Calm Pieces - Convolutional Art

My program, Convolutional Art, has a feature that not only saves the finished product, but also the "genetic code" for recreating the piece. However, this feature is not particularly good for some reasons I understand, and some I do not. While experimenting with something that did not end up working, I tried to recreate this piece:

Blue Marble

This piece is kind of unique. Usually, pieces end up being either very chaotic or completely flat and boring. This one teeters a fine balance of big contrasts, macroscopic structures, and an interesting texture. All of them work together quite well. So of course, I wanted to see if I could recreate it.

Smokestacks

Well, it's not quite the same, is it? Of the six I generated, this is the one that came the closest to the original. It has a texture that feels almost the same, and it also plays with lights and darks. At a distance, perhaps, this is even better than Blue Marble, since this one stands out more with the picture split in blacks and lights.

Anatomy of a Nightmare

This one seems similar until you notice the vastly different texture, and the way that the larger scale structures are very dissimilar to the original too. Of all these calm pieces, this is the one which is least geometric, like there actually is some sense behind its shapes.

Hole in the Dreams

This one is quite vastly different. The long brush-strokes give the piece a sense of depth, a foreground and a background with a hole in the middle. The choice of colours is very tender - and then, suddenly, out of nowhere, a black fragment rests there in the middle of the lavender white.

The last two pieces were not particularly good, so here they are, side by side. My problem with them is that it becomes quite clear what is the most important part of all these pieces. The program runs from ten input channels which it then combines in interesting ways - or that's the idea, anyway. Sometimes the input channels themselves are the ones that shine through. These above pieces bear a striking resemblance to the contour map which makes up two of the ten inputs. It is obvious that these also play a role in the original marble pieces, but there, at least, it is obscured.