SDXL

Just running the original dataset through SDXL lora training has produced something a bit jank and hard to use. I recommend starting with a weight of 0.5 and experimenting from there. Using diagram or diorama in the negative prompt can be useful.

SD 1.5

In Theory: The idea of this modal was to produce side-on cross-sections of landscapes and terrain - think D&D maps of cave systems, the way Terraria's 2D perspective works, or that info movie in Jurassic Park where the animated miners are digging for amber. I have found most models do not generate this kind of perspective easily or consistently.

In practice, however, it mostly succeeds in getting the cross-sectional perspective right and then produces strange subterranean features which may or may not be of interest. For improved results I recommend experimenting with ControlNet and/or Regional Prompter to better control composition. It also has a tendency to go overboard with roots and lattice shapes (stairs, shelves, ladders).

If you are having difficulties, make sure you aren't copying the lora name from my examples as the naming will be different.

Prompts: csu, underground, cross-section, from side

Weight: 0.3 - 1.

Weights in the 0.3 - 0.7 range seem to give some interesting depth to a composition, while the closer you get to 1 the flatter and more diagrammatic it becomes. Values over 1 can produce interesting results but by around 1.5 image quality is deteriorating.