The Stable Diffusion Rabbit Hole

I’m thoroughly terrible at documentation, so I didn’t write a single thing down. Whoot whoot, go me!

At this point, my husband has decided that we need to get Stable Diffusion set up at home, so we can make lots of pretty pictures without spending “credits” through host sites. Also because we’re the sort of nerds that enjoy making minute changes to see what effect it has, and that’s very difficult when one is working with limited numbers.

So, with this round of learning, I took the same prompt, “Yellow rose in a glass vase. Centered composition,” and changed the art style. I do remember that one of them I put down “Lisa Frank,” but I bet that you can’t guess which one. Another one was, “Magic realism.”

Also, since I used the Stable Diffusion Demo for these, I had no control over the seed or any other technical settings.

I need to get into the habit of documenting.



I joined Night Cafe for making AI art, because while each creation costs “credits,” they offer a lot more options and control over what you can make.

Every day, Night Cafe holds a contest where they announce a theme for members to submit AI art, then people vote on the entries — the one that averages the highest rating wins.

A few days ago, the theme was “Love.” After spending some time thinking about it, I decided what I wanted to do, and after a few tries, managed to create this image:

It was something of a trick, and I repeatedly specified that I wanted a “normal baby,” — with negative prompts for things like, “scary,” and “creepy.” I discovered earlier in my playing around with Stable Diffusion, that the AI has a hard time understanding what constitutes a “cute baby.”

As much as I would have liked to have a dad present in a loving family portrait, the addition was beyond my current skill level. But, as a picture of a mother and baby, I’m really pleased with what I was able to create.

Motherhood has been getting dumped on for the last 50 years, to the point where women have internalized the misogynistic messages. You aren’t “just a mom.” You aren’t “wasting your life.” You ARE contributing to society — by raising and shaping the next generation. Please, stop listening to those toxic messages and learn to love yourself and what you are doing. Motherhood really is beautiful and important.

I wasn’t the only one who submitted a picture of a mother with a baby for this contest — a huge number of other people did, too. The winning picture was of an old man and woman, illustrating that Love is spending your life with someone.

Funny how it’s artificial intelligence that is revealing what people actually feel in their hearts.


The Evolution of Carol

AI art is harder to make than you would think.

You go to Stable Diffusion for the first time, type in a random prompt like, “A horse galloping through the woods,” and get back a picture that actually looks pretty cool.

So you play around a lot more, until your confidence is bolstered enough that you start to think that you can create something specific — to approximate a solid idea that you have in mind.

And that’s where the trouble starts.

I decided that I wanted to create a character portrait of Carol, from The Scion Suit/The Scions.

My first attempt, anime style:

… Okay, I’ll give it a few more tries …

By which point, I realized that simply describing the woman I wanted a picture of wasn’t going to be enough. AI doesn’t have the same intrinsic grasp of human anatomy and faces that we do.

I joined a Discord server and did a lot of reading.

A LOT of reading.

And I quadrupled the number of keywords I was using, to include things like, “Normal face,” and “Not disfigured.”

Highlights include:

Waaaaay better, but not even close to my end goal. Too glamorous and sunkissed for Carol.

Anatomy is still off. Also seems a little vampirish.

Hair is too short.

Now we’re starting to get it.

I liked the picture, but I decided to change art styles to see if I could better grasp the essence of Carol. And I like anime, so, you know … 😀

I’m still amazed that the face isn’t horribly wrong. But it’s not quite there — the woman is too doll-like to really be Carol.


All told, it probably would have been easier and faster to draw a character portrait if I had the art skillz — but I don’t. I generated dozens of pictures, most of which I didn’t save, and probably spent more time on it than it would have taken to paint it in photoshop.

But, truth be told, I enjoyed the process quite a bit. I think very strongly in words, and my brain turned it into a puzzle game of figuring out which words in which order would produce the best results. My final change was to swap out the keyword, “Shy,” for “Demure,” which gave the right tilt to the eyebrows and mouth for her expression.

And thus, I give you, Carol.