Anime-Face-Generation

This repository contains code for generating anime character faces using GAN's

Dataset

The dataset used in this repository is the courtesy of [1] and can be easily downloaded from here. The dataloader implementation is based on this tutorial.

Implementation

Model Architecture

Currently the generator and discriminator are a slightly modified version of the architectures mentioned in [3]. For generator network, input feature size is a 128 vector, no label embedding has been used as of now. The number of sub-pixel cnns is reduced to 2, since the input images are of size 64*64. For similar reasons, in discriminator network number of residual blocks has been reduced to 8 instead of 10. The discriminator outputs binary output representing wether the image is real or fake, unlike [3].

Training

For training the modified loss function mentioned in [2] is being used. The latent variable z is being sampled from a uniform distribution.

Output

Loss function:

Faces:

References

[1] @online{chao2019/online, author = {Brian Chao}, title = {Anime Face Dataset: a collection of high-quality anime faces.}, date = {2019-09-16}, year = {2019}, url = {https://github.com/bchao1/Anime-Face-Dataset} }

[2] arXiv:1406.2661

[3] arXiv:1708.05509