Rapi — Cherished Red (NIKKE) | NoobAI-XL V-Pred

Character LoRA for Rapi in her Cherished Red outfit 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

rapicherished

Identity block — always include

rapicherished, red hair, long hair, medium breasts, 1girl, solo

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 skews busty, so the default runs fuller than the official art. medium breasts pulls it back noticeably, small breasts further still, and both respond properly. Drop the tag entirely and you get the fuller default.

What you do not need to prompt: the amber eyes, the black choker with the red diamond, and the ahoge. Those tags were deliberately cut from the captions so the trigger word carries them, and it does — across outdoor scenes, costumes and other outfits alike.

Signature outfit — the Cherished Red set

black camisole, camisole, spaghetti strap, see-through, lace trim,
pink jacket, open jacket, off shoulder, long sleeves,
pink shorts, dolphin shorts, short shorts,
barefoot, bare shoulders, collarbone

Identity and clothing are cleanly separated — swap the outfit block for serafuku, sweater, maid, miko, winter coat, business suit, whatever, and the character still holds together.

Recommended settings

• LoRA weight: 1.0

• Base model: NoobAI-XL V-Pred 1.0

• Sampler: euler + normal scheduler

• CFG: 4.5 — do not push above ~6, it is v-pred

• Steps: 28–30

• Resolution: 832×1248 — the native training size — or 1248×832

Quality tags (NoobAI style)

Positive prefix — this one matters more than usual here; the training set is dense, detailed art and the model responds to it:

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, mutated hands, extra digits,
fewer digits, missing fingers, censored, mammal, anthro, furry, ambiguous form, feral,
semi-anthro, bad eyes, asymmetrical eyes, cross-eyed, extra pupils, blurry eyes

The chain behind the sample images

Everything above works on its own. The gallery images additionally went through this, and it is worth copying — especially the ordering:

1. Generate at 832×1248, euler + normal, CFG 4.5, 30 steps

2. Upscale with an anime ESRGAN model (4x-AnimeSharp), then scale back down to 1.5× — 1248×1872

3. Hires pass: img2img over that at denoise 0.40, 24 steps, same prompt

4. Face detailer last, denoise 0.45

Do step 4 last. A detailer run before the upscale works on a tiny face, and the upscale then smears whatever it produced.

In ComfyUI this needs ComfyUI-Impact-Pack plus face_yolov8m.pt, sam_vit_b_01ec64.pth and 4x-AnimeSharp.pth. In A1111 / Forge the same thing is Hires. fix followed by the ADetailer extension — same idea, different buttons.

Copying a pose from a reference image

If you want a specific pose, use Canny ControlNet, not OpenPose. That is the opposite of the usual advice, and it is worth knowing before you waste an afternoon on it.

OpenPose is trained on photographs. Given an anime illustration — especially a seated pose with long hair covering the torso — the detector produces a garbled skeleton, and the result ignores your reference entirely. Tested it side by side: OpenPose returned a front-facing kneeling figure from a reference of someone sitting turned to the side.

Canny works, and works at low strength:

Reference → Canny (low 0.25 / high 0.65) → ControlNet Union (canny type)
strength 0.5, end_percent 0.85

Strength 0.5 is the sweet spot. The pose transfers completely at that setting, and the edge map from a soft, painterly reference is sparse enough that only the posture carries over — the face, fabric, hair and background all still come from the model. Push it to 0.9 and you start tracing the reference's silhouette for no extra benefit.

One thing that helps: leave the pose tags out of the prompt entirely when using ControlNet. Keep the character, the outfit and the setting; let the ControlNet handle the body. Otherwise the two fight each other.

Known quirks

Tested, not guessed.

• The choker's red diamond needs close framing. The choker itself is baked into the trigger and turns up on its own in most scenes. The four-diamond ornament on it, however, only renders sharply in portrait and upper-body shots — at full-body distance it softens into a red smudge, and in dark or strongly backlit scenes it disappears into shadow entirely.
  
  Two things help. Put black choker in the prompt explicitly for anything wider than a cowboy shot — it will not summon detail that the resolution cannot hold, but it does stop the choker vanishing altogether. And skip scarves and turtlenecks: they cover it. If you want a winter scene with the choker visible, use earmuffs and bare neck instead, and put scarf, turtleneck, high collar in the negative.

• She faces the camera unless you fight it. 136 of the 142 training images are tagged looking at viewer, and that bias is strong enough to override prompt tags. from behind or from side on their own will simply be ignored.
  
  For a back view, the negative prompt is what does the work — looking at viewer, front view, facing viewer — combined with from behind, facing away, looking back, back in the positive. For a body turned to the side while still facing the camera, prompting will not get you there at all; that needs ControlNet (see above).

• Bust runs fuller than the official art by default. That is what the training images look like, so it is baked in — but the tag is a genuine lever, not decoration. The captions were tagged per image rather than forced to one value, and the difference between large / medium / small breasts in the prompt is clearly visible.

• Coloured light shifts her hair. Neon and heavy colour casts pull the auburn towards magenta or purple. If you want a night-city or club scene, city lights, bokeh behaves; neon lights does not.

• Always name the garment in bedroom scenes. A large share of the training set is a pyjama set on a bed, and the model fills any gap with it. Ask for casual clothes there and you get a nightgown; leave the clothing tag out altogether — sleeping, on bed — and it improvises bare shoulders and an implied topless figure. Neither is a bug, but neither is what you asked for. Spell the outfit out (hoodie, t-shirt, sweater) and it behaves.

• The “YES” pillow. About 18% of the training images feature a white pillow with YES printed on it. It rarely appears unprompted, but if one shows up in a bed scene, put yes-no pillow in the negative.

• Blush is on by default. 97% of the training images are tagged blush. expressionless or serious clears it without a fight.

• No bedroom lock-in. Despite the indoor-heavy training set, outdoor scenes — streets, beaches, snow at night, shrines — come out clean with no coaxing.

• Small faces. At full-body framing the eyes degrade, as with any SDXL model. The upscale-then-detail chain above fixes it.

Training details

• Base: NoobAI-XL V-Pred 1.0

• Trainer: kohya sd-scripts, sdxl_train_network.py

• Dataset: 142 images — 61 signature outfit (×2 repeats) + 81 other outfits (×1)

• Captions: WD14 EVA02-Large tagger v3, unified and partly cut by hand

• Steps: 2842 — 203/epoch × 14 epochs

• Network: LoCon — LoRA dim 32 / alpha 16, conv dim 16 / conv alpha 8

• Optimizer: AdamW8bit, cosine_with_restarts (3 cycles), 100 warmup steps

• LR: unet 1e-4, text encoder 5e-5

• Resolution: 1024×1024 with bucketing; every source image was 832×1248

• Extras: min_snr_gamma 5, keep_tokens 4, v_parameterization, zero_terminal_snr, scale_v_pred_loss_like_noise_pred

• Flip augmentation: off

The conv layers are included (LoCon) rather than attention-only, which sharpens fabric and lace detail. It is still a standard .safetensors LoRA — ComfyUI and A1111 both load it with no extension needed.

On the captions. The eye colour, choker and ahoge tags were deliberately removed so the trigger word would absorb them rather than compete with them — that part worked, and those elements now come for free. Hair colour, hair length and bust were tried the same way in an earlier run and failed: without a tag to anchor them the base model's black-hair prior won, and the build drifted fuller. They are back in, unified to one value across all 142 images.

That unification mattered, because the tagger was inconsistent: it read the same auburn hair as red hair (97%), orange hair (20%) and brown hair (14%); the same amber eyes as orange eyes (82%), red eyes (71%) and brown eyes (16%), often on one image; and the signature top four different ways (camisole / black camisole / babydoll / black babydoll). Leaving any of those splits in place teaches a contradiction.

Note on the dataset: the majority of the training images are AI-generated interpretations of the official design. Small deviations from the in-game outfit are therefore possible.

Versions

• v1.0 (epoch 14) — recommended. LoCon, 2842 steps. The choker and eyes hold up in scenes far from the training distribution, and the bust tag actually works.

• Epoch 12 from the same run is available as a slightly lighter alternative.

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