Home

Awesome

ASMA-GAN

Anisotropic Stroke Control for Multiple Artists Style Transfer

Proceedings of the 28th ACM International Conference on Multimedia

The official repository with Pytorch

Top News <img width=8% src="./doc/img/new.gif"/>

2021-09-25: Training related code has been released.

[Arxiv paper]

logo

title

Methodology

Framework

Dependencies

Installation

We highly recommend you to use Anaconda for installation

conda create -n ASMA python=3.6
conda activate ASMA
conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch
pip install pyyaml paramiko pandas requests tensorboard tensorboardX tqdm

Preparation

Usage

To test with pretrained model

The command line below will generate 1088*1920 HD style migration pictures of 11 painters for each picture of testImgRoot (11 painters include: Berthe Moriso , Edvard Munch, Ernst Ludwig Kirchner, Jackson Pollock, Wassily Kandinsky, Oscar-Claude Monet, Nicholas Roerich, Paul Cézanne, Pablo Picasso ,Samuel Colman, Vincent Willem van Gogh. The output image(s) can be found in ./test_logs/ASMAfinal/

--version refers to the ASMAGAN training logs name.

--testImgRoot can be a folder with images or the path of a single picture.You can assign the image(s) you want to perform style transfer to this argument.

--specify_sytle is used to specify which painter's style is used for style transfer. When the value is -1, 11 painters' styles are used for image(s) respectively for style transfer. The values corresponding to each painter's style are as follows [0: Berthe Moriso, 1: Edvard Munch, 2: Ernst Ludwig Kirchner, 3: Jackson Pollock, 4: Wassily Kandinsky, 5: Oscar-Claude Monet, 6: Nicholas Roerich, 7: Paul Cézanne, 8: Pablo Picasso, 9 : Samuel Colman, 10: Vincent Willem van Gogh]

Training

To train your own model, first change the dataset path in ./env/config.json.

Then use:

python main.py --mode train --cuda 0 --dataloader_workers 12 --version $(your experiment name) --trainYaml train.yaml

Change the training parameters in ./train_configs/train.yaml.

To cite our paper

@inproceedings{DBLP:conf/mm/ChenYLQN20,
  author    = {Xuanhong Chen and
               Xirui Yan and
               Naiyuan Liu and
               Ting Qiu and
               Bingbing Ni},
  title     = {Anisotropic Stroke Control for Multiple Artists Style Transfer},
  booktitle = {{MM} '20: The 28th {ACM} International Conference on Multimedia, 2020},
  publisher = {{ACM}},
  year      = {2020},
  url       = {https://doi.org/10.1145/3394171.3413770},
  doi       = {10.1145/3394171.3413770},
  timestamp = {Thu, 15 Oct 2020 16:32:08 +0200},
  biburl    = {https://dblp.org/rec/conf/mm/ChenYLQN20.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Some Results

Results1

Related Projects

Learn about our other projects [RainNet], [Sketch Generation], [CooGAN], [Knowledge Style Transfer], [SimSwap],[ASMA-GAN],[Pretrained_VGG19].

High Resolution Results

<img src="./doc/img/7.png" style="zoom: 20%;"/> <img src="./doc/img/8.png" style="zoom: 20%;"/> <img src="./doc/img/9.png" style="zoom: 20%;"/> <img src="./doc/img/10.png" style="zoom: 20%;"/> <img src="./doc/img/11.png" style="zoom: 20%;"/> <img src="./doc/img/12.png" style="zoom: 20%;"/> <img src="./doc/img/13.png" style="zoom: 12%;"/> <img src="./doc/img/14.png" style="zoom: 12%;"/>