Python + CUDA "paint-looks-alike generative art" evolution algorithm
This is a Python + CUDA solution, that recreates given image in Paint-looks-alike style using the evolutionary algorithm approach
Example of work:
(here on pic: authors of the current script, extended color space)
143 figures,
time from the start to the end in this example: approx. 8 hours
The author of the original idea is Roger Alsing
How to run the script
run the main.py
Important: the size of the initial picture (height x width) should be multiple be 20, e.g. 180x240 (this limitation goes from the current cuda realization)
I strongly advise not using big pictures, the optimal size is around 200x200
This is theoretically possible to run this script without CUDA (the environment variable NUMBA_ENABLE_CUDASIM should be set to 1) however, the time of execution will increase significantly
Main realization details
There are 3 + 1 figures available (Rectangle, Ellipse, Triangle, and Line as a particular case of Rectangle). Rectangle and Ellipse can be rotated (the axis of rotation is the center of a figure)
Opacity is not supported as the main idea of the script was to generate the "paint-like" images
The possible mutations of one figure: slightly change the position of the figure, angle of rotation (if possible) and color of the figure (substitute with some similar color). Additionally, it is possible to convert one figure to another:
- Rectangle ↔ Triangle
- Rectangle ↔ Ellipse
- Ellipse → Triangle
The mutations in the scope of the whole picture include adding new figures, removing figures and shuffling the order of figures.
The examples of mutation can be found in tests/figure_test.py
The evolution itself includes mutation and simple crossovers of the best part of the population.
Evolution is sequential, e.g. we are trying to achieve as good as possible result with n figures and then allow to explore n + delta figures
The goal function is not the simple sum of pixel differences but more complex readmean distance, this distance is a more appropriate measurement from the point of the human perception of color
The possible color pallet is limited on purpose with around 60 "standard" paint colors
It is possible to add "hue" and expand the color space from 60 to around 2800 colors, just set the
variable IS_EXTENDED_COLOR_SPACE in Color.py to True
Acknowledgment
This project was done by Holovashchenko Vadym and Anna Nosach for the Evolutionary Algorithm course, University of Wrocław, 2021
(a campus of the Institute of Computer Science, University of Wrocław)