Backpropagating through Fr\'echet Inception Distance

TL;DR

This paper introduces FastFID, a method to efficiently incorporate FID as a loss function in training generative models, leading to improved FID scores.

Contribution

The paper presents FastFID, enabling the use of FID as a loss function for generative models, which was previously computationally challenging.

Findings

FastFID improves training efficiency for generative models.

Using FID as a loss enhances the quality of generated images.

FastFID achieves better FID scores compared to traditional methods.

Abstract

The Fr\'echet Inception Distance (FID) has been used to evaluate hundreds of generative models. We introduce FastFID, which can efficiently train generative models with FID as a loss function. Using FID as an additional loss for Generative Adversarial Networks improves their FID.