Rolling Watercolor Stylization

We stylize your picture into a watercolor-paint-like style!

How to Download this?

1. Download the whole project by "Code > Download ZIP" or git clone it
2. Make sure you have Python
3. Make sure you pip install'd below packages:

   pip install numpy 
   pip install dearpygui
   pip install opencv-contrib-python
   
   // or
   
   pip install -r requirements.txt
   

   • If opencv-python is installed, please uninstall it first. Only one opencv version can be in python.

     pip uninstall opencv-python 
     

     opencv-contrib-python provides all of the opencv-python functions with extra modules.
4. Run the program:

   python main.py
   

Tips

1. Apply Texture section paints your result on a paper texture

   Press select image button to select your paper image

   • The lower resolution paper image, the coarser the paper looked like in final result.
   • You can try doing Texture Synthesis to increase image resolution without scaling side effect. Here is a good implement: https://github.com/rohitrango/Image-Quilting-for-Texture-Synthesis

   If you don't have paper image, you can press start without select image. We will generate a paper-like noise map for you.

2. Adjusting sigma_s and sigma_r

   sigma_s decides how blurry the image is. The bigger value, the blurrier color lumps.
   sigma_r decides where are the edges of color blocks. The bigger value, the more differnece between edges are needed to preserve the edge, i.e. the blurrier at edges.

How Stylization is Done?

Color block effect is done by applying Rolling Guidance filter to source image.

There are 2 approaches when applying Paper Texture:

1. Have Paper Image:
   Scaling result image by grayscale value of paper image, then do gamma correction to brighten the result back. gamma is $1.2$, fixed.
2. Don't have Paper Image:
   Generate a noise texture by sampling Standard Normal Distribution to a small area, scale it to result image size, then do same sampling and scaling multiple time but to a bigger and bigger area, finally add all areas in to one. This generated texture is directly added to result image and colors' value are clipped back to $[0,255]$.