Source-linked AI summary
Fidelity Preference, Not Demographic Preference: A Pixel-Level Attribute-Sensitivity Audit of Image Aesthetic/Preference Scorers
Mingyang Xu
TL;DR
The paper investigates whether image-preference scorers encode demographic attributes as quality. It audits four scorers with pixel-level interventions across synthetic and real images, finding that fidelity preference dominates skin-lightness responses and that synthetic-only conclusions often fail to transfer.
Problem
Whether learned aesthetic and preference scorers encode demographic attributes as objective quality remains unclear, despite their role in filtering data and guiding generation.
Method
The paper applies within-image pixel interventions, artifact controls, and synthetic/real cross-validation to four scorers across skin tone and body type.
Results
Within-image skin-lightness responses are dominated by inverted-U fidelity preference, while synthetic audit effects reverse or attenuate on real faces across scorers.
Takeaways & Limitations
Synthetic-only audits can misjudge bias direction and magnitude, so real-data within-image causal isolation is needed to distinguish demographic bias from fidelity preference.
Takeaways & Limitations
The findings are limited to tested operators; body-type within-image intervention is invalid because geometric deformation confounds the response.
Abstract
from arXiv · showhide
Text-to-image systems use learned aesthetic scorers to filter training data and guide generation, but whether these scores encode demographic attributes as objective quality is unclear. We audit four scorers (LAION-Aesthetics, PickScore, ImageReward, HPSv2) using pixel-level interventions on skin tone and body type in synthetic and real images. Our key finding is that along skin-lightness, the dominant effect is fidelity preference: unaltered images score highest, and perturbations in either direction are penalized (inverted-U). Placebo arms show this penalty is not an artifact of the skin operator, as applying the same CIELAB L* shift to non-skin regions yields similar penalty magnitudes. However, the penalty is operator-dependent and holds for all operators only for LAION-Aes. Critically, audits on synthetic images alone are misleading: LAION-Aes shows strong preference for darker skin on synthetic faces, but on 1470 real faces the preference reverses and becomes much smaller, and amplification becomes non-significant. Across scorers, synthetic results do not transfer -- reversing for LAION-Aes and HPSv2, attenuating for PickScore. We contribute a reproducible benchmark with artifact control and synthetic/real cross-validation, and an auditability criterion for pixel-level causal isolation (valid for skin tone, not for body type due to deformation). Population-stratified analysis shows fidelity-penalty asymmetry is not robust across groups after FDR correction except for HPSv2. Our findings show naive synthetic audits misjudge bias direction and magnitude, and only within-image causal isolation on real data can distinguish true demographic bias from fidelity preference.
1 Introduction
The paper asks whether learned image-preference scorers treat demographic attributes as objective quality. It argues that within-image interventions distinguish demographic preference from fidelity preference and improve audit validity.
- Four mainstream scorers assign quality scores that filter training data, guide alignment, and shape generative-model behavior.
- Existing synthetic cross-image audits confound attribute changes with composition, lighting, and rendering artifacts.
- −26°ITA amplification on synthetic images collapsed to +0.2° amplification, not significant, on real faces under within-image intervention.
- The scorer penalized perturbations in either direction from original pixels, indicating pixel-level fidelity preference rather than demographic preference.
- The benchmark combines artifact calibration, synthetic/real cross-validation, and an auditability diagnostic that validates pixel isolation for skin tone but not body type.
2 Related Work and Positioning
The paper positions its contribution as a causal, within-image complement to corpus-scale and distributional audits. It separates pixel-level response mechanisms from categorical, caption-mediated, and provenance claims.
- Prior scorer-bias work documents social-preference concerns, genre shortcuts, anti-aesthetic penalties, and demographic effects in reward-guided generation.
- This study uniquely targets causal separation of an attribute effect from compositional confounds within a single image.
- Corpus-scale LAION-Aesthetics auditing uses between-image distributions and provenance analysis, whereas this paper uses graded within-image intervention on real faces.
- Agreement between the two approaches supports a rendering-fidelity mechanism, but neither method implies the other’s findings.
- The paper does not address caption-mediated gender or LGBTQ+ disparities, cultural and art-historical ranking, or provenance claims.
- A ±15 L* shift measures skin lightness rather than categorical racial presentation, so the results neither confirm nor refute categorical demographic findings.
- Pixel-level CIELAB intervention aims to change only the protected attribute while holding other image content fixed as much as possible.
3 The Audit Benchmark: Three-Pillar Protocol + Dual Arms + Four Scorers
The benchmark combines within-image interventions, artifact covariate calibration, best-of-n null correction, and parallel synthetic and real arms across four scorers.
- The reproducible audit_bench pipeline packages the audit as a config-driven, one-click benchmark.
- Pillar 1: within-image intervention: Within-image intervention changes one attribute while freezing remaining pixels, using CIELAB L* shifts for skin and geometric torso scaling for body type.
- Pillar 2: intervention-artifact covariate calibration: Artifact covariates model gamut clipping for skin interventions and resampling deformation for geometric warps.
- Pillar 3: best-of-n null calibration: Best-of-n net bias subtracts random-selection drift from argmax drift using paired bootstrap calibration with n = 4 candidates.
- Core analysis tool: the fidelity curve: The fidelity curve compares each dose with the original: negative differences on both sides indicate an inverted-U peak at the unaltered image.
- Dual arms: Synthetic SDXL counterfactuals and real FairFace faces serve as parallel arms for distinguishing scorer properties from rendering artifacts.
- Four scorers: The four scorers span CLIP-based LAION-Aes, PickScore, and HPSv2, plus BLIP-based ImageReward.
4 Skin-Tone Axis: Directional Reversal Between Synthetic and Real
Within-image skin-tone interventions reveal that synthetic audits can reverse or exaggerate apparent scorer preferences relative to real faces. The divergence varies by scorer, while placebo tests show that some inverted-U responses reflect fidelity sensitivity rather than skin semantics.
- Synthetic arm: On synthetic faces, LAION-Aes shows darker-skin preference with asymmetry −0.21 and best-of-n drift −26° ITA.A synthetic-only audit would suggest strong demographic preference and downstream amplification.
- Real arm: On real faces, LAION-Aes shows a near-symmetric inverted U: fidelity index +0.074, residual asymmetry +0.032, and amplification +0.2°, not significant.The real-face asymmetry is 6.5× smaller than the synthetic effect, while amplification is non-significant.
- Cross-arm transfer: Only 4 of 12 scorer-by-reading cells meet the strict criterion of FDR-significant, opposite-signed effects across synthetic and real arms.Other cells either collapse to non-significance, attenuate without reversing, or remain non-significant.
- Cross-arm transfer: HPSv2 reverses amplification from −14.3° ITA on synthetic faces to +4.3° on real faces, whereas PickScore attenuates without reversing.LAION-Aes amplification disappears on real faces rather than inverting; its confidence interval is [−2.3, +2.6].
- Placebo arms: For LAION-Aes and PickScore, shifting the same CIELAB operator to non-skin regions produces comparable fidelity penalties, with ratios 0.81 and 1.22.Both confidence intervals contain 1, supporting fidelity sensitivity beyond skin pixels or skin-lightness semantics.
- Placebo arms: The penalty is not operator-uniform: LAION-Aes penalizes blur about 6× more than skin-tone shifts, while PickScore rewards strong blur and ImageReward rewards hue rotation.Only LAION-Aes shows a significant fidelity penalty across all four operators; PickScore does so for 3/4, ImageReward for 1/4, and HPSv2 for 0/4.
5 Scale Calibration: Why Synthetic Audits Mislead
Synthetic audits can misread scorer behavior because cross-image differences combine attribute effects with rendering and composition confounds. Within-image causal isolation on real data is therefore needed to distinguish directional demographic effects from fidelity preference and to avoid overstating magnitude.
- Why synthetic audits mislead: Cross-image observed score differences combine pure attribute effects with composition and rendering confounds.The decomposition is stated as: cross-image observed score diff = pure attribute effect (within-image) + composition/rendering confound.
- Why synthetic audits mislead: Synthetic rendering differences can obscure the real inverted-U effect and make accompanying differences appear to be demographic-direction bias.Best-of-n selection can further amplify the misread effect.
- Why synthetic audits mislead: Within-image causal isolation on real data is required to assess the direction, magnitude, and mechanism of scorer bias.This approach distinguishes a real fidelity preference from a one-directional preference along the continuous lightness axis.
- Scope: The findings are limited to the tested intervention protocols and do not establish that models learn only local statistics.The inverted-U response is the dominant measured effect on these operators, but other global features may coexist and the effect is not operator-uniform in magnitude.
6 Body Axis: A Cross-Attribute Stress Test of the Fidelity Mechanism (Not a Body-Bias Measurement)
The body axis tests whether fidelity preference generalizes across perturbation types, but geometric deformation prevents interpreting it as body demographic bias. Results show score penalties track deformation magnitude rather than body width, motivating an auditability criterion for deciding when pixel-level interventions support bias measurement.
- Scope and rationale: Body-type perturbations are framed as a cross-attribute stress test, not a measurement of body demographic bias, because geometric warping introduces unavoidable deformation artifacts.The body axis tests whether the original state scores highest and perturbations are penalized under a non-lightness intervention.
- Scope and rationale: 378 warped images span 54 base images across seven torso-warp levels, with silhouette width_ratio measured from the synthetic body set.The design uses 6 body-type gradient levels generated from 3 scenes and 3 seeds, then applies warps from γ=0.80–1.20.
- Cross-attribute results: LAION-Aes, PickScore, and ImageReward show symmetric inverted-U fidelity curves, while HPSv2 is practically flat because its fidelity index is 21–57× smaller.All four fidelity indices survive FDR correction, but only HPSv2 has a significant asymmetry, with negligible magnitude.
- Artifact control: The deformation artifact significantly depresses scores for all four scorers, and controlling for it leaves the pure width effect mostly non-significant except for PickScore.The regression separates score effects attributable to width from those attributable to warp_artifact.
- Artifact control: Warp magnitude forms a near-perfect V around γ=1.0, explaining the mirrored inverted-U score pattern and the near-zero correlation with true body width.Median deformation at γ=0.80/1.00/1.20 is 0.149 / 0 / 0.099; within-base width correlation is effectively zero.
- Auditability criterion: A bias measurement requires all three auditability conditions; when any condition is missing, the intervention supports only a mechanistic stress test.The skin axis satisfies the criterion on 1470 real faces, whereas the body axis lacks all three conditions.
- Auditability criterion: Population-stratified amplification shows no systematic drift for the body-axis inverted U, consistent with the absence of directional tendency.Random null is approximately zero, and all four scorers’ net-bias confidence intervals straddle zero.
7 Rigor Checks: Multiple Comparisons, Measurement Validity, Robustness, Falsifiability
The benchmark treats false-positive control, measurement validity, robustness, and falsifiability as fixed audit requirements. Across these checks, the principal skin-axis conclusions survive correction, while calibration exposes the body-axis measurement defect and population differences mostly fail selection-free FDR testing.
- Multiple Comparisons: 22 of 36 headline decisions remain significant after BH-FDR correction, and the main conclusions survive correction.The correction downgrades only two raw-significant results, including the small real-face LAION-Aes discriminative effect.
- Measurement Validity: ITA° is validated on real and synthetic skin arms, with R^2 = 0.79 and 0.88, while width_ratio fails calibration at R^2 = 0.02.The skin measurement also achieves monotonic rates of 97.6% for real faces and 98.1% for synthetic faces.
- Falsifiability: The open result space supports falsifiability because synthetic and real arms conflict, calibration passes skin but vetoes body, and FDR downgrades a headline candidate.These outcomes constrain the conclusions rather than selectively confirming a preset target.
- Robustness: 100% same-sign rates across five seeds support robustness for every real-face skin-axis reading.The large synthetic LAION-Aes drift contrasts with weak real-arm drift, reinforcing the interpretation that the synthetic result is artifact-sensitive.
- Population Stratification: Only HPSv2 shows robust population heterogeneity after selection-free omnibus testing and FDR correction.LAION-Aes’s largest contrast has raw p = 0.014 but omnibus p = 0.422, illustrating the danger of cherry-picking extremes.
8 Deliverable: The Reproducible Audit Benchmark audit_bench
The deliverable is a reusable audit benchmark that combines within-image interventions, artifact calibration, and fidelity-curve analysis to separate content confounds from demographic effects.
- audit_bench: The benchmark combines pixel-level interventions, artifact covariates, fidelity curves, and best-of-n null calibration for scorer audits.Its design targets the distinction among content confounds, intervention artifacts, fidelity preference, and true demographic bias.
9 Limitations and Future Work
The paper’s main boundaries are that the skin intervention measures lightness rather than ethnicity, the body axis is confounded by deformation, and real-data coverage is limited to skin tone. Future work therefore requires a clean real-body arm and broader attribute crossings.
- Synthetic/Real Validity: Synthetic-only audits can reverse or magnify conclusions: LAION-Aes shifts from −26.1° ITA amplification synthetically to +0.2° on real faces, not significant.The figure describes this as an approximately 129× magnitude collapse rather than a sign flip.
- Fidelity Preference: The inverted-U fidelity preference is not itself demographic bias; demographic harm requires group-differentiated penalties on original states.Population-level analysis finds this harm is not robust for most scorers, although detection power is limited by 30 bases per population.
- Body Axis: The body-axis inverted-U is driven by deformation artifact rather than width semantics.The deformation coefficient is negative and dominates, while the pure width effect is weak after warp control.
- Body Axis: The body axis cannot support quantitative causal readings because geometric warps necessarily introduce deformation, with within-group score correlation −0.40 for deformation and +0.004 for width.Adding real-model images would not remove the artifact if within-image warping remains the intervention.
- Future Work: A clean real-body measurement arm remains the key follow-up, using real models with fixed background, pose, clothing, and camera.The proposed route would make body type an innate attribute rather than a warped image property.
- Skin Axis: The L* intervention isolates pixel sensitivity to skin lightness, not the full construct of ethnicity.This deliberate narrowing exchanges scope for clean causal identification.
- Coverage: The real arm covers skin tone in 1470 FairFace images, whereas the body axis currently has only a synthetic arm.The paper treats skin conclusions as anchored on the larger real-face sample and the synthetic arm as methodological evidence.
10 Ethics Statement and Broader Impact
The paper frames its contribution as a reproducible validity-check tool for scorer audits while limiting demographic claims and addressing dual-use risks. It uses public data and avoids new human-subject collection, with ethics procedures reserved for future real-body work.
- Broader Impact: The benchmark turns concerns about scorer-encoded demographic quality into a reproducible quantitative diagnosis for dataset builders and model developers.Its stated benefit is correcting the misconception that naive synthetic audits reliably identify bias direction and magnitude.
- Ethics: The study uses public datasets, including 1470 FairFace images and 311 SSP-3D frames from 62 individuals, without collecting new human-subject data.FairFace labels are restricted to population-stratified validity analysis, and no attribute-prediction model is trained.
- Demographic Labels: ITA° measures continuous skin lightness, while race labels are used only for stratification rather than population classification.The paper explicitly rejects treating FairFace categories as essentialist population divisions.
- Dual Use: The disclosed fidelity-preference mechanism is dual-use, so the paper emphasizes defensive validity checking over offensive bias amplification.The authors state that open-sourcing the protocol makes defensive use easier than offensive use.
- Future Research Ethics: Future real-body collection requires informed consent, compensation, de-identification, and institutional ethics review.Those procedures apply to the proposed human-subject Route A arm, not the present public-data study.
Generative AI Usage Statement
The authors used generative AI tools for drafting, translation, and analysis-code implementation, while retaining responsibility for experimental design, statistics, and substantive claims.
- Generative AI supported prose drafting and copy-editing, Chinese–English translation, and analysis-code implementation and refactoring.
- The authors made all experimental design decisions, statistical procedures, and substantive claims.
- Every reported number was machine-verified against released structured results by verify_ci.py.