pith. sign in

arxiv: 2606.30262 · v1 · pith:URNHO6CKnew · submitted 2026-06-29 · 💻 cs.CV

Intermediate Text Representation Guided Text-to-Image Generation for Enhancing One-and-Only Alignment

Soyoun Won , Aryan Yazdan Parast , Basim Azam , Jean Honorio , Naveed Akhtar This is my paper

Pith reviewed 2026-06-30 06:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords text-to-image generationdiffusion modelsconcept association biasone-and-only objectsintermediate representationstext encoderOAO-AttackBench
0
0 comments X

The pith

Injecting intermediate hidden states from the text encoder into early denoising steps recovers suppressed concepts for one-and-only objects.

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

The paper argues that text-to-image diffusion models suppress specific details for unique objects because the final text embedding discards concept-level information available in earlier layers of the text encoder. It supports this with an information-theoretic analysis showing reduced mutual information for the denoising process. The proposed fix feeds those intermediate states back into the conditioning signal only during the first denoising steps. This change improves how closely generated images match unusual or conflicting prompts for objects that have a single canonical visual form. The approach requires no model retraining, extra optimization, or external components.

Core claim

The central claim is that the final text embedding discards concept-level information present in the intermediate-layer text representations, and that injecting those intermediate hidden states into the conditioning signal during early denoising steps recovers the suppressed concepts for one-and-only objects without any additional training, optimization, or external models.

What carries the argument

Intermediate Text Representation (IR)-guided diffusion, which injects intermediate hidden states of the text encoder into the conditioning signal during early denoising steps.

Load-bearing premise

The concept-level information present in intermediate-layer text representations but discarded by the final embedding can be recovered and effectively utilized simply by injecting those states during early denoising steps.

What would settle it

A controlled comparison on OAO-AttackBench showing that VQAScore does not increase when intermediate states are injected versus using only the final embedding.

Figures

Figures reproduced from arXiv: 2606.30262 by Aryan Yazdan Parast, Basim Azam, Jean Honorio, Naveed Akhtar, Soyoun Won.

Figure 1
Figure 1. Figure 1: Concept association bias in one-and-only entities. SD3.0 defaults to its learned visual prior for the Earth, ignoring explicitly specified attributes (✗). SD3.0 fails to generate individual concepts, and where it does (yellow, square), it fails to appropriately associate them with the Earth. Our method faithfully reflects all specified concepts and their relationships (✓) without additional training or ext… view at source ↗
Figure 2
Figure 2. Figure 2: (Left) The frozen text encoder produces the final text embedding cf (layer L) and an intermediate hidden state ch (layer k), which carries information suppressed in cf . (Middle) The Post-Processor constructs an IR-guided embedding c˜ = cf + λch, injecting suppressed concept information back into the conditioning signal. (Right) c˜ conditions the early denoising steps (t > t∗ ) where global structure is de… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on SD3.0. (a) Whoops prompts targeting everyday ob￾jects. The baseline generates the individual concepts but fails to express the specified attribute (orange text). (b) OAO-AttackBench prompts targeting one-and-only enti￾ties. The baseline preserves canonical appearances despite explicit contradictions in the prompt; the Mona Lisa remains a painting, the Earth remains spherical, and … view at source ↗
Figure 4
Figure 4. Figure 4: Sample comparison on Gecko. Blue text denotes concepts ignored or misin￾terpreted in the baseline output. IR-guidance improves (a) general concept omission problem beyond one-and-only entities, (b) environmental context, (c) relational sce￾narios involving multiple objects, and (d) long-text prompts. sitional accuracy but also maintains, and in most cases enhances, the overall distribution quality of the g… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of IR-guidance duration. Each row shows the baseline (cf only) and three variants where c˜ is used as conditioning for progressively more denoising steps. The timeline bars at the bottom indicate the conditioning schedule. As IR-guidance is extended to more timesteps, the specified at￾tribute (blue text) is more faithfully ex￾pressed. What does IR-guidance do? To fur￾ther investigate the step-wise i… view at source ↗
Figure 6
Figure 6. Figure 6: (a) Mean squared difference of predicted images between two consecutive de￾noising timesteps. (b) Predicted generated image at each denoising timestep on the baseline (top) and IR-guided diffusion (bottom). IR-guided diffusion delays premature structural convergence, allowing concept exploration during early denoising steps. a steep increase in the mean squared difference at the transition point. Interest￾… view at source ↗
Figure 7
Figure 7. Figure 7: VQAScore vs. CLIPScore of ch from different transformer blocks. Earlier layers yield a higher VQAScore on OAO￾AttackBench. Ablation Study [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
read the original abstract

Text-to-image (T2I) diffusion models often fail to faithfully render explicit textual descriptions, instead defaulting to strongly learned visual priors due to a phenomenon referred to as concept association bias. We show that such bias is particularly strong for one-and-only (OAO) objects, entities that exist in a single canonical form, such as celestial bodies, landmarks, and artworks. The deeply ingrained visual identity for these concepts often resists modification through prompting alone. Addressing this challenge, we first identify through an information-theoretic analysis that the final text embedding discards concept-level information present in the intermediate-layer text representations, reducing the mutual information available to the subsequent denoising process. We then propose Intermediate Text Representation (IR)-guided diffusion, which injects intermediate hidden states of the text encoder into the conditioning signal during early denoising steps, recovering suppressed concepts without any additional training, optimization, or external models. To systematically evaluate the challenging task of aligning generative outputs with unusual prompts for OAO objects, we introduce OAO-AttackBench, a benchmark comprising counterfactual prompts that directly conflict with the core visual identity of OAO objects. Experiments on four benchmarks, including OAO-AttackBench, show that our method achieves up to a 19.1 percentage-point improvement in VQAScore while preserving generation fidelity and human preference. Project page: https://soyoun-won.github.io/one-and-only-ir-guidance/.

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 / 1 minor

Summary. The paper claims that text-to-image diffusion models exhibit strong concept association bias for one-and-only (OAO) objects, which resists modification via prompting. An information-theoretic analysis is used to show that the final text embedding discards concept-level information present in intermediate layers of the text encoder. The proposed IR-guided diffusion method injects these intermediate hidden states into the conditioning signal during early denoising steps, recovering the suppressed concepts in a training-free manner. A new benchmark OAO-AttackBench is introduced with counterfactual prompts, and experiments on four benchmarks report gains of up to 19.1 percentage points in VQAScore while preserving fidelity and human preference.

Significance. If the empirical results hold under rigorous controls, the work is significant for offering a lightweight, parameter-free intervention that leverages internal text-encoder states to improve prompt alignment in existing T2I models without retraining or external components. The training-free aspect and the new OAO-AttackBench benchmark are practical strengths that could aid evaluation of alignment failures on canonical visual concepts.

major comments (2)
  1. [information-theoretic analysis] Information-theoretic analysis (abstract and motivating section): The claim that the final embedding reduces mutual information is central to motivating the injection method, yet no details are provided on the computation (e.g., estimator used, number of samples, specific layers compared, or how concept-level information is quantified). This leaves the analysis as an unsupported assertion rather than a verifiable foundation.
  2. [experiments and OAO-AttackBench] Experimental evaluation (results section): The reported 19.1 pp VQAScore improvement and claims of preserved fidelity lack error bars, statistical significance tests, ablation studies on injection timing/layers, or explicit controls for prompt difficulty. Without these, it is difficult to determine whether the gains are robust or attributable to the proposed mechanism versus other factors.
minor comments (1)
  1. [method] The abstract and method description would benefit from a concise statement of the exact layers selected for injection and the precise early-denoising timestep range, to allow immediate reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point by point below and will revise the manuscript accordingly to improve clarity and rigor.

read point-by-point responses

  1. Referee: [information-theoretic analysis] Information-theoretic analysis (abstract and motivating section): The claim that the final embedding reduces mutual information is central to motivating the injection method, yet no details are provided on the computation (e.g., estimator used, number of samples, specific layers compared, or how concept-level information is quantified). This leaves the analysis as an unsupported assertion rather than a verifiable foundation.

    Authors: We agree that the information-theoretic analysis requires additional implementation details to be fully verifiable. In the revised manuscript we will expand the motivating section to specify the mutual information estimator, the number of samples drawn, the precise text-encoder layers compared, and the procedure used to quantify concept-level information. These additions will substantiate the claim that the final embedding discards intermediate-layer information. revision: yes

  2. Referee: [experiments and OAO-AttackBench] Experimental evaluation (results section): The reported 19.1 pp VQAScore improvement and claims of preserved fidelity lack error bars, statistical significance tests, ablation studies on injection timing/layers, or explicit controls for prompt difficulty. Without these, it is difficult to determine whether the gains are robust or attributable to the proposed mechanism versus other factors.

    Authors: We concur that the experimental section would benefit from greater statistical rigor and ablations. We will add error bars (standard deviation over multiple seeds), statistical significance tests, systematic ablations on injection timing and layer selection, and explicit controls for prompt difficulty. These revisions will clarify that the reported gains are robust and tied to the IR-guided mechanism rather than extraneous factors. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained empirical intervention

full rationale

The paper performs an information-theoretic analysis showing mutual information loss between intermediate and final text embeddings, then directly implements a parameter-free injection of those states into early denoising steps. Neither step reduces to a fitted parameter, self-citation chain, or definitional equivalence; the analysis is presented as original computation, the injection is an explicit architectural modification, and success is measured on external benchmarks (including a new OAO-AttackBench) rather than internal consistency. No equations or claims equate the output to the input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; full manuscript not available, so ledger entries are limited to what is explicitly stated.

axioms (1)
  • domain assumption The final text embedding discards concept-level information present in the intermediate-layer text representations
    Stated as the outcome of the information-theoretic analysis in the abstract.

pith-pipeline@v0.9.1-grok · 5797 in / 1177 out tokens · 31466 ms · 2026-06-30T06:02:39.253015+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

61 extracted references · 20 canonical work pages · 3 internal anchors

  1. [1]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Azam, B., Akhtar, N.: Plug-and-play interpretable responsible text-to-image gen- eration via dual-space multi-facet concept control. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2976–2985 (2025)

    2025

  2. [2]

    In: International Conference on Learning Representations (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. In: International Conference on Learning Representations (2018)

    2018

  3. [3]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision

    Bitton-Guetta, N., Bitton, Y., Hessel, J., Schmidt, L., Elovici, Y., Stanovsky, G., Schwartz, R.: Breaking common sense: Whoops! a vision-and-language benchmark of synthetic and compositional images. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. pp. 2616–2627 (2023)

    2023

  4. [4]

    In: 32nd USENIX security symposium (USENIX Security 23)

    Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX security symposium (USENIX Security 23). pp. 5253–5270 (2023)

    2023

  5. [5]

    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

  6. [6]

    In: International conference on learning representations

    Cho, J., Hu, Y., Baldridge, J., Garg, R., Anderson, P., Krishna, R., Bansal, M., Pont-Tuset, J., Wang, S.: Davidsonian scene graph: Improving reliability in fine- grained evaluation for text-to-image generation. In: International conference on learning representations. vol. 2024, pp. 15625–15645 (2024)

    2024

  7. [7]

    arXiv preprint arXiv:2302.00070 (2023)

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

    work page arXiv 2023

  8. [8]

    Wiley-Interscience, 2 edn

    Cover, T.M., Thomas, J.A.: Elements of Information Theory. Wiley-Interscience, 2 edn. (2006)

    2006

  9. [9]

    IEEE transactions on pattern anal- ysis and machine intelligence (2024)

    Delmas, G., Weinzaepfel, P., Lucas, T., Moreno-Noguer, F., Rogez, G.: Posescript: Linking 3d human poses and natural language. IEEE transactions on pattern anal- ysis and machine intelligence (2024)

    2024

  10. [10]

    Journal of the American Statistical Association106(496), 1602–1614 (2011)

    Efron, B.: Tweedie’s formula and selection bias. Journal of the American Statistical Association106(496), 1602–1614 (2011)

    2011

  11. [11]

    In: Forty-first international conference on machine learning (2024) 16 S

    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) 16 S. Won et al

    2024

  12. [12]

    arXiv preprint arXiv:2302.10893 (2023)

    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

  13. [13]

    Fu, X., He, M., Lu, Y., Wang, W.Y., Roth, D.: Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546 (2024)

    work page arXiv 2024

  14. [14]

    arXiv preprint arXiv:2412.03400 (2024)

    He, F., Zhang, C., Zhao, Z.: Implicit priors editing in stable diffusion via targeted token adjustment. arXiv preprint arXiv:2412.03400 (2024)

    work page arXiv 2024

  15. [15]

    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

  16. [16]

    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

  17. [17]

    Advances in Neural Information Processing Systems37, 137646–137672 (2024)

    Hu, T., Li, L., Van de Weijer, J., Gao, H., Shahbaz Khan, F., Yang, J., Cheng, M.M., Wang, K., Wang, Y.: Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems37, 137646–137672 (2024)

    2024

  18. [18]

    ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment

    Hu, X., Wang, R., Fang, Y., Fu, B., Cheng, P., Yu, G.: Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

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

    Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question an- swering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20406–20417 (2023)

    2023

  20. [20]

    In: Advances in Neural Information Processing Systems

    Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In: Advances in Neural Information Processing Systems. vol. 36 (2023)

    2023

  21. [21]

    arXiv preprint arXiv:2506.01929 (2025)

    Huberman, S., Patashnik, O., Dahary, O., Mokady, R., Cohen-Or, D.: Image gen- eration from contextually-contradictory prompts. arXiv preprint arXiv:2506.01929 (2025)

    work page arXiv 2025

  22. [22]

    arXiv preprint arXiv:2503.10687 (2025)

    Islam, K., Akhtar, N.: Context-guided responsible data augmentation with diffu- sion models. arXiv preprint arXiv:2503.10687 (2025)

    work page arXiv 2025

  23. [23]

    Expert Systems with Applications p

    Islam, K., Zaheer, M.Z., Mahmood, A., Nandakumar, K., Akhtar, N.: Genmix: effective data augmentation with generative diffusion model image editing. Expert Systems with Applications p. 132273 (2026)

    2026

  24. [24]

    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)

    2025

  25. [25]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    Kim, J., Esmaeili, E., Qiu, Q.: Text embedding is not all you need: Attention con- trol for text-to-image semantic alignment with text self-attention maps. In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 8031– 8040 (2025)

    2025

  26. [26]

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

    Krause, J., Johnson, J., Krishna, R., Fei-Fei, L.: A hierarchical approach for gen- erating descriptive image paragraphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 317–325 (2017)

    2017

  27. [27]

    Li, J., Hu, L., Zhang, J., Zheng, T., Zhang, H., Wang, D.: Fair text-to-image diffusion via fair mapping (2024),https://arxiv.org/abs/2311.17695

    work page arXiv 2024

  28. [28]

    Li, Y., Salehi, S., Ungar, L., Kording, K.P.: Does object binding naturally emerge in large pretrained vision transformers? arXiv preprint arXiv:2510.24709 (2025) Intermediate Text Representation Guided One-and-Only Alignment 17

    work page arXiv 2025

  29. [29]

    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

  30. [30]

    In: European Conference on Computer Vision

    Lin, Z., Pathak, D., Li, B., Li, J., Xia, X., Neubig, G., Zhang, P., Ramanan, D.: Evaluating text-to-visual generation with image-to-text generation. In: European Conference on Computer Vision. pp. 366–384. Springer (2024)

    2024

  31. [31]

    In: Proceedings of the 61st Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers)

    Liu, R., Garrette, D., Saharia, C., Chan, W., Roberts, A., Narang, S., Blok, I., Mical, R., Norouzi, M., Constant, N.: Character-aware models improve visual text rendering. In: Proceedings of the 61st Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers). pp. 16270–16297 (2023)

    2023

  32. [32]

    Lu, C., Krishna, R., Bernstein, M., Fei-Fei, L.: Visual relationship detection with languagepriors.In:Europeanconferenceoncomputervision.pp.852–869.Springer (2016)

    2016

  33. [33]

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

    Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7030–7038 (2023)

    2023

  34. [34]

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

    Paiss, R., Ephrat, A., Tov, O., Zada, S., Mosseri, I., Irani, M., Dekel, T.: Teaching clip to count to ten. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3170–3180 (2023)

    2023

  35. [35]

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

    Parast, A.Y., Azam, B., Akhtar, N.: Ddb: Diffusion driven balancing to address spurious correlations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17526–17535 (2025)

    2025

  36. [36]

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

    Phung, Q., Ge, S., Huang, J.B.: Grounded text-to-image synthesis with attention refocusing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7932–7942 (2024)

    2024

  37. [37]

    Pont-Tuset, J., Uijlings, J., Changpinyo, S., Soricut, R., Ferrari, V.: Connecting visionandlanguagewithlocalizednarratives.In:Europeanconferenceoncomputer vision. pp. 647–664. Springer (2020)

    2020

  38. [38]

    Advances in Neural Information Processing Systems36 (2024)

    Rassin, R., Hirsch, E., Glickman, D., Ravfogel, S., Goldberg, Y., Chechik, G.: Lin- guistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems36 (2024)

    2024

  39. [39]

    In: Breakthroughs in Statistics: Foundations and basic theory, pp

    Robbins, H.E.: An empirical bayes approach to statistics. In: Breakthroughs in Statistics: Foundations and basic theory, pp. 388–394. Springer (1992)

    1992

  40. [40]

    In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing

    Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object halluci- nation in image captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 4035–4045 (2018)

    2018

  41. [41]

    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)

    2022

  42. [42]

    In: Proceedings of the Winter Confer- ence on Applications of Computer Vision

    Sajnani, R., Vanbaar, J., Min, J., Katyal, K.D., Sridhar, S.: Geodiffuser: Geometry- based image editing with diffusion models. In: Proceedings of the Winter Confer- ence on Applications of Computer Vision. pp. 472–482 (2025)

    2025

  43. [43]

    arXiv preprint arXiv:2503.23011 (2025)

    Seo, H., Bang, J., Lee, H., Lee, J., Lee, B.H., Chun, S.Y.: Geometrical properties of text token embeddings for strong semantic binding in text-to-image generation. arXiv preprint arXiv:2503.23011 (2025)

    work page arXiv 2025

  44. [44]

    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) 18 S. Won et al

    2024

  45. [45]

    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)

    work page arXiv 2023

  46. [46]

    Sun, J., Fu, D., Hu, Y., Wang, S., Rassin, R., Juan, D.C., Alon, D., Herrmann, C., Van Steenkiste, S., Krishna, R., et al.: Dreamsync: Aligning text-to-image gener- ation with image understanding feedback. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lin- guistics: Human Language Techno...

    2025

  47. [47]

    In: Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing

    Tang, Y., Yamada, Y., Zhang, Y., Yildirim, I.: When are lemons purple? the con- cept association bias of vision-language models. In: Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Processing. pp. 14333–14348 (2023)

    2023

  48. [48]

    Diffusion lens: Interpreting text encoders in text-to- image pipelines.arXiv preprint arXiv:2403.05846,

    Toker, M., Orgad, H., Ventura, M., Arad, D., Belinkov, Y.: Diffusion lens: Inter- preting text encoders in text-to-image pipelines. arXiv preprint arXiv:2403.05846 (2024)

    work page arXiv 2024

  49. [49]

    arXiv preprint arXiv:2404.13766 (2024)

    Trusca, M.M., Nuyts, W., Thomm, J., Honig, R., Hofmann, T., Tuytelaars, T., Moens, M.F.: Object-attribute binding in text-to-image generation: Evaluation and control. arXiv preprint arXiv:2404.13766 (2024)

    work page arXiv 2024

  50. [50]

    Turc, I., Nemade, G.: Midjourney user prompts & generated images (250k) (2023)

    2023

  51. [51]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Um, S., Ye, J.C.: Minority-focused text-to-image generation via prompt optimiza- tion. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20926–20936 (2025)

    2025

  52. [52]

    Artificial Intelligence Review59(2), 57 (2026)

    Vice, J., Akhtar, N., Sigal, L., Hartley, R., Mian, A.: On the fairness, diversity and reliability of text-to-image generative models. Artificial Intelligence Review59(2), 57 (2026)

    2026

  53. [53]

    arXiv preprint arXiv:2303.02490 (2023)

    Wang, B., Vastola, J.J.: Diffusion models generate images like painters: an analyt- ical theory of outline first, details later. arXiv preprint arXiv:2303.02490 (2023)

    work page arXiv 2023

  54. [54]

    In: Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers)

    Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffu- siondb: A large-scale prompt gallery dataset for text-to-image generative models. In: Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers). pp. 893–911 (2023)

    2023

  55. [55]

    arXiv preprint arXiv:2404.16820 (2024)

    Wiles, O., Zhang, C., Albuquerque, I., Kajić, I., Wang, S., Bugliarello, E., Onoe, Y., Papalampidi, P., Ktena, I., Knutsen, C., et al.: Revisiting text-to-image eval- uation with gecko: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404.16820 (2024)

    work page arXiv 2024

  56. [56]

    arXiv preprint arXiv:2508.06044 (2025)

    Wu, H., Ma, X., Zhao, H., Zhao, Y., Li, Q.: Nep: Autoregressive image editing via next editing token prediction. arXiv preprint arXiv:2508.06044 (2025)

    work page arXiv 2025

  57. [57]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  58. [58]

    arXiv preprint arXiv:2403.02118 (2024)

    Yang, Y., Lin, Y., Liu, H., Shao, W., Chen, R., Shang, H., Wang, Y., Qiao, Y., Zhang, K., Luo, P.: Position: towards implicit prompt for text-to-image models. arXiv preprint arXiv:2403.02118 (2024)

    work page arXiv 2024

  59. [59]

    Planet Earth is visually defined by purple ocean and glowing continents

    Zhang, H., Guo, Y., Kankanhalli, M.: Joint vision-language social bias removal for clip. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 4246–4255 (2025) Intermediate Text Representation Guided Text-to-Image Generation for Enhancing One-and-Only Alignment (Supplementary Material) Soyoun Won1 , Aryan Yazdan Parast1 , Basim Az...

    2025

  60. [60]

    Identify two global characteristics in the following subcategory: shape and material

  61. [61]

    Planet Earth is visually defined by purple ocean and glowing continents

    Replace each attribute with a visually or functionally unrelated attribute that contradicts or violates the core identity. Output format: [subcategory], [replaced attribute] The replaced attribute must be applicable to the global characteristics of the given category of objects. Category: Celestial objects (e.g., The Sun, The Earth, Mercury, ...) D Implem...

    work page arXiv