working with Genetic Algorithm is difficult.

[Sam]

I was looking at these sites that draw with Genetic Algorithm, for example, this works with the Procedural-painting

Or this one with OpenCV
and Voronoi  diagrams Genetic-Art

and portrait-drawing-with-genetic-algorithm

and Procedural Paintings with Genetic Evolution Algorithm

Then I looked carefully and saw how difficult it is for someone like me to work with them and how time consuming they are (and low quality).
That means we have to really thank the SA programmers for developing such algorithms for the program so fast and easy that sometimes you can really create beautiful works in the shortest time In addition, how much each user can implement their creativity in the work.

[Synthetik]

Ok, let's look at that first one for a minute.

Now any image can be represented by a combination of basis functions that add together to generate the image.

Image compression algorithms take advantage of this by only sending some of the basis functions, so the reconstructed image is missing some information.  And they tweak things so the missing info is info you have a hard time seeing anyway.

So what the algorithm is doing is essentially plopping down basis functions until the image is reconstructed.

You can use fourier coefficients (2D sin waves), or you can use wavelets (localized basis functions), or you can use blobs of paint (think of it like a wavelet basis function). So you add them all together and you eventually get the original image again.

So that's fascinating. I was always astounded that you could add together a bunch of 2D sin waves and get an image that looked like something.  And then once you get into wavelet theory you realize that you can plop down just about anything and if you do it enough times at different sizes you get the original image back.

So yeah, a genetic algorithm that plops down just about anything on a canvas working off of a mean square error (MSE difference between what is on the canvas and the original image) will eventually  reproduce the original image (most likely blurry since MSE doesn't really care about edges, just average coverage reproduction).

And i guess that is a valid approach to painting.  Keep plopping down stuff on the canvas until you reproduce the MSE for the original image.

But this kind of approach doesn't really allow for any abstraction of the representation. And at least to me, the most interesting art is the art that toys with visual abstraction in a visually interesting and compelling way.

And in order to do that, you need a system that handles visual representation in some way that mirrors how people perceive images.  Which is what studio artist tries to do internally with it's visual attribute modeling.