AI ART: Quick look at how the image AI works (Patreon)

A bit of insight on how the image AI that I've been using, Stable Diffusion, works for those curious to find out more about it.

CONTENT

It doesn't just piece together elements of other existing image, it makes new completely original images based on what it learned about things going well together. Those images are then yours to do as you please with them!

PROCESS

It might seem as it's gonna be super complicated but it's actually crazy simple: the AI dreams. It's similar to how the human brain puts together a story that sort of makes sense to you when different neurons fire up when you sleep. The AI takes some random pixels and it tries to makes sense of them based on what it was trained on. If it was trained on cat photos, it will see cats everywhere.

Without any input, the AI starts with a random noise image like this

and it will keep trying find things that make sense to it. It will keep building and building for as long as you let it.


PROMPTS

The way people mostly use this AI is by prompts which are a bunch of words, sort of like tags. When you throw words at the AI, it will try to look for those things harder as it builds the image.

For example I gave it this text:

"25 years old woman blue eyes, blonde hair, dressed in santa costume, christmas, sexy nsfw tease sheer, white fur, red latex, (((santa hat))), posing,present, stock,  choker, cleavage, ((((makeup)))), eyeliner, eyeshadow, (blue background)"

And let it go wild for 5-10 minutes. The parenthesis are to emphasize those more, the more () you put the more the AI should focus on those terms. The post images are what the AI came up with.

It doesn't follow instructions exactly, that's where VAM will come in for posing, clothing, stuff like that. But it did a beautiful job nonetheless!

INFINITE DIVERSITY

At first glance the photos might seem like some stock photos with limited variation but everything is built on the spot by the AI and is original. It just uses what it knows about colors going well together from it learned.

I really liked the cover image but wanted to see some different options in the same style so I just clicked a button "make more like this", waited a couple of minutes and the AI did the following:

As simple as that! If you try to spot the differences you'll notice that everything is actually different.

ARTEFACTS

Similar to how dreams are kinda weird sometimes but sort of fit in together, sometimes the AI does things like that and there are weird things in the images. For example the left hand in the image above has six fingers. There are workarounds though to try to fix or avoid things like that. More on that later!