Unsupervised Image-to-Image Translation Networks

https://arxiv.org/pdf/1703.00848.pdf
Kind of like CycleGAN

Novelty

• Unsupervise
• Two domain share the same latent space

Problem

• Only two domains at a time, need to train new model for new domain

Structure

Structure

Loss

Loss

Dataset

Street scene, Dog breed, Cat Species, Attribute-based face