Arbitrary-Style-Transfer-via-Multi-Adaptation-Network

Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, Changsheng Xu

results presentation

Stylized result using Claude Monet's painting as style reference. Compared with some state-of-the-art algorithms, our result can preserve detailed content structures and maintain vivid style patterns.

Framework

System overview. For the purpose of arbitrary style transfer, we propose a feed-forward network, which contains an encoder-decoder architecture and a multi-adaptation module.

The multi-adaptation module is divided into three parts: position-wise content SA module, channel-wise style SA module, and CA module.

Experiment

Requirements

• python 3.6
• pytorch 1.4.0
• PIL, numpy, scipy
• tqdm
  

Testing

Pretrained models: vgg-model, decoder, MA_module
Please download them and put them into the floder ./model/

python test.py  --content_dir input/content/ --style_dir input/style/    --output out

Training

Traing set is WikiArt collected from WIKIART

Testing set is COCO2014

python train.py --style_dir ../../datasets/Images/ --content_dir ../../datasets/train2014 --save_dir models/ --batch_size 4

Reference

If you use our work in your research, please cite us using the following BibTeX entry ~ Thank you ^ . ^. Paper Link pdf

@inproceedings{deng:2020:arbitrary,
  title={Arbitrary Style Transfer via Multi-Adaptation Network},
  author={Deng, Yingying and Tang, Fan and Dong, Weiming and Sun, Wen, and Huang, Feiyue and Xu, Changsheng},
  booktitle={Acm International Conference on Multimedia},
  year={2020},
 publisher = {ACM},
}