Solarum Script XL writes in an invented script instead of English. It's a real writing system I designed for my world: each block is one syllable, every syllable hangs from a short roof stroke with a curl at its left end, and the marks under the roof are built from the strokes my own hand repeats most in my automatic-writing pages. Vowel marks sit above the roof.

It puts that script on real surfaces, the way people would actually make it:

- carved into stone (slabs, standing stones, lintels)

- stenciled in paint on brass and metal (doors, plates, hulls), with stencil gaps and drips

- stitched in thread on cloth banners

- painted on wood (weakest for now, see below)

How to use it

Start from this prompt at weight 0.9:

jsolscript, two rows of glyph script carved into the stone, a large stone slab set into a dry-stone wall, photograph

Change one thing at a time:

- how it's made: carved into the stone, stenciled in off-white paint, stitched in orange thread

- the surface: a brass door, the grey hull of a small spaceship, a cloth banner, a standing stone

- how much: one row, two rows, three rows

Add latin letters, english text to your negative prompt.

What it can't do (yet)

- Painted wood (boat hulls, shop shutters, market awnings) shows only a few faint glyphs. More painted scenes are coming in a later version.

- On a door with a lintel, the glyphs often land on the door instead of the lintel.

- The trigger on its own gives a dark board with a single mark.

- It doesn't spell words you type; the glyphs are real blocks of the script, arranged freely.

How it was made (full story in my article: https://civarchive.com/articles/37745/why-sdxl-kept-turning-my-invented-alphabet-into-english-and-the-two-fixes-that-worked)

The script came first (a chart of consonants, vowels and rules). Version 2 of this LoRA failed: its glyphs looked too much like Latin letters and digits, so SDXL kept "correcting" them into English, and the training pictures had the glyphs pasted on top of the surface. So I redrew every glyph around the roof-and-curl, and wrote a compositor that seats the script into each surface (carved grooves lit by the photo's own light, paint flaking along the grain, real stencil bridges and overspray, running stitches) before a light polish. 141 of those pictures are the training set: my own surfaces rendered with Juggernaut XL v9 and my own script. No other models' outputs, no photos of real places, no people.

Training: Juggernaut XL v9 base, kohya sd-scripts, LoRA dim 32 / alpha 16, UNet only, 768 px buckets, AdamW8bit, cosine, LR 1e-4, 8 repeats, released epoch 4 (of 6; 1,692 steps total).

Showcase: Juggernaut XL v9, Euler a, Karras, 24 steps, CFG 4.5, LoRA 0.9, 1216x832, hires 1.5x.

Made with AI. Part of my visual research for Solarum: https://signalofsolarum.substack.com · https://www.deviantart.com/signalofsolarum