You could be an artist with AI
This project can transform your casual sketch to beautiful painting/artwork using modern AI technology.
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
The principle behind this is called Conditional Adversarial Networks, known as pix2pix, which is able to generate image based on the given image.
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.
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.pyfor pix2pix translationapp_stylize.pyfor arbitrary style transfer
Make sure you have Python installed. And the following key packages are needed:
tensorflow
keras
pillow
flask
gevent
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.
You should installed:
# Clone the repo
$ git clone [email protected]: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.
This is a project of Digital Image Processing taught by Dr. Jia Yan.
Shuaibin Zhang, Haoran Su, Tangbo Liu, Xin Fu
Copyright © 2018, Fing
Released under the MIT License.