GLStyleNet

[update 1/12/2022]

paper: GLStyleNet: Exquisite Style Transfer Combining Global and Local Pyramid Features, published in IET Computer Vision 2020.

Arxiv paper: GLStyleNet: Higher Quality Style Transfer Combining Global and Local Pyramid Features.

Environment Required:

• Python 3.6
• TensorFlow 1.4.0
• CUDA 8.0

Getting Started:

Step 1: clone this repo

git clone https://github.com/EndyWon/GLStyleNet
cd GLStyleNet

Step 2: download pre-trained vgg19 model

bash download_vgg19.sh

Step 3: run style transfer

1. Script Parameters

• --content : content image path
• --content-mask : content image semantic mask
• --style : style image path
• --style-mask : style image semantic mask
• --content-weight : weight of content, default=10
• --local-weight : weight of local style loss
• --semantic-weight : weight of semantic map constraint
• --global-weight : weight of global style loss
• --output : output image path
• --smoothness : weight of image smoothing scheme
• --init : image type to initialize, value='noise' or 'content' or 'style', default='content'
• --iterations : number of iterations, default=500
• --device : devices, value='gpu'(all available GPUs) or 'gpui'(e.g. gpu0) or 'cpu', default='gpu'
• --class-num : count of semantic mask classes, default=5

2. portrait style transfer (an example)

python GLStyleNet.py --content portrait/Seth.jpg --content-mask portrait/Seth_sem.png --style portrait/Gogh.jpg --style-mask portrait/Gogh_sem.png --content-weight 10 --local-weight 500 --semantic-weight 10 --global-weight 1 --init style --device gpu

！！！You can find all the iteration results in folder 'outputs'！！！

3. Chinese ancient painting style transfer (an example)

python GLStyleNet.py --content Chinese/content.jpg --content-mask Chinese/content_sem.png --style Chinese/style.jpg --style-mask Chinese/style_sem.png --content-weight 10 --local-weight 500 --semantic-weight 2.5 --global-weight 0.5 --init content --device gpu

4. artistic and photo-realistic style transfer

artistic:

photo-realistic:

Citation:

If you find this code useful for your research, please cite the paper:

@article{wang2020glstylenet,
  title={GLStyleNet: exquisite style transfer combining global and local pyramid features},
  author={Wang, Zhizhong and Zhao, Lei and Lin, Sihuan and Mo, Qihang and Zhang, Huiming and Xing, Wei and Lu, Dongming},
  journal={IET Computer Vision},
  volume={14},
  number={8},
  pages={575--586},
  year={2020},
  publisher={IET}
}

Acknowledgement:

The code was written based on Champandard's code.