Awesome

Deep Plastic Surgery

This is a pytorch implementation of the paper.

Shuai Yang, Zhangyang Wang, Jiaying Liu and Zongming Guo. Deep Plastic Surgery: Robust and Controllable Image Editing with Human-Drawn Sketches, accepted by European Conference on Computer Vision (ECCV), 2020.

[Project] | [Paper] | [Human-Drawn Facial Sketches]

Please consider citing our paper if you find the software useful for your work.

Usage:

Prerequisites

• Python 2.7
• Pytorch 1.2.0
• matplotlib
• scipy
• Pillow
• torchsample
• opencv

Install

• Clone this repo:

git clone https://github.com/TAMU-VITA/DeepPS.git
cd DeepPS/src

Testing Example

• Download pre-trained models from [Google Drive] | [Baidu Cloud](code:oieu) to ../save/
• Sketch-to-photo translation
  • setting <i>l</i> to 1 to use refinment level 1.0
  • setting <i>l</i> to -1 (default) means testing with multiple levels in [0,1] with step of l_step (default l_step = 0.25)
  • Results can be found in ../output/

python test.py --l 1.0

python test.py

• Face editing with refinment level 0.0, 0.25, 0.5, 0.75 and 1.0
  • model_task to specify task. SYN for synthesis and EDT for editing
  • specify the task, input image filename, model filename for F and G, respectively
  • Results can be found in ../output/

python test.py --model_task EDT --input_name ../data/EDT/4.png \
--load_F_name ../save/ECCV-EDT-celebaHQ-F256.ckpt --model_name ECCV-EDT-celebaHQ

• Use --help to view more testing options

python test.py --help

Training Examples

• Download pre-trained model F from [Google Drive] | [Baidu Cloud](code:oieu) to ../save/
• Prepare your data in ../data/dataset/train/ in form of (I,S):
  • Please refer to pix2pix for more details

Training on image synthesis task

• Train G with default parameters on 256*256 images
  • Progressively train G64, G128 and G256 on 64*64, 128*128 and 256*256 images like pix2pixHD.
    • step1: for each resolution, G is first trained with a fixed <i>l</i> = 1 to learn the greatest refinement level for 30 epoches (--epoch_pre)
    • step2: we then use <i>l</i> ∈ {i/K}, i=0,...,K where K = 20 (i.e. --max_dilate 21) for 200 epoches (--epoch)

python train.py --save_model_name PSGAN-SYN

Saved model can be found at ../save/

• Train G with default parameters on 64*64 images
  • Prepare your dataset in ../data/dataset64/train/ (for example, provided by ContextualGAN)
  • Prepare your network F pretrained on 64*64 images and save it as ../save/ECCV-SYN-celeba-F64.ckpt
  • max_level = 1 to indicate only training on level 1 (level 1, 2, 3 --> image resolution 64*64, 128*128, 256*256)
  • use_F_level = 1 to indicate network F is used on level 1
  • Specify the max dilation diameter, training level, F model image size
  • AtoB means images are prepared in form of (S,I)

python train.py --train_path ../data/dataset64/ \
--max_dilate 9 --max_level 1 --use_F_level 1 \
--load_F_name ../save/ECCV-SYN-celeba-F64.ckpt --img_size 64 \
--save_model_name PSGAN-SYN-64 --AtoB

Training on image editing task

• Train G with default parameters on 256*256 images
  • Progressively train G64, G128 and G256 on 64*64, 128*128 and 256*256 images like pix2pixHD.
    • step1: for each resolution, G is first trained with a fixed <i>l</i> = 1 to learn the greatest refinement level for 30 epoches (--epoch_pre)
    • step2: we then use <i>l</i> ∈ {i/K}, i=0,...,K where K = 20 (i.e. --max_dilate 21) for 200 epoches (--epoch)

python train.py --model_task EDT \
--load_F_name ../save/ECCV-EDT-celebaHQ-F256.ckpt --save_model_name PSGAN-EDT

Saved model can be found at ../save/

• Use --help to view more training options

python train.py --help

Contact

Shuai Yang

williamyang@pku.edu.cn