wuga214 / IMPLEMENTATION_Wasserstein_GAN_Anime

This is a simple implementation of Wassertein GAN. The dataset used in testing is Anime that collected from Japanese anime websites.

Wassertein Generative Adversarial Network(WGAN)

This is simple demo of Wassertein GAN on Anime dataset. The model needs around 20000 iterations to get reasonable result. The running time could be around 3 hours on GTX 1070 GPU.

How to start

$python main.py

Data

Anime Data was crawled from Japanese Anime Websites randomly and preprocessed through anime face detection code, for example lbpcascade. I collected around 20000 images as training data. Please do not ask me for sharing the data due to potential copyright issue.

Stability

Before using Wassertein GAN, I tried multiple hacking technology such as mini-batch discrimination and tricks. But, unfortunately, none of them really works in general but really depends on how lucky you are.

Demo