Vesti (NIKKE) | NoobAI-XL V-Pred

Character LoRA for Vesti from Goddess of Victory: NIKKE.

Trained on NoobAI-XL V-Pred 1.0. This is a v-prediction model — make sure your setup handles v-pred / ZTSNR correctly. ComfyUI detects it from the checkpoint metadata; in A1111/Forge you may need the matching .yaml or a V-Pred-aware build. Using it on an eps-prediction checkpoint gives washed-out or burnt results.

Trigger word

vesti

Identity block — always include

vesti, grey hair, blue eyes, short hair

Those four tags lead every training caption in fixed order (keep_tokens 4), so the model expects them there.

The bust tag is a working control — use it. The training set runs a little fuller than the official art, so medium breasts is written into every caption. That means it steers: small breasts pulls the build back toward the official proportions, medium breasts holds the middle, and dropping the tag entirely gives the fuller default. All three respond properly.

The uniform is a switch, not part of the trigger

The trigger gives you Vesti — face, silver bob, blue eyes, build. It does not automatically dress her. About 45% of the training steps are her in ordinary clothes, on purpose, so you can put her in anything.

To get the signature uniform, prompt it:

vesti, grey hair, blue eyes, short hair, beret, black headwear, black jacket,
long sleeves, black shirt, black gloves, fingerless gloves, chest harness, harness,
black belt, red necktie, black cape, two-sided fabric, black shorts, short shorts,
black thighhighs, striped thighhighs, garter straps, black footwear, red footwear,
military uniform

That is the exact block the signature images were captioned with, in that order.

Back view — two extra tags

The two red suspender straps down her back only appear in shots from behind, so they were only captioned there. For a back view, add:

from behind, suspenders

The small stuff that is worth prompting

Three details are easy to lose, and the tagger did not find them on its own — they were written into the captions by hand, which means they now work as switches:

• chest harness, harness — the stacked horizontal webbing rows on her chest. This is the part most Vesti LoRAs turn into a glossy black blob.

• striped thighhighs — the two pale-blue stripes near the top of each thighhigh.

• red footwear — the red soles.

two-sided fabric is what gives the cape its red-orange inner lining.

Want it as close to the official art as possible?

Two things drift away from the source: the cape runs long, and the build comes out a little fuller than the official render. Both are fixable from the prompt, and they have to be used together — either one alone is not enough. Measured against the official full-body art, then confirmed on four seeds.

Add to the positive:

(short cape:1.3), small breasts

Add to the negative:

long cape, floor length cape, cape touching the ground, trailing cape, long coat

That takes the cape from ankle-length up to around mid-thigh — it varies a little by seed, but every seed I tried landed in the zone the official art puts it in — and narrows the silhouette to match. Keep the LoRA weight at 1.0 — the chest harness stays fully defined with these tags added, so you are not trading detail for proportion.

Stay at 1.3, and resist the temptation to raise it. A higher weight does shorten the cape further, and in a plain single-pass generation 1.5 looks perfectly clean — but the moment you put it through hires-fix or an upscale pass, 1.5 falls apart: the chest harness smears together and the collar hardware turns into a mottled mess. I compared 1.3 and 1.5 on the same seed through the same upscale chain, and 1.3 wins clearly. Push one step further to 1.6 and the image collapses into noise outright — no figure at all, on three separate seeds.

Oddly this is not a general weight limit on the model — (chest harness:1.8) and (small breasts:1.6) both render fine on the same prompt. It is this tag specifically, probably because it fights black cape in the positive and the cape terms in the negative at the same time.

I also tried capelet instead of a weighted short cape; it did not do more, and the tag was never in the training captions, so I left it out.

Suggested settings

Samplereuler, normalCFG4.5–5Steps28–30LoRA weight1.0 — do not go under 0.9Resolution832×1216 portrait, 1024×1024 square

Weight matters more than usual — measured

The chest harness is the detail this LoRA exists for, and it is the most weight-sensitive thing in it. Same seed, same prompt, only the LoRA weight changed:

weightchest harness0.5gone — plain black coat, just the necktie0.6a few straps and the round module, patchy0.7the horizontal webbing rows start to resolve0.8clean rows, buckles, round module0.9–1.0densest and best defined — use this

Below 0.9 you slide back toward the glossy black blob that every other Vesti LoRA gives you. The pale-blue thighhigh stripes, by contrast, hold all the way down to 0.5 — so if you are mixing this with other LoRAs and have to drop the weight, the harness is what you lose first.

Negative prompt used throughout testing:

worst quality, worst aesthetic, old, early, low quality, lowres, signature, username,
logo, watermark, artist name, jpeg artifacts, bad hands, mutated hands, extra digits,
fewer digits, missing fingers, bad eyes, asymmetrical eyes, bad anatomy, deformed

Known quirks

• The cape runs long. By default it reaches the ankles; the official art stops around mid-thigh. You can shorten it a lot, but not all the way. Measured on a fixed seed: short cape in the positive gets you to mid-calf, long cape, floor length cape in the negative gets you just below the knee, and both together land at the knee — the shortest I could get. Still longer than canon, so treat it as a dial, not a fix.

• The red devil-head emblem on the beret and upper back is inconsistent. There is no reliable booru tag for it, so it was left out of the captions — you get it when you get it.

• The white star on the thighhigh clip is not reproducible for the same reason.

Training

BaseNoobAI-XL V-Pred 1.0NetworkLoRA, dim 32 / alpha 16, conv 16 / 8Dataset181 uniform + 50 casual imagesSteps2648 (8 epochs)CaptionsWD14 EVA02-Large, then hand post-processing

The dataset is fully synthetic — generated from a written character sheet against the official art, then culled by hand. 17 images were thrown out for body proportion alone.

Versions

v1 — first release.

Post your images

Feel free to post whatever you generate with this — always good to see.

Character requests

Got a character you want next? Drop it in the comments — you've got nothing to lose. Worst case I don't like the idea and just don't make it. Lol