<hr /><h1 id="anchor-innocent-maid-(nikke)-or-noobai-xl-v-pred">Anchor — Innocent Maid (NIKKE) | NoobAI-XL V-Pred</h1>Character LoRA for Anchor in her Innocent Maid 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 <code>.yaml</code> or a V-Pred-aware build. Using it on an eps-prediction checkpoint gives washed-out or burnt results.<h2 id="trigger-word">Trigger word</h2><pre><code>anchormaid</code></pre>Important: the trigger word alone will not reproduce the character. Hair and eye colour tags were deliberately kept in the training captions so they stay promptable — which means you have to include them. Always prompt the identity block.<h3 id="identity-block-always-include">Identity block — always include</h3><pre><code>anchormaid, 1girl, solo,
blue hair, blue eyes, long hair, side ponytail, sidelocks, ahoge,
maid headdress, hair ribbon, blue ribbon, black ribbon,
black choker, anchor, gold pendant</code></pre><h3 id="signature-outfit-the-innocent-maid-set">Signature outfit — the Innocent Maid set</h3><pre><code>black bikini, frilled bikini, front-tie bikini top, halterneck, string bikini,
white skirt, frilled skirt,
blue jacket, open jacket, off shoulder, long sleeves, partially unzipped,
bare shoulders, collarbone, navel, blue nails</code></pre>The LoRA separates identity from clothing well — swap the outfit block freely for <code>serafuku</code>, <code>sweater</code>, <code>yukata</code>, <code>winter coat</code>, <code>maid apron</code> and the character still holds together.<h2 id="recommended-settings">Recommended settings</h2><ul><li>LoRA weight: 1.0 — drop to 0.8 if it feels too strong or you want fewer starfish</li><li>Base model: NoobAI-XL V-Pred 1.0</li><li>Sampler: euler + normal scheduler</li><li>CFG: 4.5 — do not push above ~6, it is v-pred</li><li>Steps: 28–30</li><li>Resolution: 832×1216 portrait, or 1216×832 landscape</li></ul><h3 id="quality-tags-(noobai-style)">Quality tags (NoobAI style)</h3>Positive prefix:<pre><code>masterpiece, best quality, newest, absurdres, highres, very awa</code></pre>Negative:<pre><code>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</code></pre><h2 id="known-quirks">Known quirks</h2>These are baked into the model — worth knowing before you fight them in the prompt.<ul><li>The starfish hair ornament multiplies. It often shows up two or three times, and sometimes bleeds onto clothing as orange starfish-shaped patches. Removing <code>starfish hair ornament</code> from the prompt does not fix it — it is learned from the training data. Put <code>starfish, star</code> in the negative, or lower the LoRA weight to 0.8.</li><li>Do not prompt <code>bell</code>. The bell belongs on the black hair bow, but the tag reliably migrates onto the choker and replaces the gold anchor pendant. Use <code>anchor, gold pendant</code> instead.</li><li>Maid outfit bleed. In non-canonical outfits the maid dress/apron sometimes overrides what you asked for. Add <code>apron, maid dress, white dress</code> to the negative when generating casual clothing.</li><li>White frills bleed onto other garments — hems, cuffs. Add <code>frills</code> to the negative when not generating the signature outfit.</li><li>Small faces. At full-body framing the eyes degrade, as with any SDXL model. Use closer framing, or run a hires/upscale pass — an upscale followed by a face detailer works far better than the reverse.</li></ul><h2 id="training-details">Training details</h2><ul><li>Base: NoobAI-XL V-Pred 1.0</li><li>Trainer: kohya sd-scripts, <code>sdxl_train_network.py</code></li><li>Dataset: 149 images — 77 signature outfit (×2 repeats) + 72 other outfits (×1)</li><li>Captions: WD14 EVA02-Large tagger v3, manually normalised</li><li>Steps: 2260 — 226/epoch × 10 epochs</li><li>Network: LoRA, dim 16 / alpha 8</li><li>Optimizer: AdamW8bit, cosine_with_restarts (3 cycles), 100 warmup steps</li><li>LR: unet 1e-4, text encoder 2e-5</li><li>Resolution: 1024×1024 with bucketing, 512–2048</li><li>Extras: min_snr_gamma 5, v_parameterization, zero_terminal_snr, scale_v_pred_loss_like_noise_pred</li><li>Flip augmentation: off — the side ponytail, bow and starfish are asymmetric</li></ul>Captions were post-processed: the tagger emitted the same white frilled maid headdress under four different names (<code>hat</code>, <code>white headwear</code>, <code>beret</code>, <code>mob cap</code>), which were unified to <code>maid headdress</code>; five images where the tagger missed it entirely were corrected by hand.Note on the dataset: the majority of the training images are AI-generated interpretations of the official design, with a small number of official illustrations. Small deviations from the in-game outfit are therefore possible.<h2 id="versions">Versions</h2><ul><li>v1.0 (epoch 10) — recommended. Most complete detail: starfish ornament, headdress frills, jacket anchor emblems.</li><li>Epochs 6 and 8 are also available if you prefer a lighter touch. Epoch 8 tends to flatten the headdress frills into a plain beret.</li></ul><hr /><h2 id="post-your-images">Post your images</h2>Feel free to post whatever you generate with this — always good to see.<h2 id="character-requests">Character requests</h2>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

Files with SHA256: 1f3608ba26fa7a662fbd7ed784aa61446f318b37e496cdfd80d83cb848418ca3