Eighteen ways to make a background, written down so you can make them again.
I took apart 117 free background images, one at a time, and sorted them by what you actually choose when you prompt. Each group became a recipe. Every image here is new, made from that recipe in the palette I would use: near-black, bone and one orange.
Arrow keys change the image, [ and ] change the style.
Grouped by what the image is made of, not by what it was sold as. Press one to open it above.
Two images belong to the same style when all five match. Everything else is decoration.
The study set was loud: neon purples, candy pinks, rainbow glass. I do not work like that. Same eighteen recipes, one rule: near-black, bone, warm stone, and a single vermilion that stays at about a tenth of the frame or less.
Grouping by the five variables ignores the shop's own collection names, and that changes the picture.
| Collections | 31 named collections became 18 styles. |
|---|---|
| Largest merge | Six collections turned out to be one style: glossy ribbons on black with the dye changed. |
| Largest split | One collection scattered across four styles. It was organised by file format, not by look. |
| Darkness | 58 of 117 images sit below 0.35 mean luminance. |
| Captions | Several were wrong. One shipwreck was a saxophone. |
Each style ships as a skill file: the rules, a formula with slots, a negative prompt, settings for three generators, and a short QA list that can fail.
| Resolution | I studied 1200 px previews. That is enough for composition and colour, not for judging fine grain. |
|---|---|
| Thin groups | Two of the eighteen rest on only two or three source images. I would not defend those boundaries hard. |
| My eye | The clustering is a person looking, not an algorithm. Someone else might land on fifteen groups, or twenty-two. |
| Four is a sample | Four images per style show the recipe works. They do not prove it works every time. |