CycleGAN with Better Cycles

TL;DR

This paper introduces three modifications to the cycle consistency mechanism in CycleGAN, improving image translation quality by reducing artifacts and producing more realistic images without requiring paired datasets.

Contribution

The paper proposes three simple modifications to the cycle consistency loss in CycleGAN, enhancing image quality and reducing artifacts in unpaired image-to-image translation.

Findings

Reduced artifacts in generated images

Improved realism in translated images

Fewer cycle consistency violations

Abstract

CycleGAN provides a framework to train image-to-image translation with unpaired datasets using cycle consistency loss [4]. While results are great in many applications, the pixel level cycle consistency can potentially be problematic and causes unrealistic images in certain cases. In this project, we propose three simple modifications to cycle consistency, and show that such an approach achieves better results with fewer artifacts.