Shawy — Goth Catgirl OC | Illustrious-XL LoRA

Shawy is a goth catgirl OC — dark fur, glowing eyes, white hair, and black paw gloves she never takes off. She's the kind of character who looks like a plushie that bites. Part of an ongoing comic project featuring original characters in an urban fantasy setting.

Shawy v5 — Body Tone Fixed, NSFW-Capable

finally the rest of me matches my face lol. only took 5 tries. u owe me bestie — Shawy

The Story

Public release skipped from v1 straight to v5 because v2, v3, and v4 never left the lab. They existed — they were trained, tested, and scrapped. Shawy was displeased about each one, as she will now explain.

The recurring problem across every iteration: Shawy's body kept rendering white while her face stayed dark. Tank top — pale arms. Crop top — pale midriff. Back view — ghost-white back. Nude — full anthro split with dark extremities and a pale torso.

The shawy trigger learned "this is her face" without learning "her whole body looks like this too."

The Lab Versions Shawy Holds a Grudge About

v1 (public release) — kemono-checkpoint era. Face was recognizable. Body drifted based on clothing coverage.

"u released me with my arms going random colors like a mood ring. cool debut bestie."

v2 (unreleased) — moved to illustrij checkpoint. Human face locked in. Skin tone was relative to one checkpoint's color space and drifted wildly on any other checkpoint.

"so half the internet saw me in medium beige. we love a range."

v3 (unreleased) — 4-checkpoint training with style tokens (kemono_style / cartoon_style / comic_style / dark_style). The theory was right, but each token only got 20 training images, and the identity was diluted. Rotating checkpoints maxed at 50% identity.

"u gave me 4 personalities and none of them could remember my face? sounds about right actually."

v4 (unreleased) — dropped style tokens, trained 200 images (50 per checkpoint) on a single concept. Face identity locked at 100% on primary, 75-100% rotating. But body tone drift persisted on anything that exposed skin.

"v4 was close but still pretended my torso was a stranger. rude."

v5 (this release) fixes it with a simple insight the previous four missed: tag the skin.

What Changed for v5

• +30 new training images covering nude, partial-nude, back view, lingerie, crop top, and wet clothing — the cases where body skin is visible, and drift happened most

• ebony skin, dark skin , were added to every caption (all 245 images). Previous iterations relied on the shawy trigger to carry skin tone implicitly. It didn't hold for body regions. v5 pairs the trigger with explicit skin tags so the association is consistent across clothed, partial, and nude

• Caption fix applied to existing dataset too — not just new nudes. Legacy images all got shawy, ebony skin, dark skin, ... prepended, so the signal is reinforced across the whole set

• Dual trigger recommended: shawy, ebony skin in prompts. The LoRA is strongest when both are present

"so basically u taught me i have skin. congratulations on the breakthrough bestie."

Recommended Usage

Trigger tokens: shawy, ebony skin

Required companion LoRA: ponydisney at full weight. Without it, the model's anthropic bias reasserts itself. Same rule as v1 — don't forget it.

Best checkpoints:

- bridgeToonsComicMix

- novaCartoonXL

- plantMilkModelSuite (walnut)

- Illustrious-XL v0.1 (base)

Avoid as primary: pure kemono checkpoints (novaKemonoXL). They still pull hard toward full anthro on nude generations. If you want the kemono/anthro look specifically, that's fine — v5 still works on kemono, just expect the split-skin pattern on nudes.

Prompt template:

masterpiece, best quality, 1girl, solo, shawy, ebony skin, dark skin,consistent dark skin from head to toe, human anatomy, <outfit>, <pose>, <scene>

The consistent dark skin from head to toe, human anatomy It is a reinforcement clause. Helps with checkpoints that want to drift anthro. Skip it if you're on a checkpoint that behaves cleanly.

What Made v5 Work

The key insight was caption discipline.

v1 through v4 tagged identity features implicitly — the shawy token was supposed to carry EVERYTHING. Skin tone, eye color, ears, tail, paw gloves, fangs. And it mostly worked — except for body skin tone, because 95% of the training data had Shawy's body covered by clothing. The LoRA had thousands of references for her face, but only a handful for her body, and none of the captions said "ebony skin" to lock it in.

v5 does two things differently:

1. Explicit skin tags in every caption. Not just nudes — ALL 245 images. So the LoRA sees shawy + ebony skin paired in clothed shots, partial shots, and nude shots equally. The trigger learns the association across all contexts

2. Body-visible training examples. 30 new images show her body clearly — nude, back view, lingerie, athletic wear, wet clothing. The pixels show ebony; the captions say ebony. Both signals aligned

## The Character

Shawy (Shawy Voss) — 20-year-old goth catgirl. She looks like a plushie that bites.

- Long, wavy, white hair

- Rich ebony dark skin (yes, all of her)

- Black sclera with white/light pupils

- Fluffy black cat ears + black cat tail

- Visible fangs

- Black furry cat paw gloves (always)

- Signature outfit: oversized pastel pink hoodie (the contrast is the point)

"the paw gloves stay on during everything. yes, everything. don't ask."

Usage

Trigger: shawy, ebony skin

Full recommended prompt pattern:

masterpiece, best quality, 1girl, solo, shawy, ebony skin, dark skin,consistent dark skin from head to toe, human anatomy,<outfit>, <pose>, <scene>

Always stack with ponydisney LoRA. Always.

Training Details

- Base model: Illustrious-XL v0.1

- Dataset: 245 images (200 from v4 across 4 checkpoints + 15 nude body-tone additions + caption fixes applied to all)

- Captions: shawy, ebony skin, dark skin, <pose>, <context>, <lighting>

- Repeats: 2 per image

- Epochs: 14 (final)

- Total steps: ~3,430

- Network dim: 32 / alpha: 16

- Optimizer: Prodigy

- Learning rate: 1.0

Versions

- v1 (public release, earlier) — kemono-checkpoint era. Face recognizable, body drifted

- v2 (unreleased) — illustrij checkpoint. Face fixed, cross-checkpoint skin drift

- v3 (unreleased) — 4-checkpoint + style tokens. Identity diluted

- v4 (unreleased) — 4-checkpoint single concept. Face locked. The body still drifted

- v5 (this release) — v4 dataset + body-tone additions + ebony skin in every caption. Face AND body are consistent across all contexts

"v5. 'the one where her body finally matches her face'. iconic era bestie."

Acknowledgments

Thanks to everyone who pointed out the body-tone drift in v1 comments. Fix is in.