Solarum Ring Habitat XL draws the inside of a giant rotating ring habitat: a ribbed glass roof overhead, white walls curving up on both sides, and farmland rising away from you along the curve of the ring.

This is an experimental structure LoRA. It is good at one thing: holding the shape of the ring. It is soft, and it adds only a little detail of its own.

How to use it

Keep this frame in your prompt and add one thing at a time:

jring_habitat, low view, looking along the ring, level terraced fields and straight water channels, bright daylight, landscape photograph

Weight 0.9. With the frame kept, the roof, walls and rising land held in every one of my tests. Adding things that worked:

- dense food forest of fruit trees and vines filling the terraces (tree canopy on the terraces)

- autumn orchards, orange and red leaves (dappled autumn colour)

- warm evening light, long shadows

- a river winding along the ring floor

- rolling hills, rivers and ponds in place of the terraces, for an older, grown-in ring

What it can't do (yet)

- Leave out the frame and write only about the content (a town, a forest), and the content wins: the ring disappears.

- The trigger on its own gives soft fog and a lone tree, not the ring.

- Houses come out small and few.

- Everything is soft. A second pass without the LoRA makes it sharp, but the scale collapses: the roof starts to look like a garden pergola over a lawn.

How it was made

I built a 3D model of the ring (about 10 km across, a valley about 1.3 km wide, a roof about 600 m up), laid aerial ground over it, rendered 32 views, and painted each view in six overlapping sections with Juggernaut XL v9, then stitched them back. 62 of those pictures, captioned only with what changes between them (height, flat or hilly floor, foreground, light), are the training set. No other models' outputs, no photos of real places.

Training: Juggernaut XL v9 base, kohya sd-scripts, LoRA dim 32 / alpha 16, UNet only, 768 px buckets, AdamW8bit, cosine, LR 1e-4, 10 repeats, released epoch 4 (930 steps over 6 epochs).

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

The full story of how it was made, including what failed: https://civarchive.com/articles/37456/teaching-sdxl-a-shape-it-has-never-seen-a-ring-habitat-lora-built-from-a-3d-model

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