AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer

[Paper] [PyTorch Implementation] [Paddle Implementation]

Overview

This repository contains the officially unofficial PyTorch re-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

Updates

• [2022-12-07] Upload script of user control. Please see user_specify_demo.py
• [2022-12-07] Upload inference code of video style transfer. Please see inference_frame.py. Please download checkpoints from here and extract the package to the main directory of this repo before running.

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:

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

• Inference:

  • Make a directory for checkpoints if there is not:

  shell mkdir checkpoints

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

  shell 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:

  shell bash test_adaattn.sh

  • Check the results under results/AdaAttN folder.

• Train:

  • Download 'vgg_normalised.pth' from here.

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

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

  • Before training, start visdom server:

  shell python -m visdom.server

  • Then, simply run:

  shell 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.