[Paper] [Paddle Implementation]

We have optimized the serial inference procedure to achieve better rendering quality and faster speed.

This repository contains the official PaddlePaddle implementation of paper:

Paint Transformer: Feed Forward Neural Painting with Stroke Prediction,

Songhua Liu*, Tianwei Lin*, Dongliang He, Fu Li, Ruifeng Deng, Xin Li, Errui Ding, Hao Wang (* indicates equal contribution)

ICCV 2021 (Oral)

• Linux or macOS

• Python 3.6+

• PaddlePaddle 2.0+ and other dependencies (numpy, cv2, and other common python libs)

  python -m pip install paddlepaddle-gpu

• Clone this repository:

  git clone https://github.com/wzmsltw/PaintTransformer
  cd PaintTransformer

• Download pretrained model from Google Drive and move it to inference directory:

  mv [Download Directory]/paint_best.pdparams inference/
  cd inference

• Inference:

  python inference.py

  • Input image path, output path, and etc can be set in the main function.
  • Notably, there is a flag serial as one parameter of the main function:
    • If serial is True, strokes would be rendered serially. The consumption of video memory will be low but it requires more time. Serial inference can achieve better rendering quality.
    • If serial is False, strokes would be rendered in parallel. The consumption of video memory will be high but it would be faster.
    • If animated results are required, serial must be True.

• Train:

  • You can send email to us for the training codes.

Input	Animated Output

• Do not want to run the code? Try an App 一刻相册 downloaded from here!

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

  @inproceedings{liu2021paint,
    title={Paint Transformer: Feed Forward Neural Painting with Stroke Prediction},
    author={Liu, Songhua and Lin, Tianwei and He, Dongliang and Li, Fu and Deng, Ruifeng and Li, Xin and Ding, Errui and Wang, Hao},
    booktitle={Proceedings of the IEEE International Conference on Computer Vision},
    year={2021}
  }
  

For any question, please file an issue or contact

Songhua Liu: [email protected]
Tianwei Lin: [email protected]