APDrawingGAN

We provide PyTorch implementations for our CVPR 2019 paper "APDrawingGAN: Generating Artistic Portrait Drawings from Face Photos with Hierarchical GANs".

This project generates artistic portrait drawings from face photos using a GAN-based model. You may find useful information in preprocessing steps and training/testing tips.

[Paper] [Demo]

Our Proposed Framework

Sample Results

Up: input, Down: output

Citation

If you use this code for your research, please cite our paper.

@inproceedings{YiLLR19,
  title     = {{APDrawingGAN}: Generating Artistic Portrait Drawings from Face Photos with Hierarchical GANs},
  author    = {Yi, Ran and Liu, Yong-Jin and Lai, Yu-Kun and Rosin, Paul L},
  booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition (CVPR '19)},
  pages     = {10743--10752},
  year      = {2019}
}

Prerequisites

• Linux or macOS
• Python 2.7
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

• Install PyTorch 0.4+ and torchvision from http://pytorch.org and other dependencies (e.g., visdom and dominate). You can install all the dependencies by

pip install -r requirements.txt

Quick Start (Apply a Pre-trained Model)

• Download a pre-trained model (using 70 pairs in training set and augmented data) from https://cg.cs.tsinghua.edu.cn/people/~Yongjin/APDrawingGAN-Models1.zip (Model1) and put it in checkpoints/formal_author.

• Then generate artistic portrait drawings for example photos in dataset/data/test_single using

python test.py --dataroot dataset/data/test_single --name formal_author --model test --dataset_mode single --norm batch --use_local --which_epoch 300

The test results will be saved to a html file here: ./results/formal_author/test_300/index.html.

• If you want to test on your own data, please first align your pictures and prepare your data's facial landmarks and background masks according to tutorial in preprocessing steps, then run

python test.py --dataroot {path_to_aligned_photos} --name formal_author --model test --dataset_mode single --norm batch --use_local --which_epoch 300

• We also provide an online demo at https://face.lol (optimized, using 120 pairs for training), which will be easier to use if you want to test more photos.

Train

• Download our APDrawing dataset and copy content to dataset folder
• Download models of pre-traning and auxiliary networks (for fast distance transform and line detection), from https://cg.cs.tsinghua.edu.cn/people/~Yongjin/APDrawingGAN-Models2.zip (Model2).
• Run python -m visdom.server
• Train a model (with pre-training as initialization): first copy "pre-training" models into checkpoints dir of current experiment(checkpoints/[name], e.g. checkpoints/formal), and copy "auxiliary" models into checkpoints/auxiliary.

python train.py --dataroot dataset/data --name formal --continue_train --use_local --discriminator_local --niter 300 --niter_decay 0 --save_epoch_freq 25

• Train a model (without initialization): first copy models of auxiliary networks into checkpoints/auxiliary.

python train.py --dataroot dataset/data --name formal_noinit --use_local --discriminator_local --niter 300 --niter_decay 0 --save_epoch_freq 25

• To view training results and loss plots, click the URL http://localhost:8097. To see more intermediate results, check out ./checkpoints/formal/web/index.html

Test

• Test the model on test set:

python test.py --dataroot dataset/data --name formal --use_local --which_epoch 250

The test results will be saved to a html file here: ./results/formal/test_250/index.html.

• Test the model on images without paired ground truth (please use --model test, --dataset_mode single and --norm batch):

python test.py --dataroot dataset/data/test_single --name formal --model test --dataset_mode single --norm batch --use_local --which_epoch 250

You can find these scripts at scripts directory.

Preprocessing Steps

Preprocessing steps for your own data (either for testing or training).

Training/Test Tips

Best practice for training and testing your models.

You can contact email [email protected] for any questions.

Acknowledgments

Our code is inspired by pytorch-CycleGAN-and-pix2pix.