- Joined
- Jun 19, 2019
Not an expert, just someone who has seen a lot of procedural generated stuff and noticed certain trends.I have to admit, I'm curious about the whole neural network fursona generator thing because I'm not super familiar with how those programs work and I want to know how it goes about what it's doing.
I'm mostly curious because art isn't the same as feeding it photographs because it's working with a bunch of images using a different... language, I guess? With photos of people, there's all sorts of variation in color and feature and so on, but it's all photos; the general style of how forms exist in space is the same, the way light acts on the surfaces is more or less the same, etc etc.
With these fursonas, if the output is any indication then the source images are absolutely all over the place in terms of rendering. There's absolutely flat lineart stuff with minimal cel shading, there's very cartoony disney smooth-rendered 3d, there's very painterly stuff with a much more realistic look. Here's a screenshot for an example, you can see a bunch of different 'styles' and they're all pretty distinct.
View attachment 1276015
I could totally see a program spitting out an ungodly mishmash of these different styles because it's treating things like [smooth, pleasing rounded shapes of ears rendered in x style] and [outlined spiky anime hair with no shading in y style] as individual features, but what's wierd to me is that even when they're comically a little skewed, almost all of the generated images pick one style and are consistent in applying it to the design? And it's good at it too, especially for complex structures like more realistically rendered hair or fur where it's maintaining the light source and application.
Could anyone who understands the process of neural network stuff better than me (so, anyone) help explain how this works? From an art perspective and as a total idiot when it comes to tech I'd legit love to know how that works, do they divide the input images into categories, run the results, and then jam the end results together? I know people were talking about seeing repeats so is it preloaded images rather than on the spot generation, or did someone just troll the entire furry fandom by cropping and filtering a bunch of fap material?
You've more or less hit it on the head; it is just random traits that are "matched up" on a likelihood basis with things like style and color elements, with some blending/tweaking involved to make it look more viable. A lot of it also runs on probability algorithms, which is why the vast majority of the proceedurally-generated furries spat out by the system have got pointed cat/dog ears, for instance; because the algorithm gets way more hits on pointed ears than flopped or rounded. Likewise, tan furries often have brown or red hair; longer muzzle types generally don't have tiger markings (at least not as often), "softer" art styles typically favor pastel colortones. There are some probability aspects to it regarding styles as well but that's the basic explanation.