Hairstyle Transfer — Semantic Editing GAN Latent Code

Abstract

Motivated by the success of StyleGAN, where stochastic variation is incorporated in generating realistic-looking images, we proposed to focus on the  hairstyle attributes of a face. The right hairstyle can often only be discovered through trials and errors. Thus, being able to virtually “try on” a novel hairstyle through a computer vision system seems to hold practical value in reality.

In this project, we propose an end-to-end workflow for editing hair attributes on real faces. Hairstyle Transfer leverages fixed pre-trained GAN models, GAN encoders, and manipulations of the latent code for the semantic editing. Moreover, we further confirmed the linear separability assumption of hair-related semantic attributes.

Usage

There are three colab notebooks for this end-to-end workflow.

1. StyleGAN_Encoder Generate latent representations of your own images
2. Get Attribute Score Pairs Generate pairs of latent code and scores for boundary training later
3. Train Boundaries + Face Editing with Interface GAN Semantic editing with the boundary obtained

Report

Curious to learn more? Full report is now on the blog.

Reference

This implementation is based on StyleGAN and InterFaceGAN. 🎉