1

2022
[34]

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

Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815–823 (2015) 17

2015
[35]

A benchmark of facial recognition pipelines and co-usability performances of modules.Journal of Information Technologies, 17(2):95–107, 2024

Serengil, S., Ozpinar, A.: A benchmark of facial recognition pipelines and co- usability performances of modules. Journal of Information Technologies17(2), 95– 107 (2024).https://doi.org/10.17671/gazibtd.1399077,https://dergipark. org.tr/en/pub/gazibtd/issue/84331/139907718

work page doi:10.17671/gazibtd.1399077 2024
[36]

doi: 10.18653/v1/2024.naacl-long.353

Seshadri, P., Singh, S., Elazar, Y.: The bias amplification paradox in text-to- image generation. In: Duh, K., Gomez, H., Bethard, S. (eds.) Proceedings of the 2024 Conference of the North American Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers). pp. 6367–6384. Association for Computational L...

work page doi:10.18653/v1/2024.naacl-long.353 2024
[37]

Finetuning text-to- image diffusion models for fairness.arXiv preprint arXiv:2311.07604, 2023

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) 2, 3 16 V. T. Sambandham et al

work page arXiv 2023
[38]

Very Deep Convolutional Networks for Large-Scale Image Recognition

Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 18

work page internal anchor Pith review Pith/arXiv arXiv 2014
[39]

Wan, Y., Chang, K.W.: The male ceo and the female assistant: Evaluation and mitigation of gender biases in text-to-image generation of dual subjects (2025), https://arxiv.org/abs/2402.110891

work page arXiv 2025
[40]

Wan, Y., Subramonian, A., Ovalle, A., Lin, Z., Suvarna, A., Chance, C., Bansal, H., Pattichis, R., Chang, K.W.: Survey of bias in text-to-image generation: Definition, evaluation, and mitigation (2024),https://arxiv.org/abs/2404.010302, 6

work page arXiv 2024
[41]

Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language (2022), https://arxiv.org/abs/2205.1410019

work page arXiv 2022
[42]

Advances in Neural Information Processing Systems 36, 35331–35349 (2023) 2, 4, 7

Wang, Z., Gui, L., Negrea, J., Veitch, V.: Concept algebra for (score-based) text- controlled generative models. Advances in Neural Information Processing Systems 36, 35331–35349 (2023) 2, 4, 7

2023
[43]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Zhang, C., Chen, X., Chai, S., Wu, C.H., Lagun, D., Beeler, T., De la Torre, F.: Iti-gen: Inclusive text-to-image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3969–3980 (October
[44]

Car- penter

Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stack- gan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In: Proceedings of the IEEE international conference on computer vi- sion. pp. 5907–5915 (2017) 1 Title Suppressed Due to Excessive Length 17 Appendix Table 2: Performance comparison of...

2017