Respecting low-level components of content with skip connections and semantic information in image style transfer (CVMP’19)

Man M. Ho, Jinjia Zhou, Yibo Fan.

[ DEMO Video], [Comparison Video]

This repository is to support our Paper/DOI: https://doi.org/10.1145/3359998.3369403.

News

Date News
2020/04/04 Code and trained models now available.
2019/12/15 Add additional results.

Environment Setup

We run our DEMO on Anaconda Environment with following packages:

Get started

Clone this repo:

git clone https://github.com/minhmanho/respecting_content_ist.git
cd respecting_content_ist

Download models

bash ./models/fetch_models.sh

Stylize a folder of images:

python run.py --img_folder <folder_dir> --model <model_dir>

Run GUI DEMO:

python run.py --img <img_dir> --model <model_dir> --auto_update

If you have a powerful GPU, let’s add –auto_update for the better experience. The application will update every change you make automatically if auto_update is activated; otherwise, you have to click the button [Update].

Addtional comparison results

Content

Most of the content images are collected from Unsplash

Compare to Johnson’s work + Instance Normalization

Source code

[Their work] [Ours] [Their work] [Ours]

: Candy

: Udnie

Compare to recent works

[AdaIn] [WCT] [AvatarNet] [Ours]

: Candy

: Watercolor painting portrait of a woman

: Seated Nude by Pablo Picasso

: Watercolor painting of rabbit

: A sketch

: Udnie

License

This repository (as well as its materials) is for non-commercial uses and research purposes only.

Cite

@inproceedings{ho2019respecting,
  title={Respecting low-level components of content with skip connections and semantic information in image style transfer},
  author={Ho, Minh-Man and Zhou, Jinjia and Fan, Yibo},
  booktitle={European Conference on Visual Media Production},
  pages={1--9},
  year={2019}
}

Contact

If you have any suggestions, questions, or the use of these images infringe your copyrights/license, please contact me man.hominh.6m@stu.hosei.ac.jp. I will take action ASAP.

Acknowledgements

We refer the Interactive Colorization to build our demonstration. This work is supported by the State Key Laboratory of ASIC & System, Fudan University, China.