Sym-Parameterized Dynamic Inference for Mixed-Domain Image Translation

This repository is official Pytorch implementations of SGN in the following paper.

Simyung Chang, SeongUk Park, John Yang and Nojun Kwak, "Sym-Parameterized Dynamic Inference for Mixed-Domain Image Translation", ICCV2019, arXiv

This code is built on Pytorch-CycleGAN and tested on Pytorch 0.4.1.

Dynamic Inference for Mixed-Domain Image Translation.

The concept of sym-parameter.

Overall Structure of SGN for Three Different Losses.

Video translation result of SGN : SGN-Video

Train

Prepare training data

./download_dataset <dataset_name>

<dataset_name> for the models in the paper.

Model 1: vangogh2photo

Model 2: ukiyoe2photo

Model 3 : summer2winter_yosemite, monet2photo

Train

python train.py --dataroot <data_dir> --style_image <image_file>

To train the Model 1

python train.py --dataroot 'datasets/vangogh2photo/' --style_image 'images/style-images/udnie.jpg'

Test

The pre-trained models can be downloaded from Google Drive.

Copy the models in pretrained/

Test

python test.py --dataroot <data_dir> --generator <trained_file>

Test Model1 with pre-trained check point.

python test.py --dataroot 'datasets/vangogh2photo' --generator 'pretrained/Model1.pth'

Citation

@article{chang2018image,
  title={Image Translation to Mixed-Domain using Sym-Parameterized Generative Network},
  author={Chang, Simyung and Park, SeongUk and Yang, John and Kwak, Nojun},
  journal={arXiv preprint arXiv:1811.12362},
  year={2018}
}

Acknowledgments

Our code is built on Pytorch-CycleGAN.