Quiry (NIKKE) | NoobAI-XL V-Pred

NoobAI-XL V-Pred character LoRA for Quiry from Goddess of Victory: NIKKE — A.C.P.U. squad, Elysion. Pink twintails, green eyes, and a pink-and-charcoal military school uniform with a small flat-topped cap.

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

It also works on eps-prediction checkpoints — tested on NoobAI-XL 1.1 (eps) and Illustrious-XL 1.0 at CFG 6.0, where the character comes through cleanly. What you lose there is contrast: dark and night scenes wash out to grey, which is exactly what V-Pred is for. Costume micro-details drift a little too.

Trigger word

quiry

Identity block (always include)

quiry, pink hair, green eyes, twintails

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

The dark pointed ribbon bows at the base of her twintails are part of her, not part of an outfit — they appear in the casual images too. Add hair ribbon, black ribbon if you want them sharp.

The switchable state

Hairstyle

twintailstwo tails gathered high at her temples, dark pointed ribbon bows at the basehair downeverything loose, one long mass well past her hips, no bows

hair down was trained with a ×3 repeat so it would hold its own against the twintailed majority — the raw dataset is 240 twintails images against 41 loose ones. It holds cleanly in every framing, and it was already clean at epoch 1.

Signature outfit

The uniform is a switch, not baked into the trigger. Paste this block to get it:

hat, black headwear, mini hat, hair flower, pink flower, collared shirt, white shirt,
puffy short sleeves, frills, purple necktie, chest harness, belt, chain, corset,
black corset, pleated skirt, black skirt, miniskirt, white thighhighs, garter straps,
pink footwear, platform footwear

That is the exact vocabulary the captions used, in the order they used it. Rewriting it in your own words makes the outfit worse — the model learned this wording.

The trigger alone still leans toward the uniform, because 191 of the 281 training images are in it. You will usually get a white collared shirt and a dark ribbon without asking. For the full uniform, paste the block.

Recommended settings

sampler   : euler
scheduler : normal
CFG       : 4.5        (do not go above ~6 on V-Pred)
steps     : 28-30
size      : 832x1248   (the training bucket ratio)
strength  : 1.0

Quality tags (NoobAI style)

masterpiece, best quality, newest, absurdres, highres, very awa

Negative:

worst quality, worst aesthetic, old, early, low quality, lowres, signature, username,
logo, watermark, artist name, jpeg artifacts, bad hands, bad anatomy

If the cap comes out with a brim, add peaked cap, police cap, baseball cap, visor, brim to the negative. Her cap has no brim and no visor.

The chain behind the gallery images

Base generation at 832×1248 → 4×-AnimeSharp upscale, back down to 1.5×, hires img2img at denoise 0.40 → FaceDetailer last → final 4×-AnimeSharp upscale, scaled to 4K.

The detailer goes last on purpose. Run it first and it works on a small face, and the upscale afterwards smears the result.

Known quirks

Most of her small details do come through. The A.C.P.U. emblem (teal star, pink laurel) lands on her right breast and on the cloth at her hip, the pink cross-lacing on the black corset renders cleanly, and the pink heart-shaped rings on her garter straps are there on both thighs. Those were cut from the captions entirely rather than left half-tagged, so the model treats them as part of her rather than as optional accessories.

What is genuinely off:

The cloth at her hip is a rectangle, not a triangle. In the official art it is a folded cloth hanging point-down. It comes out as a rounded rectangular panel. This one was already wrong in the training data and no amount of prompting fixed it.
Rare poses need weighting. The tagger only wrote salute on 10 of the 281 images and from behind on 39, so the LoRA barely reinforced either. Spell them out and weight them: (salute:1.4), (arm up:1.2), hand up, one hand only, right hand at the side of her cap or (from behind:1.5), back, (looking back:1.2).
Her signature squint — half-lidded, unimpressed — is only on 1.8% of the captions, but the base model knows it well. Use (half-closed eyes:1.3), jitome, (v-shaped eyebrows:1.15).
Add very long hair if you want her twintails at full length. Without it they come out around chest length; the captions carry it, so the model expects it.

Training details

base          : NoobAI-XL V-Pred 1.0
network       : LoCon, dim 32 / alpha 16, conv 16 / conv_alpha 8
dataset       : 281 images
                191 signature x1
                 49 casual, twintails x1
                 41 casual, hair down x3
steps         : 363 per epoch, 8 epochs = 2904
optimizer     : AdamW8bit, cosine_with_restarts, 3 cycles
lr            : 1e-4 unet / 5e-5 text encoder
keep_tokens   : 4
v-pred        : --v_parameterization --zero_terminal_snr --scale_v_pred_loss_like_noise_pred

The dataset was built from a handful of official reference images, expanded with generated training images across a wide spread of poses, framings, lighting and settings.

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