What FID Hides: Detecting, Ranking, and Diagnosing Deviations in Generative Evaluation

论文概要

研究领域: ML
作者: Hao Chen
发布时间: 2026-08-25
arXiv: 2608.24881

中文摘要

生成模型通常通过Frechet Inception Distance (FID)和Kernel Inception Distance (KID)进行排序，但FID的前两矩摘要可能遗漏分布差异，且报告的标量差距本身并不是针对采样变化的校准检验。FID的矩限制有具体后果：在ImageNet上，仅优化以匹配参考Inception均值和协方差的视觉不可识别图像获得FID 24.7，而留出真实图像为58.6（越低越好）。此外，FID和KID是标量差异，在两个样本交换时保持不变，因此不编码离散度变化的方向：模式崩溃中的欠离散与过离散。我们引入ZID（Z-resolved Integrated Diagnostic），它结合了来自秩图（RISE）和高斯核（两种带宽的GPK）的六个标准化位置敏感和离散度敏感分支。ZID报告三个关联输出：用于排序偏离幅度的指数、用于检验分布相等性的置换p值、以及用于诊断的有符号离散度读数。在控制实验中，ZID检测到广泛的偏离，其分数跟踪相应扫描中的严重程度增加，包括FID持平或反转的情况。

原文摘要

Generative models are commonly ranked by Fréchet Inception Distance (FID) and Kernel Inception Distance (KID), yet FID's first-two-moment summary can miss distributional differences, and a reported scalar gap alone is not a calibrated test against sampling variation. FID's moment restriction has concrete consequences: on ImageNet, visually unrecognizable images optimized only to match the reference Inception mean and covariance obtain FID $24.7$ versus $58.6$ for held-out real images (lower is better). Moreover, FID and KID are scalar discrepancies that are unchanged when the two samples are exchanged and therefore do not encode the direction of a dispersion change: under-dispersion, as can occur in mode collapse, versus over-dispersion. We introduce \textbf{ZID} (\emph{Z-resolved Integrat...

自动采集于 2026-08-27

#论文 #arXiv #ML #小凯