imfing / sketch-to-art

🖼 Create artwork from your casual sketch with GAN and style transfer

Sketch to Art 🎨

You could be an artist with AI

[Live Demo] Note: Only frontend is available now

Or you can Run with Docker in minutes

• Introduction
• Run with Docker
• Theories
  • Sketch Reconstruction
  • Style Transfer
• Manual Installation
  • Backend
  • Frontend
• Acknowledgments
• Authors
• License

Introduction

This project can transform your casual sketch to beautiful painting/artwork using modern AI technology.

Screenshots

Run with Docker

With Docker, you can quickly build and run the entire application in minutes 🐳

# 1. First, clone the repo
git clone https://github.com/mtobeiyf/sketch-to-art.git
cd sketch-to-art

# 2. Build Docker image
docker build -t sketch-to-art:dev .

# 3. Run!
docker run -it --rm -p 8080:8080 -p 5001:5001 -p 5002:5002 sketch-to-art:dev

Then, go to localhost:8080 and play with the demo! 🎉

Theories

To achieve the goal, there are mainly two steps in the pipeline:

• Reconstruct and generate real image from the sketch
• Arbitary style transfer to beautify the result with given result

Sketch Reconstruction

The principle behind this is called Conditional Adversarial Networks, known as pix2pix, which is able to generate image based on the given image.

Style Transfer

It became known to us with the appearance of Prisma app. Typically, we generate an individual model for each pre-defined style. Here, we want to go further by using any new picture as the style. So, we adopted the method, Universal Style Transfer via Feature Transforms proposed in NIPS2017, which enables us to perform arbitary style transfer.

Manual Installation

Backend

The server side is powered by Python and Flask. You can see this simpler example. Navigate to the server directory and all the files concerning the service and neural networks are there. The two main files:

• app_pix.py for pix2pix translation
• app_stylize.py for arbitrary style transfer

Prerequisites

Make sure you have Python installed, and some packages are needed: tensorflow, keras, pillow, flask, gevent. You can use pip to install them:

pip install -r server/requirements.txt

Run

# Simply run with python
python app_xxx.py

And you could see the output indicating the port it's listening (5001 and 5002). Go to http://localhost:5001 and you should see the returned information.

Frontend

You should installed:

• Node.js
• Yarn

# Clone the repo
git clone [e-mail]:mtobeiyf/sketch-to-art.git
cd sketch-to-art

# Install dependencies
yarn  # or npm install

# Run
yarn dev  # or npm run dev

Open your favorite browser at http://localhost:8080, the site is there.

Acknowledgments

This is the final project of Digital Image Processing instructed by Prof. Jia Yan.

• WCT-TF
• pix2pix-tensorflow
• pytorch-CycleGAN-and-pix2pix
• Neural-Style-Transfer-Papers
• Vue.js

Authors

Xin Fu, Shuaibin Zhang, Tangbo Liu, Haoran Su

License

Copyright © 2018, Fing

Released under the MIT License.