---

Overview

This project is fork of TensorFlow implementation of StarGAN. StarGAN is a unified Generative Adversarial Networks (GANs) for multi-domain image-to-image translation. What this means is, StarGAN can flexibly translate an input image to any desired target domain using only a single generator and a discriminator. You can take a look at the demo video for StarGAN here.

Requirements

• Tensorflow 1.8
• Python 3.6

Usage

Downloading the dataset

> python download.py celebA

├── dataset
   └── celebA
       ├── train
           ├── 000001.jpg 
           ├── 000002.jpg
           └── ...
       ├── test (It is not celebA)
           ├── a.jpg (The test image that you wanted)
           ├── b.png
           └── ...
       ├── list_attr_celeba.txt (For attribute information) 

Train

• python main.py --phase train

Test

• python main.py --phase test
• The celebA test image and the image you wanted run simultaneously

Pretrained model

• Download checkpoint for 128x128

Summary

Results (128x128, wgan-gp)

Women

Men

Reference

• StarGAN paper
• Author pytorch code

Author

Junho Kim