Manifold Alignment for Semantically Aligned Style Transfer

[Paper]

Getting Started

MAST has been tested on CentOS 7.6 with python >= 3.6. It supports both GPU and CPU inference. If you don't have a suitable device, try running our Colab demo.

Clone the repo:

git clone https://github.com/NJUHuoJing/MAST.git

Prepare the checkpoints:

download checkpoints in checkpoints.zip and unzip it into the root path of the project.

Install the requirements:

conda create -n mast-env python=3.6
conda activate mast-env
pip install -r requirements.txt

# If you want to use post smoothing as the same as PhotoWCT, then install the requirements below;
# You can also just skip it to use fast post smoothing, remember to change cfg.TEST.PHOTOREALISTIC.FAST_SMOOTHING=true
pip install -U setuptools
pip install cupy
pip install pynvrtc

Running the Demo

Artistic style transfer

First set MAST_CORE.ORTHOGONAL_CONSTRAINT=false in configs/config.yaml. Then use the script test_artistic.py to generate the artistic stylized image by following the command below:

# not use seg
python test_artistic.py --cfg_path configs/config.yaml --content_path data/default/content/4.png --style_path data/default/style/4.png --output_dir results/test/default

# use --content_seg_path and --style_seg_path to user edited style transfer
python test_artistic.py --cfg_path configs/config.yaml --content_path data/default/content/4.png --style_path data/default/style/4.png --output_dir results/test/default --content_seg_path data/default/content_segmentation/4.png --style_seg_path data/default/style_segmentation/4.png --seg_type labelme --resize 512

Photo-realistic style transfer

First set MAST_CORE.ORTHOGONAL_CONSTRAINT=true in configs/config.yaml. Then use the script test_photorealistic.py to generate the photo-realistic stylized image by following the command below:

# not use seg
python test_photorealistic.py --cfg_path configs/config.yaml --content_path data/photo_data/content/in1.png --style_path data/photo_data/style/tar1.png --output_dir results/test/photo --resize 512

# or use --content_seg_path and --style_seg_path to user edited style transfer
python test_photorealistic.py --cfg_path configs/config.yaml --content_path data/photo_data/content/in1.png --style_path data/photo_data/style/tar1.png --output_dir results/test/photo --content_seg_path data/photo_data/content_segmentation/in1.png --style_seg_path data/photo_data/style_segmentation/tar1.png --seg_type dpst --resize 512

GUI For Artistic style transfer and User Editing

We provide a gui for user-controllable artistic image stylization. Just use the command below to run test_gui.py

python test_gui.py --cfg_path configs/config.yaml

Features

1. You can use different colors to control the style transfer in different semantic areas.
2. The button Expand and Expand num respectively control whether to expand the selected semantic area and the degree of expansion.

See the gif demo for more details.

Google Colab

If you do not have a suitable environment to run this project then you could give Google Colab a try. It allows you to run the project in the cloud, free of charge. You may try our Colab demo using the notebook we have prepared: Colab Demo

Citation

@inproceedings{huo2021manifold,
    author = {Jing Huo and Shiyin Jin and Wenbin Li and Jing Wu and Yu-Kun Lai and Yinghuan Shi and Yang Gao},
    title = {Manifold Alignment for Semantically Aligned Style Transfer},
    booktitle = {IEEE International Conference on Computer Vision},
    pages     = {14861-14869},
    year = {2021}
}

References

• The post smoothing module is borrowed from PhotoWCT