SC-FEGAN

SC-FEGAN : Face Editing Generative Adversarial Network with User's Sketch and Color

Youngjoo Jo, Jongyoul Park

arXiv: https://arxiv.org/abs/1902.06838

Overview

We learn to edit face image with a deep network. Our network SC-FEGAN is well suited to generate high quality synthetic image using intuitive user inputs with sketch and color. We used SN-patchGAN discriminator and Unet-like generator with gated convolutional layers.

Dependencies

• tensorflow
• numpy
• Python3
• PyQt5
• opencv-python
• pyyaml

Usage

First, download the model from Google drive.

Basic usage is:

mv /${HOME}/SC-FEGAN.ckpt.* /${HOME}/ckpt/
python3 demo.py

Select the number of GPU by editing demo.yaml file (not support multi-GPUs).

GPU_NUM: 1 (the number you want to use)
#GPU_NUM: (if you want to use only CPU, erase the number)

You can use our network with simple GUI. The only erased regions are filled.

The buttons in GUI:

• Open Image: Open the image you want to edit and reset sketch and color.
• Mask: Click this button and draw the mask on the left viewer.
• Sketches: Click this button and draw the line for sketch on the left viewer.
• Color: Click this button and draw the line for color. If you click this first time, you have to choose the color from palette.
• Palette: Click this button to change color. If you choose the color, then click the Color button to change.
• Save Img: Click this button to save results. It is saved as 'name.jpg'.
• Arrange: Arrange the editing works.
• Undo: Undo previous editing work.
• Complete: Complete the image and show it on the right.

We recommend using this in the following steps:

1. Draw the sketch plausibly referring to the original image.
2. Draw the mask on the sketched region.
3. Click the `Arrange` button.
4. Draw the color on the masked region.
5. Click `Complete'.

Example Results

Face editing

Edit earring

Face restoration

Face restoration (with only sketch and color)

License

CC 4.0 Attribution-NonCommercial International

The software is for educational and academic research purpose only.

Notes

• This is developed on Linux machine running Ubuntu 18.04.1
• Provided model and sample code is under a non-commercial creative commons license.

Citing

@article{jo2019sc,
  title={SC-FEGAN: Face Editing Generative Adversarial Network with User's Sketch and Color},
  author={Jo, Youngjoo and Park, Jongyoul},
  journal={arXiv preprint arXiv:1902.06838},
  year={2019}
}

Next

• Update training code

Acknowledgement

We acknowledge the official code DeepFillv1. We thank all the researchers at ETRI, especially Kimin Yun and Jinwoo Jung, for insightful discussions.