AI ART: turning VAM images to digital art with AI (Patreon)

I'll post a step by step guide on what software you can use, from where you can get and how you can use it in the next days. Before that I wanted to cover a bit some of the logic behind it and what you should know so that it makes more sense when you try it.

If you haven't checked already my previous post about this, it's here and it covers the basics on how the AI decided to build the images.

Long story short, the AI sort of "dreams" the images based on what it has available:
- words that you give it
- an image to guide it
- or just randomness 

MODELS

The simplest way to understand AI models is to think of them as vam addons. Stable  Diffusion comes with its own model that's quite amazing for general purposes. But there are other models you can download that other people make, some go in nsfw directions, some go towards anime, some towards art.

A few people do models by training them, giving the AI images to learn from. Then they share the models either as free or on patreon or both, same as with the addons here.

They're one file, take about 2-8gb each, you hoard them like vam addons in a folder and you can switch them from a dropdown when doing the images.

MODELS IN ACTION

In this example I made this quick & ugly image in vam. I added the blue cube to demonstrate how the AI incorporates and tries to make sense of it

First I didn't tell it anything, just gave it the image and hit go on the standard model:

Better looking than some of my vam models but no prize winner. So I gave it also some text to help it:
"35 years old woman, mature, brown hair on one side, makeup, red cape, black shiny roman armor, stockings, red pouches, posing, standing,photorealistic, forest background, blue cube" and it came up with these:

and

The standard model (the default AI that builds the images) is trained on just about any type of image you can imagine: photo, drawing, 3d, anime. So the AI will sometimes go for a photorealistic style, sometimes for cartoons. But you can control that with prompts.

You can also let it run for longer and it will fix more things, like the mouth in the first image. The less you allow it to work on it, the more glitches are gonna be in the image. For these I let it run 25 iterations which is kinda the recommended minimum & it takes seconds or up to a minute or two. For a good photo people usually let it run 50-100 or even more.

PROMPTS & NEGATIVE PROMPTS

The words you give the AI to help it, like "black shiny roman armor, stockings," are the prompt. There's also a negative prompt you can give it, which is things you don't want the AI to think about when making the image. You can type in for example: "anime, cartoon, drawing" and the AI will avoid using things like that and go just for realistic photos instead

If you like anime and want images like that you can do the opposite, add "photography" as a negative prompt and add anime as positive one. But for anime there's also an awesome model that's great with that, named Anything v3:

I tried it at first again with no prompt. Just the image to see what the AI does with it:

it thought of the blue cube as part of the person in this one particular case. I gave it another go with a prompt to help it (mature woman,shiny black armor,etc) and it made these:

and

It sometimes goes where you want, sometimes it's very stubborn. You can control how much the AI goes crazy with it with two sliders:
- one for how much the AI tries to respect your prompt words
- and one for how much it tries to respect the original image (if you gave it one)

That's all the settings there are:
- prompt & negative prompt
- an optional image (like the vam one i gave it)
- and the two sliders to control how Kanye the AI goes

I tested it also on a different model named HassanBlend (v1.4) that does photography the best at the moment from what I've found and got the cover pic

The setting was for the AI to go a bit crazy with it (I didn't mind one bit!) but I wanted it to keep it closer to the original so I turned it down a bit and got this:

Very close to the original image and it also followed the prompt finally in regard to the age. Which brings me to the downsides:

DOWNSIDE 1: AI nightmares too

The AI can make sort of any image and depending on what the model was trained on it might go in disturbing places and freak you out or maybe even get you in trouble.

Some Virtamate looks for example already look barely legal. Running some of these general AI models (that were not trained with erotic things mind) on them might make the characters look too young. To avoid any problems like that you have to be careful to what you're asking the AI to do and even force it to avoid things like that by mentioning prompts like "mature" or adding negative prompts like "young".

There's also freakish things it's gonna do quite often that might catch you off guard and disturb you. It doesn't just mix the images it learned, it makes its own images and it can go alien or horror quite easily.

DOWNSIDE 2: softcore only

There are lots of AI models out there, it's a bit of a boom happening right now but NSFW ones are not too great, not yet anyway and probably not soon, AI hates fingers & gens. There are some tricks though that I'll get into later.

DOWNSIDE 3: some models are not safe

There's a site like a hub where all the good models are (huggingface.co) and get curated. But there are also other models floating around the net that can be very dangerous, especially promising NSFW. I've seen that people are a bit in panic mode over that so you have to be extra careful to what you download and from where.