This is the third of three (four?) posts detailing the process of programming a convolutional art generator. For the results, take a look at these Newer pieces or Early pieces.
Where we left off, the generator had finally become sophisticated enough to create paintings that do not just look like perlin-noise, first of all by utilising brush-like polygons for actually applying the colours, and secondly, by feeding the outputs of previous neural layers into new perlin-maps. This has gotten us to things like this:
Now that I had reached a result with which I could be happy, the goal became to readd all of the variance, that had once been my bane. I think it says something about the development process when it comes to random content - when things aren’t working out, you try to reclaim control, but for final ascendance, control has to be given back to the program. So I added an option for the pieces to use the original RGB-channel approach to colours.
Here we can see that each method has its own strengths and weaknesses. The pigment approach creates visually pleasing, harmonic colours, mimicking how actual paintings are made. The digital method, however, is more wild, and in that wildness, sometimes creates something great. It’s also a natural at rainbows.
There is one thing I am neglecting to mention. Curation. You can work a thousand years creating an intelligent AI (which I have not), but you’ll still have to throw out the bogus pieces, like entirely blank canvases, or things like this:
Which isn’t terrible, just directionless and bland. The colours are too flat - but it is way easier to have a human pick out the more aesthetically pleasing pictures than having the computer somehow learn what aesthetics are.
There is another option though. Choice over curation. So, yeah, I added back in the quartering of the image. You are presented with four images like so:
Each is a slight mutation on the same neural network, with one of 27 rules changed somehow, as well as a slight change of colour, and a change of one of the other variables used in the calculations (like the constants of the size of layered Perlin-maps). The strength of this approach is that the picture then chosen can have its neural network promoted to the new main network for four new pictures to be created. It is a form of evolution by artificial selection.
I cannot emphasise how important this is. In one part, the curation of better pieces. Already through four generations of four options, though the user only sees 16 images and makes four choices, in actuality, there were 256 potential outcomes, of which, hopefully, the most aesthetically pleasing is promoted.
That’s just one part. The other thing is that this generator now is an interactive experience. I love it for this - it is a tranquil, thoughtful experience. Also, I have seen several examples of people not just choosing pictures, but starting to think about what makes them choose - to consider their own sense of aesthetics.
In the rest of the post, I will take you through one single whirl of the machine, completely uncurated. Perhaps it will go terribly, maybe it’ll be boring. Let’s see what happens.
The first selection is made to have higher variance than the future ones. No consistent theme has appeared yet. I’ll choose the bottom right one, because I like its texture.
Hm, it really took to heart what I chose. Three of the four mutations are quite similar to the original. It seems to be quite a stable configuration. I think I’ll choose the top right one this time, mostly because of the dark splotch just left-of-center, which creates some sense of depth.
It is suggesting a more peach-flushed colour, or a version of the one I chose that is just the slightest bit calmer. I think I’ll take the bottom-left.
Sometimes, people feel that they are building towards something and steer away from drastic changes. I personally feel no allegiance to the rust-ish colour scheme of the last three picks. Notice how the composition of the bottom-left one can also be glimpsed at in the previous choice, oh, and in the very first options in the top-right. Anyway, that’s the one I’ll pick. So after the four quartered images, the last choice will be rendered in full 1920x1080 and then saved onto the computer.
Looking at it in full size, I kind of regret not picking one of the more chaotic pieces, but I only have myself to blame for that. That’s a weakness of this approach - pictures work differently depending on their size and context. But that's not just a problem for my program, but for all digital art.
Comments
Post a Comment