One of my noise paintings started as someone's wedding portrait. Randos from Facebook. The starting piece was a real photo, but it had this really weird uncanny quality to it that moved me.
Imagine a couple from the deep South on their wedding day, standing on a very normal front porch. They're both in their early 40s, and you can tell from their expressions that something is very wrong.
So, I took the photo into Krita, and cut it into three sections, top and bottom along the middle, cutting a sparse middle section that was only a couple of pixels tall.
After that, I collaged in a line drawing I made and blended it in with the Overlay mix mode on a new layer, so that it was transparent.
I do that because it gives the image texture that can be moved around or bent later.
And that gave me my noise painting or "blotter" image.
It's been my experience that a good blotter contains lines and artifacts that go outside of the distribution you would consider "normal."
Then I ran the blotter through Stable Diffusion.
Sometimes, I'll use ComfyUI when I want to use a more complex technical process. But this time, I used AUTOMATIC1111 because it's simpler, only needed two layers of conditioning, and most of the technical work for this one was done in Krita.
There, I ran it through at a low blur setting, with high semantic similarity (CFG 16) which constrains what the machine can do.
And I made sure to use a verbose bacterial butterfly prompt for the conditioning paintover using RunDiffusion's Juggernaut X model. I like how well X responds to the weird contours of my blotters.
The idea when you do it this way is to stack the different pieces of the machine against each other.
You want to break the distance calculation (the steps needed to get from where you are to the described thing), the classifier (how the machine understands what you give it), and the denoiser (the thing that turns the output from static to final image).
It's sort of like telling the machine, "Don't believe what your eyes and instructions are telling you!"
Yes yes.
You can also think of it like setting up a little robot fight.
The less freedom you give it to wriggle around and give you a straightforward interpretation of the prompt, the more it has to sort of paint inside the lines, if that makes sense.
You sort of develop a feel for how the machine is going to react to these things, the more you do.
And that gave me the preliminary render, which I curated from a batch of seventeen candidates.
I picked the render because it preserved my photo warping, aged the groom up about 20 years, and ate the bride completely.
There's a poetry in that, I think.
Anyway, the reason people have mostly stopped using Automatic is because the stock version has limitations that make the outputs look unfinished.
That's fine.
If it was too finished at this point in the process, the outputs wouldn't be interesting.
So what I did from there was pull it into GPT-Image 2.5, and told it to run minor corrections and an enhancement pass.
Then I told it to use the sketchy style from an earlier piece on the redraw.
And that's how I made the finished piece.