Discovering Divergent Representations between Text-to-Image Models

Bryan Russell; Fabian Caba Heilbron; Josef Sivic; Joseph E. Gonzalez; Lisa Dunlap; Trevor Darrell

REVIEW 1 major objections 1 minor 54 references

Discovering Divergent Representations between Text-to-Image Models

T0 review · 1 major / 1 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CompCon discovers when two text-to-image models diverge on the same prompt, finding the visual attribute one model adds and the prompt concepts that trigger it.

desk verdict CompCon is a genuinely new, honestly-reported method for discovering prompt-dependent visual differences between T2I models; the main caveats are the synthetic benchmark and a CLIP-based scorer that the authors themselves show can misfire. read the letter →

arxiv 2509.08940 v1 pith:CIQZBRW3 submitted 2025-09-10 cs.CV

Lisa Dunlap , Joseph E. Gonzalez , Trevor Darrell , Fabian Caba Heilbron , Josef Sivic , Bryan Russell This is my paper

classification cs.CV

keywords text-to-imagemodelsmodelcomparisondivergencediscoveryevolutionarysearchCLIPscoringbiasdetectionpromptdescriptionbenchmark

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that systematic, input-dependent differences between two text-to-image models can be discovered automatically as pairs of a visual attribute (e.g., “flames”) and a prompt description (e.g., “strong emotions”). It introduces CompCon, an evolutionary algorithm that proposes candidate attributes with a VLM, scores them with CLIP, then iteratively refines prompt descriptions until they reliably trigger the difference. The authors build a benchmark, ID2, with 60 known divergences, and show CompCon outperforms LLM-only, TF-IDF, and VisDiff baselines. Applied to PixArt and SD-Lightning, it uncovers concrete divergences and biases that single-model auditing misses. The method gives developers and users a principled way to ask “what does this model see that the other doesn’t?”

What carries the argument

The divergence score z(a,I1(p),I2(p)) = [s(a,I1)>t] × [s(a,I1)−s(a,I2)>δ] — a CLIP cosine-similarity test that the attribute is present above threshold t in one model’s images and exceeds the other model’s by margin δ. This score classifies prompts as diverging or non-diverging; the evolutionary search’s prompt bank H and LLM mutation loop maximizes the expected score over newly generated prompts. The central mechanism is this classifier-in-the-loop search that couples visual attribute discovery with prompt-description refinement.

What would settle it

Take any discovered divergence and ask humans to classify 100 new prompts from the description: if the attribute is not present in a clear majority of one model’s images and absent from the other’s, the divergence score is not tracking real visual content. The paper’s own Figure 13 already shows examples where CLIP mislabels prompts as diverging when both or neither image contains the attribute.

Watch

Extended reading notes

Core claim

CompCon models the task as learning a mapping F from text prompts to diverging visual attributes for a given model pair. It first uses a vision-language model to surface candidate attributes from image grids, ranks them with a CLIP-based divergence score requiring one model’s image set to contain the attribute and the other’s not, then runs an evolutionary search: an LLM writes a description of diverging vs. non-diverging prompts, new prompts are generated or retrieved from that description, and the description that maximizes the fraction of confirmed diverging prompts is kept. The authors claim this finds input-dependent differences such as PixArt rendering emotion prompts with flames while

Load-bearing premise

The CLIP-based divergence score—requiring attribute presence above a threshold in one model’s images and a margin over the other—correctly identifies when a generated image really contains the attribute, so that prompts labeled diverging are genuinely diverging.

Editorial extensions

If this is right

CompCon can be applied to any pair of text-to-image models to produce interpretable, natural-language descriptions of where they differ, not just aggregate quality scores.
Discovered divergences can reveal biases—e.g., SD3.5 generating African American people for media professions, PixArt showing old men for traditional trades—giving developers a targeted list to audit.
The iterative search improves divergence hit-rate substantially (e.g., wet streets from 8% to 52% of generated prompts), so the method is tunable in cost vs. precision.
The ID2 benchmark and LLM-as-judge evaluation (human-validated with comparable kappa) provide a reusable standard for future divergence-discovery methods.
Because the pipeline works with smaller open-source VLMs/LLMs (still beating baselines), it is reproducible without proprietary models.

Reading between the lines

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

The same search loop could be adapted to compare video generators, audio generators, or any conditional generative model pair where a CLIP-like scorer exists, not just images.
Discovered divergences could serve as a debugging signal during training: if a new checkpoint diverges from its predecessor on a known attribute, that points to a data or architecture change to inspect.
The method could be turned around to generate prompts that intentionally split models—useful for building ‘trap’ prompts to stress-test alignment or for watermarking model outputs.
If the Platonic representation hypothesis is right, the set of divergences between two models might shrink as both improve; CompCon’s outputs could serve as a direct empirical measure of representational convergence.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read the arXiv version. My take: this is a useful, honestly-written paper that does something new—it moves attribute discovery between image sets from static to input-dependent, using evolutionary search to find prompt descriptions that trigger visual differences between two T2I models. The qualitative results (flames for strong emotion, wet streets for bleakness) are plausible, and the paper includes a real benchmark, code, open-model results, and human validation. Credit where due: the ID2 dataset, despite its synthetic origins, is a sensible way to get ground-truth divergences, and the human agreement data (kappa 0.635 vs 0.667) suggests the LLM judge is not wildly off.

The soft spots are real but not fatal. First, Table 1 reports single-run scores with no error bars or significance tests; given the variance in LLM/VLM outputs, the claimed margins (attribute 0.60 vs 0.47) might not be stable. Second, the divergence score in Eq. (2) uses hand-tuned thresholds, and the paper itself (Section F, Fig. 13) shows CLIP misclassifies prompts when both or neither image has the attribute. That is a load-bearing weakness because the evolutionary search optimizes exactly this score; it could in principle converge to descriptions that exploit CLIP's blind spots rather than true visual differences. The robustness experiments (error injection, hallucinated attributes) help but don't fully close the gap, since they don't test z's accuracy on real attributes at the operating points used. Third, ID2 was generated with GPT-4o, the same family used for CompCon's mutations and judge; the human validation mitigates but doesn't eliminate the concern that the benchmark rewards GPT-4o-like phrasing.

That said, the paper is transparent about these issues, and the central contribution stands: input-dependent divergence discovery is a real task, the method is a reasonable first solution, and the qualitative findings suggest it works on real model pairs. The stress-test's worry about CLIP artifacts is worth taking seriously, but I don't think it sinks the paper.

Who is it for: anyone building or auditing T2I models, and researchers working on interpretability or bias discovery. I'd bring it to a reading group and cite it. Recommendation: send to peer review, but with a request for error bars and either a more robust scorer or a per-attribute validation of the CLIP thresholds.

Referee Report

1 major / 1 minor

Summary. The paper introduces CompCon, an evolutionary search algorithm that, given a pair of text-to-image models, discovers pairs of a diverging visual attribute (e.g., 'flames') and a diverging prompt description (e.g., 'prompts with strong emotions') such that prompts following the description cause one model to depict the attribute while the other does not. The pipeline first uses a VLM to propose candidate visual attributes, ranks them with a CLIP-based divergence score (Eq. 2), then iteratively refines a natural-language prompt description by generating candidate prompts, classifying them as diverging/non-diverging with the same CLIP score, and returning the description with the highest diverging-prompt proportion. The authors introduce ID2, a simulated benchmark of 60 input-dependent differences, and evaluate CompCon against LLM-only, TF-IDF, and VisDiff baselines on attribute and description scores. They also present qualitative comparisons of PixArt vs. SDXL-Lightning, including discoveries such as 'wet streets' and bias-related attributes such as 'African American people' for media-profession prompts.

Significance. If the reported results hold, the paper makes a useful contribution: it defines a concrete discovery task for input-dependent representational differences between generative models, ships a benchmark (ID2), and provides an end-to-end method with human validation of both the benchmark and the LLM judge. The appendices are unusually thorough (hyperparameters, prompts, open-model variants, error-sensitivity experiments), and the code is released. The human validation of ID2 and the LLM judge, plus the open-model variant in Table 4, are real strengths. The central claim, however, is that CompCon outperforms the baselines on ID2; that claim is currently supported by a single run of scores whose statistical reliability and benchmark-referee independence are not established.

major comments (1)

[§5.1] The thresholds t and δ in Eq. 2 are free hyperparameters whose values are not derived. For the qualitative model comparison, the paper states the authors 'manually inspect prompts labeled as diverging to set thresholds t and δ' per discovered representation. This means the reported discoveries (e.g., 'flames', 'wet streets') may reflect thresholds adjusted after seeing the outputs of a specific run, undermining the predictive claim of the method. Please report a sensitivity analysis over t and δ, or use a pre-registered thresholding rule, for the qualitative experiments.

minor comments (1)

[§C.3] Early stopping at max-average-divergence 0.1 within 5 iterations is an additional free parameter; please clarify whether this threshold was fixed before the eval runs or tuned on ID2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the discovery algorithm and benchmark evaluation are not equivalent to the algorithm's inputs.

full rationale

CompCon's core derivation is self-contained. The diverging visual attributes are proposed by a VLM from image grids and then ranked by the CLIP-based divergence score z in Eq. (2); the prompt-description search maximizes the expectation of z in Eq. (3). Nothing in these equations is fitted to the ID2 ground-truth labels, and the reported attribute/description scores are computed by an LLM judge against independently created ground-truth tuples, not by the CLIP scorer used in the search. The hand-set thresholds t and delta and the documented CLIP misclassifications (Fig. 13) are empirical reliability limitations, not circular reductions. The use of GPT-4o for both generation and judging is a benchmark-design concern, but it is not a derivation that reduces to its own inputs; moreover, the paper includes human validation (weighted kappa 0.635 vs. inter-human 0.667) and open-model results. Self-citation of VisDiff (Dunlap et al.) is only as a baseline and does not carry any load-bearing premise. No equation or selection rule is equivalent by construction to the claimed discoveries, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method depends on hand-set CLIP thresholds, sampling sizes, and the reliability of off-the-shelf VLMs and LLMs. No new physical entities are introduced. The synthetic ID2 benchmark is a domain assumption whose validity for real model divergences is untested beyond qualitative examples.

free parameters (5)

divergence threshold t = 0 for discovery, 0.2 for benchmark, manually tuned for qualitative runs
Threshold for CLIP similarity to attribute; chosen by hand in Section 5.1.
margin threshold delta = 0.05
CLIP score margin; manually set in Section 5.1.
k candidate prompts per iteration = 25
Number of new prompts generated by the LLM each iteration; fixed.
B sampled prompts = 25
Number of prompts sampled from diverging and non-diverging banks for description mutation.
early stopping threshold = 0.1 within 5 iterations
Jobs not reaching average divergence score above 0.1 are killed; Section C.3.

assumptions (4)

domain assumption CLIP cosine similarity between attribute text and image is a valid indicator of attribute presence.
Equations (2) and (4) use this to label prompts as diverging; Section F documents misclassification cases.
domain assumption ID2's simulated differences via a single model with injected attributes represent real model divergences.
The benchmark uses SD-3.5-Turbo prompt pairs differing by attribute mention, which may not cover implicit, training-driven divergences.
domain assumption GPT-4o as LLM judge yields scores that match human judgment.
Cohen's kappa 0.635 vs humans, comparable to inter-human 0.667, but moderate.
domain assumption Off-the-shelf VLM reliably lists visual attributes from image grids.
Section 3.1; authors test robustness to errors but rely on VLM quality.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Discovering Divergent Representations between Text-to-Image Models." pith.science (2026). https://pith.science/paper/CIQZBRW3

@misc{pith2026250908940,
  author       = {Pith},
  title        = {Pith review of: Discovering Divergent Representations between Text-to-Image Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIQZBRW3}},
  note         = {Machine review of arXiv:2509.08940}
}

read the original abstract

In this paper, we investigate when and how visual representations learned by two different generative models diverge. Given two text-to-image models, our goal is to discover visual attributes that appear in images generated by one model but not the other, along with the types of prompts that trigger these attribute differences. For example, "flames" might appear in one model's outputs when given prompts expressing strong emotions, while the other model does not produce this attribute given the same prompts. We introduce CompCon (Comparing Concepts), an evolutionary search algorithm that discovers visual attributes more prevalent in one model's output than the other, and uncovers the prompt concepts linked to these visual differences. To evaluate CompCon's ability to find diverging representations, we create an automated data generation pipeline to produce ID2, a dataset of 60 input-dependent differences, and compare our approach to several LLM- and VLM-powered baselines. Finally, we use CompCon to compare popular text-to-image models, finding divergent representations such as how PixArt depicts prompts mentioning loneliness with wet streets and Stable Diffusion 3.5 depicts African American people in media professions. Code at: https://github.com/adobe-research/CompCon

Figures

Figures reproduced from arXiv: 2509.08940 by the authors.

**Figure 1.** Discovering divergent representations with COMPCON. Left: COMPCON takes as input a pair of text-to-image models and outputs a diverging prompt description to produce a diverging visual attribute appearing in one model but not the other. Right: We show the discovered diverging visual attribute ‘flames’ appearing in PixArt but not SDXL-Lightning over different diverging prompts. Abstract In this paper, we investigate … view at source ↗

**Figure 2.** COMPCON overview. We illustrate our approach for discovering diverging visual attributes (top) and diverging prompt descriptions (bottom). Given two text-to-image models and a set of prompts, we use a VLM to identify visual differences. For each diverging attribute, we iteratively refine diverging prompt descriptions by generating candidate prompts P can a from the description, classifying them as diverging (Hdiv) … view at source ↗

**Figure 3.** ID2 creation. Given a diverging prompt description dpa and diverging visual attribute a, we use an LLM to generate prompt pairs where one of the prompts mentions the diverging visual attribute. Both prompts are then passed to the same text-toimage model to generate image pairs with the visual difference a. exhibit controlled, systematic differences. This simulation approach allows us to create ground truth data wi… view at source ↗

Figures from the paper (11 more)

**Figure 4.** Figure 4: ID2 example. Top: We show dataset prompts and corresponding generated images, where the second image row depicts the diverging visual attribute. Bottom: We show the ground truth diverging visual attribute and diverging prompt description, along with outputs from our a…

**Figure 5.** Figure 5: COMPCON results comparing PixArt and SD-Lightning. PixArt associates negative emotions / desolation in urban environments with ‘wet streets,’ while SD-Lightning struggles to depict run-down or dilapidated scenes, where PixArt instead conveys ‘decay.’ A person who work…

**Figure 7.** Figure 7: Further qualitative results comparing PixArt-Alpha to SDXL-Lightning using a templated prompt bank. Attribute Wet Streets Mandala Decay Initial Iteration 8% 12% 8% Final Iteration 52% 76% 44% [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Qualitative results comparing PixArt-Alpha to SDXL-Lightning using an LLM generated initial prompt bank. 3 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Finding bias. COMPCON discovers age bias present in PixArt images for prompts about traditional professions. Digital art of an enraged opera singer PixArt SD-Lightning A woman who serves as an academic dean Playground 2.5 Dreamlike 2.0 Unique to PixArt and Playground: …

**Figure 10.** Figure 10: Diverging attributes discovered by CompCon across [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

**Figure 11.** Figure 11: Example image grid input to VLM during the diverging [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: Example input-dependent difference in ID [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

**Figure 13.** Figure 13: Example of CLIP classifying prompts as diverging [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

**Figure 14.** Figure 14: Finding bias in PixArt-Alpha and SDXL-Lightning. 11 [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]

**Figure 15.** Figure 15: COMPCON results comparing PixArt and SD-Lightning over initial and final iterations. Our evolutionary search improves results over the initial iteration, where the diverging prompt description induces diverging prompts that cause one model to generate the diverging vi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 8 linked inside Pith

[1]

Introducing stable diffusion 3.5.https: //stability.ai/news/introducing- stable- diffusion-3-5, 2024

Stability AI. Introducing stable diffusion 3.5.https: //stability.ai/news/introducing- stable- diffusion-3-5, 2024. Stable Diffusion 3.5 Large, Large Turbo, and Medium models. 5, 7

2024
[2]

Llama 3 model card

AI@Meta. Llama 3 model card. 2024. 2

2024
[3]

Claude 3.5 sonnet, 2024

Anthropic. Claude 3.5 sonnet, 2024. 5

2024
[4]

D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba. Network dissection: Quantifying interpretability of deep vi- sual representations. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6541–6549, 2017. 2

2017
[5]

Easily acces- sible text-to-image generation amplifies demographic stereo- types at large scale

Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily acces- sible text-to-image generation amplifies demographic stereo- types at large scale. InProceedings of the 2023 ACM Confer- ence on Fairness, Accountability, and Transparency (FAccT),

2023
[6]

Pixart-α: Fast training of dif- fusion transformer for photorealistic text-to-image synthesis,

Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of dif- fusion transformer for photorealistic text-to-image synthesis,
[7]

Reproducible scal- ing laws for contrastive language-image learning

Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scal- ing laws for contrastive language-image learning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818–2829, 2023. 6

2023
[8]

Tibet: Identifying and evaluating biases in text-to-image generative models, 2023

Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models, 2023. 2

2023

Show all 54 references

[9]

Evolving interpretable visual classifiers with large language models

Mia Chiquier, Utkarsh Mall, and Carl V ondrick. Evolving interpretable visual classifiers with large language models. InECCV, 2024. 2

2024
[10]

Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers.arXiv preprint arXiv:2202.04053, 2022

Jaemin Cho, Maarten Sap, Mark Yatskar, Yejin Choi, and Dan Schwartz. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers.arXiv preprint arXiv:2202.04053, 2022. 2

2022 arXiv
[11]

J. Cho, A. Zala, and M. Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image genera- tion models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3043–3054,
[12]

Weighted kappa: Nominal scale agreement provision for scaled disagreement or partial credit.Psycho- logical Bulletin, 70(4):213–220, 1968

Jacob Cohen. Weighted kappa: Nominal scale agreement provision for scaled disagreement or partial credit.Psycho- logical Bulletin, 70(4):213–220, 1968. 6

1968
[13]

Openbias: Open-set bias detection in text-to-image generative models

Moreno D’Inc `a, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

2024
[14]

Gradbias: Un- veiling word influence on bias in text-to-image generative models, 2024

Moreno D’Inc `a, Elia Peruzzo, Massimiliano Mancini, Xingqian Xu, Humphrey Shi, and Nicu Sebe. Gradbias: Un- veiling word influence on bias in text-to-image generative models, 2024. 2

2024
[15]

Efros, and Assaf Shocher

Amil Dravid, Yossi Gandelsman, Alexei A. Efros, and Assaf Shocher. Rosetta neurons: Mining the common units in a model zoo. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1934–1943,

1934
[16]

Dreamlike photoreal 2.0.https:// huggingface.co/dreamlike- art/dreamlike- photoreal-2.0, 2023

Dreamlike Art. Dreamlike photoreal 2.0.https:// huggingface.co/dreamlike- art/dreamlike- photoreal-2.0, 2023. A photorealistic Stable Diffusion 1.5–based model trained on 768×768 images. 1

2023
[17]

Gonzalez, and Serena Yeung-Levy

Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E. Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. InConference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 7

2024
[18]

Friedrich, P

F. Friedrich, P. Schramowski, M. Brack, L. Struppek, D. Hin- tersdorf, S. Luccioni, and K. Kersting. Fair diffusion: In- structing text-to-image generation models on fairness.arXiv preprint arXiv:2302.10893, 2023. 2

2023 arXiv
[19]

Efros, and Jacob Steinhardt

Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting clip’s image representation via text-based de- composition, 2023. 2

2023
[20]

Efros, and Jacob Steinhardt

Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting the second-order effects of neurons in clip, 2024. 2

2024
[21]

Ghosh and A

S. Ghosh and A. Caliskan. Person==light-skinned, western man, and sexualization of women of color: Stereotypes in stable diffusion. InFindings of the Association for Compu- tational Linguistics: EMNLP 2023, pages 6971–6985, 2023. 2

2023
[22]

Tiam - a metric for evaluating alignment in text- to-image generation

Paul Grimal, Herv ´e Le Borgne, Olivier Ferret, and Julien Tourille. Tiam - a metric for evaluating alignment in text- to-image generation. InProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 2890–2899, 2024. 2

2024
[23]

Hamidieh, H

K. Hamidieh, H. Zhang, T. Hartvigsen, and M. Ghassemi. Identifying implicit social biases in vision-language models. arXiv preprint arXiv:2411.00997, 2023. 2

2023 arXiv
[24]

Hessel, A

J. Hessel, A. Holtzman, M. Forbes, R.L. Bras, and Y . Choi. Clipscore: A reference-free evaluation metric for image cap- tioning.arXiv preprint arXiv:2104.08718, 2021. 2

2021 arXiv
[25]

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

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. InAdvances in Neural Information Processing Systems (NeurIPS), pages 6626–6637, 2017. 2

2017
[26]

Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering.arXiv preprint arXiv:2303.11897, 2023

Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Os- tendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering.arXiv preprint arXiv:2303.11897, 2023. 2

2023 arXiv
[27]

The platonic representation hypothesis

Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. InICML,
[28]

Open- clip, 2021

Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, 9 Vaishaal Shankar, Hongseok Namkoong, John Miller, Han- naneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Open- clip, 2021. If you use this software, please cite it as below. 6

2021
[29]

Building and better understanding vision- language models: insights and future directions., 2024

Hugo Laurenc ¸on, Andr´es Marafioti, Victor Sanh, and L ´eo Tronchon. Building and better understanding vision- language models: insights and future directions., 2024. 8, 2

2024
[30]

Holis- tic evaluation of text-to-image models

Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Min- guk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. Holis- ti...

2023
[31]

Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image genera- tion, 2024

Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Lin- miao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image genera- tion, 2024. 7, 1

2024
[32]

Sdxl- lightning: Progressive adversarial diffusion distillation,

Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxl- lightning: Progressive adversarial diffusion distillation,
[33]

Visual instruction tuning

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InNeurIPS, 2023. 7

2023
[34]

Organizing unstructured im- age collections using natural language.arXiv preprint arXiv:2410.05217, 2024

Mingxuan Liu, Zhun Zhong, Jun Li, Gianni Franchi, Sub- hankar Roy, and Elisa Ricci. Organizing unstructured im- age collections using natural language.arXiv preprint arXiv:2410.05217, 2024. 2

2024 arXiv
[35]

Stable bias: Analyzing soci- etal representations in diffusion models

Alexandra Sasha Luccioni, Cynthia Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing soci- etal representations in diffusion models. InProceedings of the 37th Conference on Neural Information Processing Sys- tems (NeurIPS), 2023. 8, 1

2023
[36]

Hello gpt-4o.https : / / openai

OpenAI. Hello gpt-4o.https : / / openai . com / index / hello - gpt - 4o/, 2024. (Accessed on 06/05/2024). 6

2024
[37]

Benchmark for compositional text- to-image synthesis

Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Dar- rell, and Anna Rohrbach. Benchmark for compositional text- to-image synthesis. InProceedings of the Neural Informa- tion Processing Systems (NeurIPS). University of California, Berkeley, 2024. 2

2024
[38]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the...

2021
[39]

Photorealistic text-to-image diffusion models with deep language understanding.arXiv preprint arXiv:2205.11487, 2022

Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding.arXiv preprint arXiv:...

2022 arXiv
[40]

Improved techniques for training gans

Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), pages 2234–2242, 2016. 2

2016
[41]

LAION-5b: An open large-scale dataset for train- ing next generation image-text models

Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev....

2022
[42]

A statistical interpretation of term specificity and its application in retrieval.Journal of Doc- umentation, 28(1):11–21, 1972

Karen Sp ¨arck Jones. A statistical interpretation of term specificity and its application in retrieval.Journal of Doc- umentation, 28(1):11–21, 1972. 6

1972
[43]

Smith, Luke Zettlemoyer, and Tao Yu

Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instruction-finetuned text embeddings. 2022. 6

2022
[44]

Mass- producing failures of multimodal systems with language models, 2023

Shengbang Tong, Erik Jones, and Jacob Steinhardt. Mass- producing failures of multimodal systems with language models, 2023. 2

2023
[45]

T2iat: Measuring valence and stereotypical biases in text-to-image generation, 2023

Jialu Wang, Xinyue Gabby Liu, Zonglin Di, Yang Liu, and Xin Eric Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation, 2023. 2

2023
[46]

To- wards fairness in visual recognition: Effective strategies for bias mitigation

Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. To- wards fairness in visual recognition: Effective strategies for bias mitigation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

2020
[47]

Menacing appearance

Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gun- jan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin- fei Yang, Burcu Karagol Ayan, et al. Scaling autoregres- sive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2022. 2 10 Discover...

2022 arXiv
[48]

PROMPT 2 ... C.2. Sampling As detailed in Section 3, we randomly sampleBdiverg- ing and non-diverging prompts from the prompt bank to create our diverging prompt description. In the initial it- eration (i= 0), this sampling is entirely random. For subsequent iterations (i+ 1),...
[49]

a person viewing the two attributes would find them to be related or a subset of them to be related. 5
[50]

- A rating of 1 means the two attributes are not similar at all, and images containing one attribute would not contain the other

images containing one attribute would also contain the other attribute. - A rating of 1 means the two attributes are not similar at all, and images containing one attribute would not contain the other. Example of a rating of 1: (”nature”, ”dark clouds”) - A rating of 2 means t...
[51]

a person viewing the two sets of concepts would find them to be related or a subset of them to be related
[52]

vines” Prompt Description: “technology

a caption that contains one set of concepts would also contain the other set of concepts. Here is a general guideline for each rating: - A rating of 1 means the two sets of concepts are not similar at all, and a caption containing one set of concepts would not contain the othe...
[53]

Visual Attribute: ’watercolor painting’ Semantic Attributes: [’sadness’, ’loneliness’, ’mellow’]
[54]

Menacing appear- ance

Visual Attribute: ’bright lights’ Semantic Attributes: [’wooden chest’, ’dresser’] Please adhere to the format above and provide a list of visual attributes and semantic attributes that are indicative of the visual attributes. F. Limitations and Failure Cases We outline a few ...

1990

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Introducing stable diffusion 3.5.https: //stability.ai/news/introducing- stable- diffusion-3-5, 2024

Stability AI. Introducing stable diffusion 3.5.https: //stability.ai/news/introducing- stable- diffusion-3-5, 2024. Stable Diffusion 3.5 Large, Large Turbo, and Medium models. 5, 7

2024

[2] [2]

Llama 3 model card

AI@Meta. Llama 3 model card. 2024. 2

2024

[3] [3]

Claude 3.5 sonnet, 2024

Anthropic. Claude 3.5 sonnet, 2024. 5

2024

[4] [4]

D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba. Network dissection: Quantifying interpretability of deep vi- sual representations. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6541–6549, 2017. 2

2017

[5] [5]

Easily acces- sible text-to-image generation amplifies demographic stereo- types at large scale

Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily acces- sible text-to-image generation amplifies demographic stereo- types at large scale. InProceedings of the 2023 ACM Confer- ence on Fairness, Accountability, and Transparency (FAccT),

2023

[6] [6]

Pixart-α: Fast training of dif- fusion transformer for photorealistic text-to-image synthesis,

Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of dif- fusion transformer for photorealistic text-to-image synthesis,

[7] [7]

Reproducible scal- ing laws for contrastive language-image learning

Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scal- ing laws for contrastive language-image learning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818–2829, 2023. 6

2023

[8] [8]

Tibet: Identifying and evaluating biases in text-to-image generative models, 2023

Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models, 2023. 2

2023