FIDGAN: A Generative Adversarial Network with An Inception Distance

Abstract

Two evaluation metrics for GAN models have been proposed in existing studies: Inception score (IS) and Fréchet Inception distance (FID). We propose a new GAN model based on the idea that backpropagating the FID score would guide the GAN to efficiently learn the distribution of real images and generate high-quality images. Based on such an idea, we propose a training loss for the generator to minimize a modified FID loss. Trained with the CIFAR-10 dataset, FIDGAN exhibited an FID of 11.78, which corresponds to a reduced FID compared to an existing model called BigGAN by 20.0%.