WCT2 (ICCV 2019 accepted)

Photorealistic Style Transfer via Wavelet Transforms | paper | supplementary materials | video stylization results

Jaejun Yoo*, Youngjung Uh*, Sanghyuk Chun*, Byeonkyu Kang, Jung-Woo Ha

Clova AI Research, NAVER (* equal contributions)

PyTorch implementation for photorealistic style transfer that does not need any further post-processing steps; e.g. from day to sunset, from summer to winter, etc. This is the first end-to-end model that can stylize 1024×1024 resolution image in 4.7 seconds, giving a pleasing and photorealistic quality without any post-processing.

The code was written by Jaejun Yoo and Byeongkyu Kang.

Getting Started

Dependency

• PyTorch >= 0.4.1
• Check the requirements.txt

pip install -r requirements.txt

Installation

• Clone this repo:

git clone https://github.com/clovaai/WCT2.git
cd WCT2

• Pretrained models can be found in the ./model_checkpoints

• Prepare image dataset
  • Images can be found in DPST repo
    • You can find the entire content and style images (with paired segmentation label maps) in the following link DPST images. input folder has the content images and the style folder has the style images. Every segmention map can be found in the segmentation folder.
  • To make a new dataset with label pairs, please follow the instruction of PhotoWCT repo
  • Put the content and style images with their segment label pairs (if available) into the example folder accordingly.
    • Currently there are several example images so that you can execute the code as soon as you clone this repo.
• Finally, test the model:

python transfer.py --option_unpool cat5 -a --content ./examples/content --style ./examples/style --content_segment ./examples/content_segment --style_segment ./examples/style_segment/ --output ./outputs/ --verbose --image_size 512 

The test results will be saved to ./outputs by default.

Arguments

• --content: FOLDER-PATH-TO-CONTENT-IMAGES
• --content_segment: FOLDER-PATH-TO-CONTENT-SEGMENT-LABEL-IMAGES
• --style: FOLDER-PATH-TO-STYLE-IMAGES
• --style_segment: FOLDER-PATH-TO-STYLE-SEGMENT-LABEL-IMAGES
• --output: FOLDER-PATH-TO-OUTPUT-IMAGES
• --image_size: output image size
• --alpha: alpha determines the blending ratio between content and stylized features
• --option_unpool: two versions of our model (sum, cat5)
• -e, --transfer_at_encoder: stylize at the encoder module
• -d, --transfer_at_decoder: stylize at the decoder module
• -s, --transfer_at_skip: stylize at the skipped high frequency components
• -a, --transfer_all: stylize and save for every composition; i.e. power set of {-e,-d,-s})
• --cpu: run on CPU
• --verbose

Photorealistic Style Transfer

• DPST: "Deep Photo Style Transfer" | Paper | Code
• PhotoWCT: "A Closed-form Solution to Photorealistic Image Stylization" | Paper | Code
• PhotoWCT (full): PhotoWCT + post processing

Schematic illustration of our wavelet module

Component-wise Stylization

• Only for option_unpool = sum version
• Full stylization

python transfer.py --option_unpool sum -e -s --content ./examples/content --style ./examples/style --content_segment ./examples/content_segment --style_segment ./examples/style_segment/ --output ./outputs/ --verbose --image_size 512

• Low-frequency-only stylization

python transfer.py --option_unpool sum -e --content ./examples/content --style ./examples/style --content_segment ./examples/content_segment --style_segment ./examples/style_segment/ --output ./outputs/ --verbose --image_size 512

Results

• option_unpool = cat5 version

python transfer.py --option_unpool cat5 -a --content ./examples/content --style ./examples/style --content_segment ./examples/content_segment --style_segment ./examples/style_segment/ --output ./outputs/ --verbose --image_size 512

Acknowledgement

• Our implementation is highly inspired from NVIDIA's PhotoWCT Code.

Citation

If you find this work useful for your research, please cite:

@inproceedings{yoo2019photorealistic,  
  title={Photorealistic Style Transfer via Wavelet Transforms},
  author={Yoo, Jaejun and Uh, Youngjung and Chun, Sanghyuk and Kang, Byeongkyu and Ha, Jung-Woo},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year={2019}
}

Contact

Feel free to contact me if there is any question (Jaejun Yoo jaejun.yoo@navercorp.com).

License

Copyright (c) 2019 NAVER Corp.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.