FID
Better aligned with human perception compared to pixel-based metrics like SSIM and MSE
While Frechet Distance was originally a concept for measuring distance between curves, the Frechet Distance used in Frechet Inception Distance (FID) is a modified concept applied to measuring distance between probability distributions. It measures the Wasserstein-2 distance (W2) by comparing the mean and covariance of probability distributions
Since Frechet Distance inherently accounts for noise in its definition, it is widely used in vision models that employ noise-based modeling.