XOGAN: One-to-Many Unsupervised Image-to-Image Translation

TL;DR

XOGAN is a novel unsupervised image translation model that learns one-to-many mappings, allowing for diverse and controllable image generation across domains without paired data.

Contribution

Introduces the XOGAN model with an additional variable and XO-structure to enable one-to-many unsupervised image translation with controllable variations.

Findings

XOGAN generates diverse, plausible images with controllable variations.

It outperforms state-of-the-art methods in diversity of generated images.

Effective on edges-to-objects and facial image translation tasks.

Abstract

Unsupervised image-to-image translation aims at learning the relationship between samples from two image domains without supervised pair information. The relationship between two domain images can be one-to-one, one-to-many or many-to-many. In this paper, we study the one-to-many unsupervised image translation problem in which an input sample from one domain can correspond to multiple samples in the other domain. To learn the complex relationship between the two domains, we introduce an additional variable to control the variations in our one-to-many mapping. A generative model with an XO-structure, called the XOGAN, is proposed to learn the cross domain relationship among the two domains and the ad- ditional variables. Not only can we learn to translate between the two image domains, we can also handle the translated images with additional variations. Experiments are performed on unpaired image generation tasks, including edges-to-objects translation and facial image translation. We show that the proposed XOGAN model can generate plausible images and control variations, such as color and texture, of the generated images. Moreover, while state-of-the-art unpaired image generation algorithms tend to generate images with monotonous colors, XOGAN can generate more diverse results.