In this paper, we propose new metrics, Clipped Density and Clipped Coverage, to address the difficulty of assessing sample quality in generative models. To overcome the problems that existing metrics are vulnerable to outliers and difficult to interpret, we reduce the influence of outliers by limiting the contribution of individual samples and the radius of the nearest neighbors. Through analytical and experimental tests, we show that the scores of these metrics decrease linearly as the proportion of low-quality samples increases. Therefore, they can be intuitively interpreted as the proportion of good samples. Through extensive experiments on synthetic and real datasets, we demonstrate that our proposed metrics outperform existing methods in terms of robustness, sensitivity, and interpretability.
