Sign In

Solarum Lattice-sight XL — psychedelic ink over magenta & gold, with a crumpled-paper switch (SDXL)

100

Not recently loaded

Download

1 variant available

fp16 SafeTensor

jsolarum_latticesight_xl_v3.safetensors

Half precision, best balance • 162.64 MB

Verified:

Type
LoRA
Stats

20

40

10

Reviews

No reviews yet

Published

Nov 5, 2026

Base Model
Training
Steps: 1,160Epochs: 4
Usage Tips
Strength: 0.9
Hash
AutoV216B84684B8
Trigger Words
jlatsyt
V1_s2026110602.png

In my world, Solarum, people in the Undercity take a spore walk: a trip into the Lattice, the living network under everything. When they come back, they draw what they saw. This LoRA is those drawings.

What it does: turns any subject into a black ink drawing over big swirling fields of magenta, acid yellow, plum and gold, with a few small coloured shards. The subject stays readable: people, animals, buildings, streets, planets.

The paper switch: add , on crumpled paper to your prompt and it comes out on creased, handled paper, like a walker's sketch. Leave it out for a smooth poster look.

How to use it

- Weight 0.9. Start from: jlatsyt, a lighthouse on a rocky shore at night, on crumpled paper

- Then swap the subject. Simple subjects read best.

What it can't do (yet)

- The trigger alone doesn't give a reliable picture; always add a subject.

- Some subjects barely change at low weight; go to 0.9–1.0.

- It doesn't do the rainbow-mosaic look of my hand-made seed pictures.

How it was made

I started from 40 pictures I made on Civitai with a stack of my own LoRAs. One LoRA in that stack I don't use for training, so those 40 became the target, not the training set. Instead: each subject was drawn first with my own Acrylic on Paper LoRA, coloured in a tint step written in code (magenta / yellow / plum / gold swirls, ink kept, small shards), then painted over with my own clean LoRAs and an edge guide so the subject held. The paper creases come from my own first-pass pictures. 33 subjects, each flat and on paper (66 pictures), captioned only with the subject.

Training: Juggernaut XL v9 base, kohya sd-scripts, LoRA dim 32 / alpha 16, UNet only, 1024 px buckets, AdamW8bit, cosine, LR 1e-4, 10 repeats, released epoch 4 of 6. No other people's art, no Flux / Google / OpenAI material.

Made with Claude (Anthropic): Claude Opus 5.5 in Cowork planned the method and the tint code; Claude Code with Claude Sonnet 5.5 on my desktop ran the renders and training.