FairJudge: Abstention-Aware Multimodal Judges for Fairness and Alignment Evaluation in Text-to-Image Models

Ioannis Patras; Maimuna Nowaz; Maryam Fetanat; Matthew Purver; Zahraa Al Sahili

REVIEW 4 major objections 4 minor 16 references

FairJudge: Abstention-Aware Multimodal Judges for Fairness and Alignment Evaluation in Text-to-Image Models

T0 review · 4 major / 4 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 A constrained, evidence-seeking judge protocol outperforms CLIP on text-to-image fairness attributes, the paper claims.

desk verdict A sensible protocol for abstention-aware MLLM judging, but the headline gains on religion/disability and alignment rest on an unverified prompt-derived yardstick and a lenient abstention-as-correct rule. read the letter →

arxiv 2510.22827 v3 pith:LLFUZ3O5 submitted 2025-10-26 cs.CV cs.LG

Zahraa Al Sahili , Maimuna Nowaz , Maryam Fetanat , Ioannis Patras , Matthew Purver This is my paper

classification cs.CVcs.LG

keywords text-to-imageevaluationmultimodalLLMjudgefairnessauditingsocialattributepredictionabstentionprompt–imagealignmentDIVERSIFYlabel-constrainedprompting

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 a text-to-image (T2I) evaluator built from instruction-following multimodal large language models (MLLMs) can be made both more accurate and more auditable than standard proxies by constraining its outputs to closed label sets, requiring it to cite visible evidence, and letting it abstain with an explicit 'unspecified' answer. Across four attribute-prediction benchmarks and three profession/alignment benchmarks, these constrained judges outperform CLIP and DeepFace on gender and age, close large gaps on religion and disability, and raise mean prompt–image alignment while keeping profession accuracy high. The authors argue the gains come from converting open-ended judging into a parseable, evidence-grounded procedure, and they show in ablations that dropping label constraints or disabling abstention sharply degrades race, age, and religion accuracy. The paper also contributes DIVERSIFY and DIVERSIFY-Professions, synthetic corpora designed to probe culturally situated and non-facial attributes that face-only benchmarks miss. A sympathetic reader would care because current evaluation proxies either force predictions or reward surface correlation, masking failure modes that careful human review would catch.

What carries the argument

The central mechanism is the FairJudge protocol: (i) closed label sets for each target attribute with an explicit 'unspecified'/'Neutral' fallback; (ii) a strict JSON output schema that makes responses parseable and auditable; (iii) a required per-field evidence rationale logged for audit but not used in scoring; and (iv) for alignment, a 1–5 rubric mapped by s=(r−3)/2 to [−1,1], making judge scores comparable to CLIP cosine. The same protocol is applied uniformly across models (GPT-4.1, Gemini 1.5 Pro, LLaMA-4) and across datasets. Its work is to convert an open-ended MLLM judgment into a calibrated decision that can abstain, and to force the judge to separate visible evidence from inferenc

What would settle it

Take a random sample of DIVERSIFY images (e.g., 100) and have human annotators label the six attributes from pixels alone, without seeing the generating prompt; then score judges against these pixel-grounded labels. If religion and disability accuracy, or the alleged gains over CLIP, collapse relative to the prompt-derived labels — or if judges with the closed label set still predict labels whose cues are absent from the pixels (e.g., predicting 'Sikh' when no turban is rendered) — the protocol is measuring prompt intent, not visual evidence. A second check: generate prompt variants that inten

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that social-attribute prediction in T2I should be framed as a constrained judging problem rather than a similarity-scoring problem. The FairJudge protocol gives an MLLM a fixed taxonomy per attribute (gender, race, age, religion, culture, disability), demands strict JSON output with a rationale grounded in visible cues, and permits 'unspecified' when evidence is insufficient; for alignment, it replaces cosine similarity with a 1–5 rubric linearly rescaled to [−1,1]. Evaluated on FairFace, PaTA, FairCoT, and the new DIVERSIFY, judge models deliver double-digit improvements over CLIP on gender and age, and on DIVERSIFY raise religion accuracy to 0.689–0.817 v

Load-bearing premise

The load-bearing premise is that DIVERSIFY's ground-truth labels, taken verbatim from the prompts that generated the images, match what is actually rendered in the pixels; the paper concedes in §A.2 and §A.6 that no human verification was done and that generators may fail to render requested cues, so judge accuracy against these labels could partly reflect prompt-taxonomy guessing rather than visual grounding.

Editorial extensions

If this is right

If correct, fairness auditing of T2I models can be done with a lightweight, model-agnostic protocol that needs no training of attribute recognizers and no human labels beyond the prompt taxonomy.
Evaluators gain coverage of attributes that are weakly encoded in pixels — religion, culture, disability — where face-centric and contrastive baselines were previously the only automated options.
The protocol's scoring scale is designed for direct comparison with CLIP, so existing CLIP-based pipelines can be upgraded to judge-based verification without changing reporting conventions.
Ablations imply a practical cascade: use contrastive retrieval for high-recall candidate generation, then a judge for verification and abstention, escalating uncertain cases to a stronger or second judge.
The released per-image outputs, rationales, and parser logs permit re-scoring under alternative rubrics and taxonomies, turning evaluation into a reproducible audit artifact.

Reading between the lines

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

DIVERSIFY's labels are prompt-derived (the paper says so in §A.2 and §A.6), so the reported religion/disability accuracies may partly measure how well the judge recovers the prompt taxonomy's prior rather than what is visible in pixels; a human-annotated pixel-grounding subset would separate these.
Because culture is diffuse and frequently non-visual, the protocol's abstention behavior is itself a finding about the limits of image-only evaluation; extending the judge to short video clips or multiple views could test whether contextual cues resolve current abstentions.
The template-perturbation check is narrow; a systematic sweep over instruction wording, label order, and label-set granularity would quantify how much of the performance gap is protocol design versus the base model's priors.
The cost table makes a progressive-judging cascade an obvious engineering choice; a direct experiment measuring accuracy and alignment under an 'escalate abstentions only' policy would test whether most of the gain survives at lower cost.

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

FairJudge is a well-designed protocol for making MLLM judges abstain when evidence is weak, and the gender/age results on human-annotated benchmarks look real. But the paper’s broadest claims — double-digit gains on religion and disability, and higher alignment — are not yet supported, because the DIVERSIFY yardstick is prompt-derived, the disability scoring counts “unspecified” as correct on an 87%-unspecified set, and the abstract promises VIEScore/VQAScore baselines that never appear.

What’s actually new: the protocol itself — closed labels, explicit abstention, evidence rationales, rubric mapping — is a useful, lightweight package for T2I fairness auditing. The authors are honest about limitations; §A.6 and §7 admit the label–image mismatch and the possibility that high scores reflect dataset regularities. The FairFace/PaTA results with human labels show genuine gains over CLIP and DeepFace for gender and age. The two synthetic corpora, if labels were human-verified, would fill a real gap.

The load-bearing problem is the DIVERSIFY benchmark. Labels are text-derived from prompts, no pixel verification. A judge matching the prompt-derived label may be recovering prompt intent, not visual content. The disability scoring makes it worse: Table 11 shows 87.2% of images are “unspecified” for disability, and §4.5 says “unspecified” is scored as correct when ground truth is “no disability.” So a judge that mostly abstains gets ~87% accuracy before looking at the image. GPT-4.1’s 93.2% is not yet meaningful. CLIP is forced to choose a visible label, so the comparison is not on equal footing. This doesn’t invalidate the face-centric benchmarks, but it guts the religion/disability generalization claims and the DIVERSIFY alignment numbers. Also, the alignment comparison maps a 1–5 rubric to [-1,1] with no human calibration; “raises mean alignment” hasn’t been shown to mean better alignment. And the abstract’s mention of VIEScore/VQAScore as baselines is simply false — they don’t appear in Section 4. No error bars anywhere, and artifacts are promised only upon acceptance.

So who is this for? People building evaluation pipelines for T2I fairness will find the protocol worth trying, and the paper is a useful cautionary example of benchmark validity. It deserves a serious referee, but it needs major revision: drop or fix the missing baselines, human-verify a sample of DIVERSIFY labels, redo disability scoring, add uncertainty, and calibrate alignment against human judgment.

I’d send it to review — the core idea is sound and the authors are transparent — but I’d expect the reviewers to push for the label-verification study and the scoring fix before acceptance.

Referee Report

4 major / 4 minor

Summary. The paper proposes FairJudge, an evaluation protocol for text-to-image (T2I) models in which an MLLM acts as a structured judge. The judge is constrained to closed label sets for gender, race, age, religion, culture, and disability; must provide evidence-based rationales; may abstain via an explicit 'unspecified' option; and rates prompt–image alignment on a 1–5 rubric that is linearly rescaled to [−1,1]. The authors evaluate attribute prediction on FairFace, PaTA, FairCoT, and a new DIVERSIFY corpus, and profession/alignment on IdenProf, FairCoT-Professions, and DIVERSIFY-Professions. They compare against CLIP and DeepFace and report large gains, especially on religion and disability, while also introducing two synthetic corpora whose labels are derived from the generating prompts.

Significance. The protocol is simple, lightweight, and potentially useful for auditable T2I evaluation. The gender and age gains over CLIP and DeepFace on human-annotated benchmarks such as FairFace and PaTA are large and credible, and the release of prompts, per-image outputs, and parser logs would support reproducibility. If the DIVERSIFY-based claims were validated, the new corpora would be a valuable stress test for non-facial attributes. However, the current evidence for the headline religion/disability gains and for the alignment superiority is weakened by label-provenance and scoring-scale issues, and the abstract promises baselines that do not appear in the experiments.

major comments (4)

[Abstract; §4.1; Tables 1–4] The abstract states that FairJudge 'outperforms or complements CLIP, DeepFace, VIEScore, and VQAScore,' but VIEScore and VQAScore never appear in §4.1 or in Tables 1–4. The experimental comparison is only against CLIP and DeepFace. The promised baselines are load-bearing for the headline claim of superiority; either add these baselines or revise the abstract and all related claims.
[§4.5, Table 11, §A.2, §A.6] The DIVERSIFY labels are prompt-derived, not human-verified (§A.2: 'All labels are text-derived from the prompt templates; no human annotation of rendered pixels is required'), and §A.6 concedes possible label–image mismatch. The disability scoring policy in §4.5 counts 'unspecified' as correct whenever the ground-truth label is 'no disability'; Table 11 shows 87.2% of DIVERSIFY images have 'unspecified' disability. GPT-4.1's 93.2% disability accuracy in Table 2 is therefore close to a majority-class baseline of 87.2%, and the comparison against a forced-choice CLIP is confounded. The religion gains (Table 2: CLIP 0.345 vs GPT-4.1 0.689) may likewise reflect recovering prompt intent rather than visual grounding if cues were not rendered. Please provide human verification of at least a random subset of labels, report accuracy on non-unspecified images, and give CLIP an equivalent abstenti
[§3.2.2, Table 3] The alignment claim rests on comparing CLIP cosine similarity with the rubric-based score s=(r−3)/2. Both lie in [−1,1], but the distributions are not calibrated: CLIP cosine values for text–image pairs are typically concentrated near 0.2–0.3, while rubric ratings map to 0.5–0.8. Table 3 reports CLIP alignment 0.250 vs GPT-4.1 0.810 on DIVERSIFY-Professions; this 'substantial increase' may be a scale artifact. Please show rank correlations or distribution-matched comparisons, or treat the two families of scores as not directly comparable.
[Table 4, §4.5] The ablation table is incomplete for the central claim about abstention: the GPT-4.1 'without unspecified' row is entirely dashes, yet the text states that disabling abstention 'disproportionately harms attributes with weak or intermittent evidence.' No quantitative support is shown for GPT-4.1. Please fill this row or restrict the claim to models with data. Also, Table 4 uses 'Gemini 2.5 Pro' for ablations while §4.1 lists Gemini 1.5 Pro as a judge; clarify which model was used.

minor comments (4)

[Table 4] The GPT-4.1 'Without labels' Age entry is reported as 0.86%, which looks like a formatting error (0.86 rather than 86%). Please check and correct.
[Listing 1; §3.2.1] Culture is described as a free-text tag in the prompt ('Culture: free-text tag OR "unspecified"'), while §3.2.1 says the protocol uses a closed label set. Clarify whether culture is closed or open, since this affects the 'label-constrained' claim.
[§A.2, Tables 6–11] Several categories have zero counts (Child, Indigenous, etc.). The text says 'Demographic coverage is balanced by design,' but the tables show empty strata; please state these zeros explicitly and explain their implications.
[§4.5] Rationales are logged but not used for scoring. A small human audit of rationales would strengthen the evidence-grounding claim and help rule out spurious correlations.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline disability result on DIVERSIFY is partly built into the scoring rule: 'unspecified' is defined as 'no disability' and 87.2% of the benchmark's disability labels are 'unspecified', so a judge that abstains earns the majority-class base rate by construction; the other evaluation claims have independent content.

self definitional [Sec. 4.5 (Ablation; scoring policy) and Table 11 (DIVERSIFY disability distribution)]
"for disability, unspecified is interpreted as predicting no disability (i.e., scored as correct when the ground truth is “no disability” and as an error otherwise). ... Table 11: Unspecified 409 87.2%."

The scoring policy makes the judge's abstention output 'unspecified' definitionally equal to the correct answer whenever the ground-truth disability label is 'unspecified'. Table 11 shows that this is true for 409/469 (87.2%) of DIVERSIFY images. Thus any judge that follows the protocol's instruction to output 'unspecified' when evidence is insufficient receives ~87.2% accuracy before looking at pixels. The reported GPT-4.1 disability accuracy of 0.932 (Table 2) is therefore close to the base rate of the 'unspecified' class, while CLIP—which cannot abstain and must choose a visible disability label—is scored wrong on the same 87.2% majority by construction. The claimed 'large margins' on disability are an artifact of the scoring rule, not of visual grounding.

full rationale

The core FairJudge protocol (label-constrained JSON, evidence rationales, abstention, linear mapping of the 1–5 rubric to [-1,1]) is not circular: it is a definition of a scoring procedure, and the comparisons on gender, age, race, religion and profession/alignment rest on standard labeled benchmarks or on the rubric's own output rather than on fitting the evaluation target. Self-citations (FairCoT, 'Data matters most') are background/mitigation work and are not load-bearing for the judge-protocol derivation. The main circular step is localized to the DIVERSIFY disability results: under Sec. 4.5's scoring policy, 'unspecified' is interpreted as 'no disability' and Table 11 shows 87.2% of DIVERSIFY ground-truth disability labels are 'unspecified', so the judge's abstention is scored correct by definition on the majority of the benchmark. This makes the disability accuracy numbers and the comparison to a non-abstaining CLIP baseline partly tautological. The paper's separate limitation that DIVERSIFY labels are prompt-derived without human pixel annotation (Sec. A.2, A.6) is a data-validity concern rather than a circular derivation; it reduces confidence in the religion/culture results but does not by itself make the evaluation reduce to its inputs. Other benchmark claims retain independent content, so the overall circularity is partial rather than total.

Assumptions & free parameters 2 free parameters · 4 assumptions · 2 invented entities

The protocol itself adds no free physical parameters; the hand-chosen knobs are the abstention scoring rules and the rubric/anchor design. The dominant axioms are dataset-label provenance (prompt-derived labels treated as ground truth) and uncalibrated alignment comparability. No new physical entities; the DIVERSIFY corpora are new synthetic assets whose labels are not independently verified.

free parameters (2)

Attribute-specific abstention scoring policy = disability: unspecified counted as no-disability; religion/culture: unspecified excluded; gender/race/age: unspecified =
Introduced in §4.5. On DIVERSIFY, 87.2% of disability labels are 'unspecified' (Table 11), so a judge that always abstains scores ~87% disability accuracy. This hand-chosen rule directly shapes the headline disability gains.
Alignment rubric anchors + linear map s=(r−3)/2 = s = (r−3)/2; anchors 1..5 defined in §3.2.2
The 1–5 rubric definitions and the rescaling are designer choices. The claim that the rescaled scores are 'directly comparable' to CLIP cosine is asserted without calibration, yet Table 3's margins (0.81 vs 0.25) rest on this assumption.

assumptions (4)

ad hoc to paper DIVERSIFY's prompt-derived labels equal the attributes actually present in the rendered pixels
§A.2 states labels are text-derived with no pixel annotation; §A.6 admits 'generators may fail to render requested cues, leading to label–image mismatch'. Accuracy on DIVERSIFY therefore measures agreement with prompt intent, not verified visual content.
domain assumption The MLLM 1–5 rubric score tracks human alignment judgments without calibration
§4.4 compares mean rubric-rescaled scores with CLIP cosine and claims judges 'surpass' CLIP; §7 defers human triangulation to future work. Without human ratings, a higher mean rubric score is not shown to mean better alignment.
domain assumption Judge outputs are stable across API versions and sampling temperature
No temperature or API-version information is reported (§4.1, §C), so the leaderboard swaps across datasets in Table 1 could be noise rather than 'complementary strengths'.
ad hoc to paper CLIP cosine and rubric-rescaled scores on [−1,1] are comparable as alignment measures
§3.2.2 says the rescaling makes MLLM scores 'directly comparable to cosine-based CLIP alignment'. The two instruments have different calibration and variance, so absolute differences in Table 3 are not evidence of better alignment.

invented entities (2)

DIVERSIFY corpus (469 synthetic images + prompt-derived labels)
purpose: Benchmark for religion/culture/disability and demographic attribute judging beyond face-visible cues (§3.3)
Images are not released; only prompts/metadata and regeneration scripts are promised (§A.7), and the generator is unnamed. Labels have no independent verification against pixels, so the falsifiable handle outside the paper is weak.
DIVERSIFY-Professions corpus (1,200 synthetic images)
purpose: Profession grounding and alignment benchmark with non-iconic scenes (§3.3)
Same prompt-derived label provenance as DIVERSIFY (§B.2). Includes labels outside the judge's closed set (e.g., Taoist religion, Table 14), making some instances unanswerable by construction and deflating coverage/accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FairJudge: Abstention-Aware Multimodal Judges for Fairness and Alignment Evaluation in Text-to-Image Models." pith.science (2026). https://pith.science/paper/LLFUZ3O5

@misc{pith2026251022827,
  author       = {Pith},
  title        = {Pith review of: FairJudge: Abstention-Aware Multimodal Judges for Fairness and Alignment Evaluation in Text-to-Image Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLFUZ3O5}},
  note         = {Machine review of arXiv:2510.22827}
}

abstract

Evaluating text-to-image (T2I) systems requires judging not only whether an image matches a prompt, but also whether socially salient attributes are represented faithfully and without unsupported inference. Existing automated evaluators typically rely on face-centric recognizers or contrastive image--text similarity, which provide limited diagnostic feedback and often force predictions even when visual evidence is ambiguous or absent. For fairness-sensitive attributes such as religion and disability, where cues may be contextual, indirect, or intentionally unspecified, these evaluators can therefore miss failure modes that careful human reviewers would notice. We introduce \textsc{FairJudge}, an abstention-aware evaluation protocol that uses instruction-following multimodal LLMs as structured judges for social-attribute prediction, profession grounding, and prompt--image alignment. The protocol constrains outputs to closed label sets, requires visible-evidence rationales, supports an explicit \textsc{unspecified} decision when cues are insufficient, and maps rubric-based alignment judgments to $[-1,1]$. These constraints turn MLLM judging from open-ended assessment into a parseable, auditable evaluation procedure. Across four attribute-prediction benchmarks and three profession/alignment benchmarks, \textsc{FairJudge} outperforms or complements CLIP, DeepFace, VIEScore, and VQAScore. Ablations show that closed labels, abstention, and evidence reporting are central to reliability. We further introduce \textsc{DIVERSIFY} and \textsc{DIVERSIFY-Professions}, two context-rich resources for evaluating social representation and profession grounding beyond face-visible or iconic cues. We release code, prompts, datasets, parser logs, and per-image judge outputs to support reproducible auditing.

Figures

Figures reproduced from arXiv: 2510.22827 by the authors.

**Figure 2.** FairJudge: instruction-following MLLMs as fair judges. (a) Social-attribute prediction with label [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Overview of the DIVERSIFY benchmark. A mosaic of example images illustrating diverse professions, cultures, scenes, and non-iconic viewpoints. The dataset is designed to reduce shortcut cues and stress-test judges on religion, culture, and disability situated, context-dependent signals [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 4 linked inside Pith

[1]

LLM-as-a-judge

Introduction Text-to-image (T2I) systems are increasingly used to generate and retrieve people-centric imagery in creative, assistive, and analytical settings. Yet these models inherit—and can amplify—social biases from their data and training objectives, yielding skewed portrayals across gender, skin tone, geo-culture, religion, disability, andprofession...

arXiv 2023
[2]

LLM-as-a-judge

Related Work 2.1. Bias in Text-to-Image Models Text-to-image (T2I) systems inherit and sometimes amplify social biases present in their training data. Early large-scale analyses show systematic disparities across gender, skin tone, geo-culture, and professions, and propose diagnostic protocols and scorecards for auditing representation (Luccioni et al., 2...

2023
[3]

Methods 3.1. Background: CLIP and CLIP-based Alignment CLIP.Contrastive Language–Image Pretraining (CLIP) learns an image encoder fθ(·)and a text encoder gϕ(·)such that paired image–text embeddings have high cosine similarity while mismatched pairs are pushed apart. Given an image I and a textual prompt t, the canonical alignment score is the cosine CLIPA...
[4]

Gender (male/female/unspecified)
[5]

Race (Black, White, Asian, Latino_Hispanic, Indigenous, unspecified)
[6]

Age (child, young adult, middle-aged, elderly, unspecified)
[7]

Religion (Christian, Muslim, Hindu, Buddhist, Jewish, Sikh, Shinto, Neutral, unspecified)
[8]

Culture (e.g., British, African, East Asian, Latin American, or unspecified)

Show all 16 references

[9]

Gender",

Disability (mobility impairment, blind/low vision, deaf/hard of hearing, dwarfism, vitiligo, unspecified) Output strict JSON with keys: "Gender", "Race", "Age", "Religion", "Culture", "Disability". If evidence is insufficient, use "unspecified". Calibration and neutrality.We a...
[10]

a photo of

Experimental Results We evaluate three questions: (i)social-attribute predictionfor gender, race, and age on FairFace,PaTA,FairCoT, andDIVERSIFY; (ii)generalizationto religion, culture, and disability; and (iii)prompt–image alignment onIdenProf,FairCoT-Professions, and DIVERSI...
[11]

task intent with evidence,

Discussion Our results show thatjudge-stylemultimodal LLMs provide a more faithful and conservative read of sensitive attributes than contrastive or face-centric baselines. The strongest gains appear where visualevidenceisdiffuseorcontext-dependent. On gender and age, judges r...
[12]

Conclusion Weintroducedalightweight,calibratedprotocolthat treats instruction-following MLLMs asfair judges for text–to–image (T2I) evaluation. By constraining outputs to label sets, requiring explicit visual evidence, and encouraging principled abstention, the protocol shifts...
[13]

First, ambiguity in annotation undermines ceiling performance, particularly for culture where definitions are fluid and culturally situated

Limitations Despite the gains, several limitations deserve emphasis. First, ambiguity in annotation undermines ceiling performance, particularly for culture where definitions are fluid and culturally situated. Even when judges abstain appropriately, measured accuracy will refl...
[14]

Bibliographical References Zahraa Al Sahili, Ioannis Patras, and Matthew Purver. 2025. Data matters most: Auditing social bias in contrastive vision–language models. Transactions on Machine Learning Research. Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, and Benyou ...

2025 arXiv
[16]

Base prompt (doctor)

Dear: Debiasing vision-language models with additive residuals. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820–6829. Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. 2023. Finetuning text-to-imag...

2023 arXiv
[2023]

JackHessel,AriHoltzman,MaxwellForbes,Ronan Le Bras, and Yejin Choi

Fair diffusion: Instructing text-to-image generation models on fairness.arXiv preprint arXiv:2302.10893. JackHessel,AriHoltzman,MaxwellForbes,Ronan Le Bras, and Yejin Choi. 2021. Clipscore: A reference-free evaluation metric for image captioning. InProceedings of the 2021 conf...

2021 arXiv

Pith tools

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

[1] [1]

LLM-as-a-judge

Introduction Text-to-image (T2I) systems are increasingly used to generate and retrieve people-centric imagery in creative, assistive, and analytical settings. Yet these models inherit—and can amplify—social biases from their data and training objectives, yielding skewed portrayals across gender, skin tone, geo-culture, religion, disability, andprofession...

arXiv 2023

[2] [2]

LLM-as-a-judge

Related Work 2.1. Bias in Text-to-Image Models Text-to-image (T2I) systems inherit and sometimes amplify social biases present in their training data. Early large-scale analyses show systematic disparities across gender, skin tone, geo-culture, and professions, and propose diagnostic protocols and scorecards for auditing representation (Luccioni et al., 2...

2023

[3] [3]

Methods 3.1. Background: CLIP and CLIP-based Alignment CLIP.Contrastive Language–Image Pretraining (CLIP) learns an image encoder fθ(·)and a text encoder gϕ(·)such that paired image–text embeddings have high cosine similarity while mismatched pairs are pushed apart. Given an image I and a textual prompt t, the canonical alignment score is the cosine CLIPA...

[4] [4]

Gender (male/female/unspecified)

[5] [5]

Race (Black, White, Asian, Latino_Hispanic, Indigenous, unspecified)

[6] [6]

Age (child, young adult, middle-aged, elderly, unspecified)

[7] [7]

Religion (Christian, Muslim, Hindu, Buddhist, Jewish, Sikh, Shinto, Neutral, unspecified)

[8] [8]

Culture (e.g., British, African, East Asian, Latin American, or unspecified)