stylegan3-anime-generator icon indicating copy to clipboard operation
stylegan3-anime-generator copied to clipboard

Official PyTorch implementation of StyleGAN3

StyleGAN3 anime picture generator
Training with LoliHouse dataset

Introduction

The code is based on the official implementation stylegan3, with some details modified.

Using a large number of full body anime character images to train which get an anime character generator.

Please refer to the official code for the usage of the model.

Pre-trained models

Config

kimg FID link
StyleGAN3‑R 10800 7.51 kimg_10800.pkl
StyleGAN3‑R (stage2) 10800+1600 6.97 kimg_10800+1600.pkl

LoliHouse Dataset

The training dataset was constructed based on the image package provided by LoliHouse. The 83,786 images obtained by randomly coring the images in the package by square boxes were used as the training dataset.

The resolution of the dataset is 1024x1024.

Dataset links: (Ali cloud disk has restrictions on sharing, so it is spread over two cloud disks)

Original image package: LoliHouse

Cite the Dataset

@misc{
  lolihouse2022,
  author = {ZiYi Dong},
  title = {LoliHouse Anime Character Dataset},
  howpublished = {\url{https://github.com/7eu7d7/stylegan3-anime-generator}},
  url = {https://github.com/7eu7d7/stylegan3-anime-generator},
  type = {dataset},
  year = {2022},
  month = {February}
}