In my world, Solarum, people take a spore walk into the Lattice, the living network under everything, and draw what they saw when they come back. A grey walk is the cheapest and shallowest: sounds, colours, echoes. Things surface for a second and dissolve. A grey walk gives you a sketch; a gold walk gives you an icon. This LoRA is the grey sketch.
It's built from my own drawings. I do automatic drawing: pencil, a set pressure, letting the hand and the paper lead. The base of this LoRA is seven of those drawings, cropped and turned (automatic drawing has no "up"), and every subject picture in the set was grown out of one of them.
What it does: turns any subject into a pencil-and-ink drawing on paper with a pale mint-green wash and small orange touches. It works on things it never saw in training: an elephant, a cellist, a kitchen at night, a boat under a planet.
How to use it
- Weight 0.9. Start from: jlatgry, a small boat under a huge ringed planet
- Just name the subject. No style words needed.
- Negative: photograph, realistic, 3d, vivid colors, saturated, text, letters.
What it can't do (yet)
- It comes out cleaner than my originals: more ink and engraving than loose scribble. Animals, faces and the cat are the loosest.
- At 0.7 some subjects stay half-photographic. Use 0.9.
- Lanterns and lights keep some of their own colour.
How it was made
v1 learned only "abstract" (every caption said abstract, so that's all it learned). v2 added subject pictures grown out of my drawings with img2img, and got too clean. v3 added 60 more subjects (rooms, people, animals, objects, places) at a gentler setting, picked by hand. 215 pictures in all: my drawings plus 104 subject pictures, captioned with the subject only. No other people's art, nothing from Flux / Google / OpenAI.
Training: Juggernaut XL v9 base, kohya sd-scripts, LoRA dim 32 / alpha 16, UNet only, 1024 px buckets, AdamW8bit, cosine, LR 1e-4, released epoch 4 of 6 (1,584 steps).
The family: Solarum Lattice-sight XL (ink over magenta and gold): https://civitai.red/models/3038144 · Solarum Lattice-sight Gold XL (gold line art on black): https://civitai.red/models/3039807
Made with Claude (Anthropic): Claude Opus 5.5 in Cowork cropped my drawings and planned the method; Claude Code with Claude Sonnet 5.5 on my desktop ran the renders and training.
More: https://signalofsolarum.substack.com · https://www.deviantart.com/signalofsolarum
Full story: Building Solarum, part 7, https://civitai.red/articles/38640/i-trained-a-lora-on-my-own-automatic-drawings-it-learned-abstract-and-nothing-else-at-first