/Yumis_Cells_Generator

Repository for 2021 CUAI Conference Mobility A Team Project. About StyleGANs.

Primary LanguagePython

Yumi's Cells Generator

About Project

Yumi's Cells Genertor : Training various GANs with small Yumi dataset.
이보림 : Chung-ang Univ. C.S.E
조원 : Chung-ang Univ. E.E
최명균 : Chung-ang Univ. E.E
in CUAI 2021 Conference

Short Paper
Video

Abstract

GAN is hard to training model with less datasets. Also, it takes many time.
There are many pre-trained model with famous and large datasets ( FFHQ, AFHQ, CelebA-HQ ... ). However, in reality, it is very difficult to collect good datasets of interest.
We trained small datasets with popular Webtoon 'Yumi's Cells'. Then, We generated Cells' images and compared various stylegan models.
Finally, we found the effective way to train models and select model with small Yumi's dataset.

Short Results

StyleGAN + FreezeD

050000

StyleGAN2

image image

image image

image

StyleMapGAN

  • local editing ( Synthesized )

2 2

=>

2

Related Projects

StyleGAN1 : https://arxiv.org/abs/1812.04948
StyleGAN2 : https://arxiv.org/abs/1912.04958
FreezeD : https://arxiv.org/abs/2002.10964
FreezeG : https://github.com/bryandlee/FreezeG
StyleMapGAN : https://arxiv.org/abs/2104.14754