A GAN Demo: ACG-Style Faces Generating

Introduction

This is a GAN demo for creating anime character faces from random noise.

Prerequisite

Codes

• main.py
• gan.py
• utils.py
• generate.py
• dcgan.py
• nets.py
• show.html
• show.js
• spider.py

Main Codes

main.py contains training configurations. gan.py defines net structure. utils.py contains some auxiliary functions. generate.py generates anime faces.

Not Available Now

dcgan.py and nets.py are rebuild versions of main.py and gan.py (not completed). show.html and show.js are for future presentation on web via Keras.js.

Others

spider.py collects and downloads training images from the Internet, thanks for the provider Acokil!

Datasets

Datasets are not uploaded.

• faces.zip
• hqface.zip

faces.zip is an anime face dataset with the image shape of (96, 96, 3). They are collected from Konachan.

hqface.zip is also collected from Konachan, but contains images with higher quality.

Environment

OS

• Linux CentOS
• Windows 10

The linux mainly serves as training platform. Windows is for coding.

GPU

• Nvidia Tesla K40M

GPUs are from ZJUSPC. Thanks for the authorization of the usage of K40M from ZJUSPC!

References

Code references: GAN-Zoo

Paper references:

• Generative Adversarial Nets (arxiv id: 1406.2661v1)
• Conditional Generative Adversarial Nets (arxiv id: 1411.1784v1)
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (arxiv id: 1511.06434v2)

Results

Here are some generated faces. I used a 300-d random noise as input and trained for 40,000 iterations. In each iteration, I used 64 images to train the model.

As you can see, the quality of these faces is not good enough. In fact, for most of the generated images you can only recognize blurry faces so I just picked out some well-performed results. It is hard but worth to improve the model's performance.

Future

Try to improve performance via these approaches:

• Use high quality training images
• Use larger training images

I adjusted the GAN. Then I used dataset hqface with the shape of (112, 112, 3) of each image to train a new model. It performs better on generating images of higher resolution.

Here are some new examples generated by the new GAN.