Another example of my noise painting pipeline.
The idea is that you start with a hand painted image or collage, but that you design it in such a way as that it messes up the classification and denoising systems in a diffusion model.
This one was mostly hand drawn in Krita, using different brushes and cloned/inverted 3 times on the same canvas using layers, and light use of the edge detection filter. For texture, I added a picture of a 1983 IBM all in one PC, and edge traced it for texture.
Taking it to Stable Diffusion, this piece was simple, so it only had two inputs and a default workflow in Automatic 1111. Automatic actually still makes sense when you don't need a complex workflow setup.
I used Juggernaut X for the model, because the way it's trained is fascinating.
The robot fight
Rendering:
First the prompt conditioning, and then the image conditioning.
The prompt itself wasn't complicated. Just enough to get the artistic neighborhood for the render. No legacy artists were directly named by name in this one. But the key to the whole process is the robot fight.
When processing inputs, the idea is to pit the classification system, language processing, and the denoiser against each other, so results are unpredictable between iterations. But in practice, this makes the system feel more fluid and flexible.
So most of the perplexity needs to happen in the input image itself.
After you've been doing it for a minute, you develop a sense for how the variables sit, and what the tolerances are going to be for any given piece, because the blotter/noise painted image has resistances all its own.
That's where the craft lives.
Settings and refinement
If memory serves, I melted this one using the DPM3 algorithm, or some variation of it. And since DPM3 is more resistant to suggestion than Euler Ancestral (the one I usually use for these), I turned both classification free guidance, and denoising up higher than usual. CFG at about 15, and Denoise/blur to around 56%.
This time, only four candidates, but two of them were the same, because I wanted to see what happened if I turned the blur from 56% to 61%. It was about the same, honestly. DPM is fickle.
The last step was semantic refinement using GPT-Image-2.5. And that consists of asking ChatGPT for targeted subtle edits and corrections to the piece.
And that my children, is how the slop is made.
I'm actually pretty happy with the way this one turned out.