Aesthetic-Aware Image Style Transfer

by Zhiyuan Hu, Jia Jia, Bei Liu, Yaohua Bu and Jianlong Fu

Paper link: http://hcsi.cs.tsinghua.edu.cn/Paper/Paper20/MM20-HUZHIYUAN.pdf

This is the pytorch implementation of Aesthetic-Aware Image Style Transfer [Hu et al., MM2020].

Our work propose a novel problem called Aesthetic-Aware Image Style Transfer(AAST) which aims to control the texture and color of an image independently during style transfer and thus generate results with more diverse aesthetic effects.

Our code framework refers to the framework of pytorch-AdaIN.

Results

Results on Aesthetic Aware Image Style Transfer
Results on parameter Interpolation

Usage

1.Prerequisite

python 3.7+
pytorch 1.4+
torchvision 0.5+
Pillow

(Optional)

CUDA 10.0
tqdm
TensorboardX

2. Download Pre-trained model

Download vgg_normalized.pth(required for training) and net_final.pth(for testing only) and put them under models/

vgg_normalized.pth:

Google Drive: https://drive.google.com/file/d/1YOL3eyavRvxgM0GA19mWQUSg-oofAKb1/view?usp=sharing

net_final.pth:

Google Drive:https://drive.google.com/file/d/1e7ph-4cf8OKCp8nitXqHcOT0C9_otM1O

3. Test

Use --content_dir, --texture_dir and --color_dir to specify directories that save content images, texture reference images and color reference images. The model will iterate over all combinations between content, texture and color.

Use --test_opt to specify the type of test you want to conduct:

TC: Transfer texture and color together.
ST: Traditional style transfer.
T: Texture only transfer.
C: Color only transfer.
INT: Parameter interpolation for texture and color. For this type of test, you need to specify the interpolation num by --int_num

python main.py --mode test --test_opt <TEST_OPTION> --content_dir <CONTENT_DIR> --texture_dir <TEXTURE_DIR> --color_dir <COLOR_DIR>

For more detailed configurations, please refer to --help option.

4. Train

Use --content_dir, --texture_dir and --color_dir to specify directories that save content images, texture reference images and color reference images.

In each iteration, the model will randomly sample a batch of content-texture-color pair for training. The training will stop when it reaches the maximum iteration num (specified by --max_iter). Usually the training will not iterate over the whole dataset as the num of combinations of content, texture and color is really large.

python main.py --mode train --content_dir <CONTENT_DIR> --texture_dir <TEXTURE_DIR> --color_dir <COLOR_DIR>

For more detailed configurations, please refer to --help option.

Citation

If you find this repo useful in your research, please consider citing the following papers:

@inproceedings{hu2020aesthetic,
  title={Aesthetic-Aware Image Style Transfer},
  author={Hu, Zhiyuan and Jia, Jia and Liu, Bei and Bu, Yaohua and Fu, Jianlong},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={3320--3329},
  year={2020}
}

Contact

If you have any questions or suggestions about this paper, feel free to contact me (z8hu@ucsd.edu).

cyinen/AAST