跳转至

文章背景与核心概要

文本到图像生成模型在很大程度上依赖于学习得来的美学与偏好评分器（如 LAION-Aesthetics、PickScore、ImageReward 和 HPSv2）来过滤训练数据集并引导生成结果。在 AI 对齐与公平性领域中，一个至关重要的问题是：这些评分器是否将人口统计学属性编码为图像质量的客观标志？

本文通过对真实和合成图像上的肤色及体型进行像素级审计，深入探讨了这一问题。研究揭示了一个关键洞察：评分器表现出强烈的是“保真度偏好”（fidelity preference），而非人口统计学偏好。具体而言：未修改的图像始终得分最高，任何对像素值的蓄意扰动（无论方向如何）都会导致评分下降。此外，纯粹在合成图像上进行的审计极具误导性，例如 LAION-Aesthetics 表面上偏爱合成人脸的较深肤色，但在真实人脸数据上这种偏好会显著逆转并减弱。该研究的方法论启示在于：要区分真正的Демографический（人口统计学）偏好与基础保真度保持，必须在真实世界数据上进行图像内的因果隔离，并对伪影和操作算子进行严格控制。

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

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

Authors: Mingyang Xu
Published: August 17, 2026
arXiv: 2608.23593 [cs.CV]

Authors: Mingyang Xu
Published: August 17, 2026
arXiv: 2608.23593 [cs.CV]

📌 Summary

📌 Summary

文本到图像模型深度依赖于学习得来的美学和偏好评分器（如 LAION-Aesthetics、PickScore、ImageReward 和 HPSv2）来过滤训练数据集并指导生成结果。在 AI 对齐和公平性领域，一个核心问题是这些评分器是否将人口统计学属性编码为了图像质量的客观标记。

Text-to-image models heavily rely on learned aesthetic and preference scorers (such as LAION-Aesthetics, PickScore, ImageReward, and HPSv2) to filter training datasets and steer generation outcomes. A critical question in AI alignment and fairness is whether these scorers encode demographic attributes as objective markers of image quality.

本文通过在合成图像与真实世界图像上对肤色及体型执行像素级审计，对这一问题进行了研究。研究揭示了一个关键见解：评分器表现出强烈的保真度偏好，而非人口统计学偏好。具体包括： - 倒 U 型惩罚： 未经改动的图像始终获得最高分；任何对像素值的微调扰动（不论其方向如何）都会导致评分下降。 - 具有误导性的合成审计： 纯粹在合成图像上进行的审计具有极大的误导性。例如，LAION-Aesthetics 表面上似乎更偏爱合成人脸上的较深肤色，但这种偏好在真实人脸图像上大体上会发生逆转并减弱。 - 方法论启示： 要将真正的人口统计学偏见与基本的保真度保持区分开来，必须基于真实世界数据开展图像内因果隔离，同时配合严谨的伪影与算子控制。

This paper investigates this by conducting pixel-level audits on skin tone and body type using both synthetic and real-world images. The study reveals a crucial insight: scorers exhibit a strong fidelity preference rather than a demographic one. Specifically: - Inverted-U Penalty: Unaltered images consistently score the highest; any artificial perturbation in pixel values (regardless of direction) leads to a score penalty. - Misleading Synthetic Audits: Audits performed purely on synthetic images are highly misleading. For example, LAION-Aesthetics appears to favor darker skin tones on synthetic faces, but this preference largely reverses and diminishes on real human faces. - Methodological Takeaway: Distinguishing genuine demographic bias from basic fidelity preservation requires within-image causal isolation on real-world data, alongside careful artifact and operator control.

🔍 Key Findings

🔍 Key Findings

  1. 保真度与人口统计学偏见的对比： 对肤色的干预表明，导致评分下降的主导驱动因素是对像素本身的操纵，从而形成了一条倒 U 型偏好曲线，其中原始图像总是更受青睐。安慰剂对照组（将偏移应用于非皮肤区域）证实，这种评分惩罚与图像的修改有关，而不仅仅是肤色操作算子本身。
  2. 合成域与真实域的差异： 合成评估结果无法泛化迁移到真实世界数据中：
  3. LAION-Aes & HPSv2： 在合成测试集与真实测试集之间，偏好发生了完全逆转。
  4. PickScore： 偏好显著减弱。
  5. 因果隔离的挑战： 尽管像素级因果隔离对于肤色调整是可行的，但由于会产生不自然的结构变形，它在体型干预方面被证明是不可靠的。
  1. Fidelity vs. Demographic Bias: Interventions on skin tone demonstrate that the dominant driver of score drops is the manipulation of pixels itself, forming an inverted-U preference curve where original images are always favored. Placebo arms (applying shifts to non-skin regions) confirm this penalty is tied to image alteration rather than the skin-tone operator alone.
  2. Synthetic vs. Real Domain Discrepancy: Synthetic evaluation results fail to transfer to real-world data:
  3. LAION-Aes & HPSv2: Preferences completely reverse between synthetic and real test sets.
  4. PickScore: Preferences significantly attenuate.
  5. Causal Isolation Challenges: While pixel-level causal isolation is viable for skin tone adjustments, it proves unreliable for body type interventions due to unnatural structural deformations.

🔗 Metadata & Resources

🔗 Metadata & Resources