AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer

[Paper] [PyTorch Implementation] [Paddle Implementation]

Overview

This repository contains the official PyTorch implementation of paper:

AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer,

Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, Errui Ding

ICCV 2021

Prerequisites

• Linux or macOS
• Python 3
• PyTorch 1.7+ and other dependencies (torchvision, visdom, dominate, and other common python libs)

Getting Started

• Clone this repository:

  git clone https://github.com/Huage001/AdaAttN
  cd AdaAttN

• Inference:

  • Make a directory for checkpoints if there is not:

    mkdir checkpoints

  • Download pretrained model from Google Drive, move it to checkpoints directory, and unzip:

    mv [Download Directory]/AdaAttN_model.zip checkpoints/
    unzip checkpoints/AdaAttN_model.zip
    rm checkpoints/AdaAttN_model.zip

  • Configure content_path and style_path in test_adaattn.sh firstly, indicating paths to folders of testing content images and testing style images respectively.

  • Then, simply run:

    bash test_adaattn.sh

  • Check the results under results/AdaAttN folder.

• Train:

  • Download COCO dataset and WikiArt dataset and then extract them.

  • Configure content_path and style_path in train_adaattn.sh, indicating paths to folders of training content images and training style images respectively.

  • Before training, start visdom server:

    python -m visdom.server

  • Then, simply run:

    bash train_adaattn.sh

  • You can monitor training status at http://localhost:8097/ and models would be saved at checkpoints/AdaAttN folder.

  • You may feel free to try other training options written in train_adaattn.sh.

Citation

• If you find ideas or codes useful for your research, please cite:

  @inproceedings{liu2021adaattn,
    title={AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer},
    author={Liu, Songhua and Lin, Tianwei and He, Dongliang and Li, Fu and Wang, Meiling and Li, Xin and Sun, Zhengxing and Li, Qian and Ding, Errui},
    booktitle={Proceedings of the IEEE International Conference on Computer Vision},
    year={2021}
  }
  

Acknowledgments

• This implementation is developed based on the code framework of pytorch-CycleGAN-and-pix2pix by Junyan Zhu et al.