AniGAN: Style-Guided Generative Adversarial Networks for Unsupervised Anime Face Generation

Bing Li¹, Yuanlue Zhu², Yitong Wang², Chia-Wen Lin³, Bernard Ghanem¹, Linlin Shen⁴

¹Visual Computing Center, KAUST, Thuwal, Saudi Arabia
²ByteDance, Shenzhen, China
³Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan
⁴Computer Science and Software Engineering, Shenzhen University, Shenzhen, China

conda create -n AniGAN python=3.7
conda activate AniGAN
conda install pytorch==1.4.0 torchvision==0.5.0 -c pytorch
pip install PyYAML

• Download pre-trained weights (trained on our face2anime dataset) from Google Drive and put it in src/checkpoints/try4_final_r1p2/.
• Our method with pretrained weights can be evaluated using the test.py:

python test.py \
    --source_img SOURCE_IMG_PATH \
    --reference_img REFERENCE_IMG_PATH \
    --output_dir SAVE_PATH

• SOURCE_IMG_PATH: path of the source image. Two examples are provided: input_img_examples/1795111.png and input_img_examples/2773005.png.
• REFERENCE_IMG_PATH: path of the reference image. Two examples are provided: input_img_examples/imgHQ11831.png and input_img_examples/imgHQ11853.png.
• SAVE_PATH: directory path to save the result image. Optional and default value is result_dir.

We build a new dataset called face2anime, which is larger and contains more diverse anime styles (e.g., face poses, drawing styles, colors, hairstyles, eye shapes, strokes, facial contours) than selfie2anime. The face2anime dataset contains 17,796 images in total, where the number of both anime-faces and natural photo-faces is 8,898. The anime-faces are collected from the Danbooru2019 dataset, which contains many anime characters with various anime styles. We employ a pretrained cartoon face detector to select images containing anime-faces. For natural-faces, we randomly select 8,898 female faces from the CelebA-HQ dataset. All images are aligned with facial landmarks and are cropped to size 128 × 128. We separate images from each domain into a training set with 8,000 images and a test set with 898 images.

You can download the face2anime dataset from Google Drive.

If you find this work useful or use the face2anime dataset, please cite our paper:

@article{li2021anigan,
  title={AniGAN: Style-Guided Generative Adversarial Networks for Unsupervised Anime Face Generation},
  author={Li, Bing and Zhu, Yuanlue and Wang, Yitong and Lin, Chia-Wen and Ghanem, Bernard and Shen, Linlin},
  journal={IEEE Transactions on Multimedia},
  year={2021},
  publisher={IEEE}
}