Conditional Image Generation with One-Vs-All Classifier

TL;DR

This paper introduces GAN-OVA, a novel conditional image generation model using a One-Vs-All discriminator to improve training stability, speed, and quality across datasets like MNIST and CelebA-HQ.

Contribution

The paper proposes replacing the traditional real/fake discriminator with a One-Vs-All classifier in GANs for enhanced conditional image generation.

Findings

GAN-OVA improves training stability over regular conditional GANs.

GAN-OVA accelerates class-specific image generation.

GAN-OVA enhances image quality in experiments.

Abstract

This paper explores conditional image generation with a One-Vs-All classifier based on the Generative Adversarial Networks (GANs). Instead of the real/fake discriminator used in vanilla GANs, we propose to extend the discriminator to a One-Vs-All classifier (GAN-OVA) that can distinguish each input data to its category label. Specifically, we feed certain additional information as conditions to the generator and take the discriminator as a One-Vs-All classifier to identify each conditional category. Our model can be applied to different divergence or distances used to define the objective function, such as Jensen-Shannon divergence and Earth-Mover (or called Wasserstein-1) distance. We evaluate GAN-OVAs on MNIST and CelebA-HQ datasets, and the experimental results show that GAN-OVAs make progress toward stable training over regular conditional GANs. Furthermore, GAN-OVAs effectively accelerate the generation process of different classes and improves generation quality.