Multimodal image-to-image translation via a single generative adversarial network

Our model architecture is defined as depicted bellow, please refer to the paper for more details:

Usage Guidance

Dependencies

1. python 3.x
2. pytorch 4.0+

Testing

• Runing the following command to translate edges to shoes&handbags (the pretrained models are stored in ./checkpoints/edges_shoes&handbags directory):

python ./test.py --name edges_shoes&handbags --d_num 2

Then the translated samples are stored in ./checkpoints/edges_shoes&handbags/edges_shoes&handbags_results directory. By default, it produce 5 random translation outputs.

Training

• Download the dataset you want to use and move to ./datasets. For example, you can use the horse2zebra dataset provided by CycleGAN. Please make sure that you have the following directory tree structure in your repository

├── datasets
│   └── horse2zebra
│       ├── trainA
│       ├── testA
│       ├── trainB
│       ├── testB

The Animals With Attributes (AWA) dataset can be downloaded from hear.

• Start training with the following command:

python ./train.py --name horse2zebra --d_num 2

Intermediate image outputs and model binary files are stored in ./checkpoints/horse2zebra/web

Results

Edges ↔ Shoes&handbags:

Horse ↔ Zebra:

Cat ↔ Dog ↔ Tiger:

Leopard ↔ Lion ↔ Tiger:

Photos ↔ Vangogh ↔ Monet ↔ Cezanne:

bibtex

If this work helps to easy your research, please cite this paper :

@article{huang2022multimodal,
  title={Multimodal image-to-image translation via a single generative adversarial network},
  author={Huang, Shihua and He, Cheng and Cheng, Ran},
  journal={IEEE Transactions on Artificial Intelligence},
  year={2022},
  publisher={IEEE}
}

Acknowledgment

The code used in this research is based on SingleGAN and CycleGAN.