Fidelity Preference, Not Demographic Preference: A Pixel-Level Attribute-Sensitivity Audit of Image Aesthetic/Preference Scorers

偏好保真度而非人口统计学特征：图像美学/偏好评分器的像素级属性敏感度审计

Abstract: 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.

摘要： 文本生成图像系统使用学习到的美学评分器来过滤训练数据并指导生成，但这些评分是否将人口统计学属性编码为客观质量尚不明确。我们通过对合成图像和真实图像中的肤色和体型进行像素级干预，审计了四种评分器（LAION-Aesthetics、PickScore、ImageReward、HPSv2）。

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.

我们的主要发现是，在肤色明度方面，主导效应是“保真度偏好”：未修改的图像得分最高，而向任何方向的扰动都会受到惩罚（呈现倒U型曲线）。安慰剂对照组显示，这种惩罚并非肤色处理算子的人为产物，因为将相同的 CIELAB L* 偏移应用于非皮肤区域也会产生相似幅度的惩罚。然而，这种惩罚具有算子依赖性，且仅在 LAION-Aes 中对所有算子均成立。

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.

至关重要的是，仅对合成图像进行审计具有误导性：LAION-Aes 在合成人脸中表现出对深色皮肤的强烈偏好，但在 1470 张真实人脸中，这种偏好发生了逆转且变得微乎其微，放大效应也不再显著。在不同评分器之间，合成图像的结果无法迁移——对于 LAION-Aes 和 HPSv2 结果发生逆转，对于 PickScore 则出现减弱。

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.

我们贡献了一个具有伪影控制和合成/真实交叉验证的可复现基准，以及一个用于像素级因果隔离的审计标准（该标准对肤色有效，但由于形变问题对体型无效）。人口分层分析显示，在进行 FDR 校正后，除 HPSv2 外，保真度惩罚的不对称性在各群体间并不稳健。我们的研究结果表明，单纯的合成审计会误判偏见的指向和幅度，只有在真实数据上进行图像内因果隔离，才能将真正的人口统计学偏见与保真度偏好区分开来。