Skip to content

Latest commit

 

History

History
83 lines (47 loc) · 3.2 KB

README.md

File metadata and controls

83 lines (47 loc) · 3.2 KB
Raw

AesUST (ACM MM 2022)

[update 8/28/2022]

Official Pytorch code for "AesUST: Towards Aesthetic-Enhanced Universal Style Transfer"

Introduction:

AesUST is a novel Aesthetic-enhanced Universal Style Transfer approach that can generate aesthetically more realistic and pleasing results for arbitrary styles. It introduces an aesthetic discriminator to learn the universal human-delightful aesthetic features from a large corpus of artist-created paintings. Then, the aesthetic features are incorporated to enhance the style transfer process via a novel Aesthetic-aware Style-Attention (AesSA) module. Moreover, we also develop a new two-stage transfer training strategy with two aesthetic regularizations to train our model more effectively, further improving stylization performance.

show

Environment:

  • Python 3.6
  • Pytorch 1.8.0

Getting Started:

Clone this repo:

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

Test:

  • Download pre-trained models from this google drive. Unzip and place them at path models/.

  • Test a pair of images:

    python test.py --content inputs/content/1.jpg --style inputs/style/1.jpg

  • Test two collections of images:

    python test.py --content_dir inputs/content/ --style_dir inputs/style/

Train:

  • Download content dataset MS-COCO and style dataset WikiArt and then extract them.

  • Download the pre-trained vgg_normalised.pth, place it at path models/.

  • Run train script:

    python train.py --content_dir ./coco2014/train2014 --style_dir ./wikiart/train

Runtime Controls:

Content-style trade-off:

python test.py --content inputs/content/1.jpg --style inputs/style/1.jpg --alpha 0.5

show

Style interpolation:

python test.py --content inputs/content/1.jpg --style inputs/style/30.jpg,inputs/style/36.jpg --style_interpolation_weights 0.5,0.5

show

Color-preserved style transfer:

python test.py --content inputs/content/1.jpg --style inputs/style/1.jpg --preserve_color

show

Citation:

If you find the ideas and codes useful for your research, please cite the paper:

@inproceedings{wang2022aesust,
  title={AesUST: towards aesthetic-enhanced universal style transfer},
  author={Wang, Zhizhong and Zhang, Zhanjie and Zhao, Lei and Zuo, Zhiwen and Li, Ailin and Xing, Wei and Lu, Dongming},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={1095--1106},
  year={2022}
}

Acknowledgement:

We refer to some codes from SANet and IEContraAST. Great thanks to them!