Welcome to Space Station 13...

This is an experimental PonyDiffusion v6 LoRA is designed to replicate the 32x32 pixel art style sprites of the timeless open-source 2dspessman simulator, Space Station 13.

The goal of this model is to allow you to create customizable human(-oid) sprites on a clean background for easy processing into 32x32 assets.

As the original assets had been provided under a non-commercial license, you may not monetize this LoRA directly (although charging for hosting is fair game) or circumvent this restriction by merging with other models released under different terms.

Artstyle/Datasources:

As SS13 community is a tangled mess of codebases, the specific source artstyle is specifically roughly mid-to-late-2010s UristMcStation (of which I am a maintainer), containing a mix of custom, Baystation12, and classic tgstation assets.

The training dataset is a curated, hand-annotated batch export of rescaled sprites of humans and humanoids (meaning alien races, human subspecies, humanlike enemies, and clowns), nude or dressed in a variety of equipment, from civilian clothes, through combat gear, to spacesuits.

Usage:

ss13 pixel art sprite of <desc>

The captions include body features (e.g. hair color), accessories (e.g. hats, clothing) and held weapons (use holding or wielding) as well as alien race names in the UMcS convention where applicable, so the LoRA should recognize them.

EDIT Aug 16th 2024:

After playing around with it some more, additional hints:

1. The v0.1 LoRA generalized astonishingly well, and this synergizes nicely with Pony's broad familiarity with pop-culture characters. Try making a 2dspessman out of your waifu!

2. Expanding on the above, while the style is pretty strong, Pony characters work fine and you can force all sorts of more abstract Pony concepts like poses back in; using Comfy, I recommend keeping model strength at 1.0 and boosting the "foreign" concept's tags' strength to about 1.5.

This has reliably yielded Pony poses in the SS13 artstyle for me.

3. The training was done without the score_<X> tags, but score_9 at least seems to make the shading better.