Multi-Style Photo Cartoonization

Pytorch implement of Yezhi Shu, Ran Yi, Mengfei Xia, Zipeng Ye, Wang Zhao, Yang Chen, Yu-Kun Lai, Yong-Jin Liu. GAN-based Multi-Style Photo Cartoonization. IEEE Transactions on Visualization and Computer Graphics, DOI (identifier) 10.1109/TVCG.2021.3067201, 2021.

Our Proposed Framework

Results

Installation

Install PyTorch 1.2.0+ and torchvision from http://pytorch.org and other dependencies could be downloaded by

pip install -r requirements.txt

Train

• Prepare your own dataset and put it under the dataset folder. Note that your own dataset should contain subfoloders: train0, train1[style1name], train2[style2name], ..., edge1[style1name], edge2[style2name], ..., test0. Different style names corresponding to different styles you have collected.

  ├─datasets
  ├─dataset name
    ├─ edge1_[style1name]
    ├─ edge2_[style2name]
    ├─ edge3_[style3name]
    ├─ test0
    ├─ train0
    ├─ train1_[style1name]
    ├─ train2_[style2name]
    └─ train3_[style3name]

• Download the vgg19.pth from BaiduYun (https://pan.baidu.com/s/1zHx06jMfIUIdBY_X3ko6LA, extract code is z9p7) or Googledrive (https://drive.google.com/drive/folders/1jOSl_sBpGzm1wAvLgU5vDJOV9tg2PzIc?usp=sharing). Put it under Multi-Style Photo Cartoonization folder.
• Train the model using the commond. Set --init to use initialization phase. If you have pretrained model, you can set --continue_train and --which_epoch simultaneously.

  python train.py --init --name train --dataroot [your dataset path]

Test

• You can use your own trained model to test the model, also you can download pretrained models by the link below.
• Put test real-world photo in test0 folder of your dataset.

  python test.py --name train --serial_test --dataroot [your dataset path] --which_epoch 145

Citation

Reference code

• CycleGAN (https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix)
• CartoonGAN (https://github.com/FlyingGoblin/CartoonGAN)
• ComboGAN (https://github.com/AAnoosheh/ComboGAN)