<h1 id="enikk-(nikke)-or-noobai-xl-v-pred">Enikk (NIKKE) | NoobAI-XL V-Pred</h1>Character LoRA for Enikk from Goddess of Victory: NIKKE. Gothic-techwear signature outfit, dark skin, platinum-white hair and a floating black-violet thorn halo.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 <code>.yaml</code> or a V-Pred-aware build. On an eps-prediction checkpoint you will get washed-out or burnt results.<hr /><h2 id="trigger-word">Trigger word</h2><pre><code>enikk</code></pre><h3 id="identity-block-(always-include)">Identity block (always include)</h3><pre><code>enikk, white hair, dark skin, dark-skinned female, twintails</code></pre>Those five tags lead every training caption in fixed order (<code>keep_tokens 5</code>), so the model expects them there. The last one is the hairstyle — see below.What you do not need to prompt: the amber-gold eyes, the blunt fringe, the hime-cut sidelocks, and the thorn halo. Those were deliberately cut from the captions so the trigger word carries them, and it does — the halo shows up in outfits and scenes nowhere near the training distribution.The halo is not optional. It is part of the character, so it was trained as identity rather than as a switchable accessory. There is no reliable way to prompt it away.<hr /><h2 id="the-two-switchable-states">The two switchable states</h2>Both of these are real, tested switches — not hopeful tags.<h3 id="hairstyle">Hairstyle</h3><pre><code>twintails → two tails gathered at the temples, plain dark ties
hair down → everything loose, one long mass below the hips</code></pre><code>hair down</code> was trained with a ×2 repeat specifically so it would hold its own against the twintailed majority. It does, cleanly, in every framing.<h3 id="veil">Veil</h3><pre><code>(veil:1.4), see-through</code></pre>The veil is the sheer drape that hangs from her headdress. Two things matter:<ul><li>Include <code>see-through</code>. It is what makes the veil hang in front of her face rather than sweeping to the sides.</li><li>Leave <code>looking at viewer</code> out. That tag is far more common on the unveiled training images and quietly pulls the whole look back toward "no veil".</li></ul><code>(veil:1.2)</code> is a light drape, <code>(veil:1.6)</code> a heavy, long one that reaches past the shoulders.<hr /><h2 id="signature-outfit">Signature outfit</h2>Upper body:<pre><code>white dress, detached sleeves, white gloves, elbow gloves, black necktie,
bare shoulders, white headwear, harness, multiple straps, belt, buckle</code></pre>Add this for anything full-body:<pre><code>white bodysuit, white pants, tight pants, tight clothes,
white footwear, high heels, see-through skirt, zipper</code></pre>Without that second block the model draws dark bare legs and separate shoes. On the actual design the lower body is one continuous white garment — thigh to heel, no seam, the shoe is part of the suit. The tagger never described it, so you have to.The <code>harness, multiple straps, belt, buckle</code> part is the same story for the torso: there is an extra ribbed layer over the abdomen with horizontal straps, a buckle and a hanging tag. Leave those tags out and the torso renders as a smooth white surface.Identity and clothing are cleanly separated — swap the whole outfit block for <code>serafuku</code>, <code>maid</code>, <code>miko</code>, <code>hoodie</code>, <code>business suit</code>, <code>bikini</code>, whatever, and the character still holds together.<hr /><h2 id="recommended-settings">Recommended settings</h2>LoRA weight<code>1.0</code>Base modelNoobAI-XL V-Pred 1.0Sampler<code>euler</code> + <code>normal</code> schedulerCFG<code>4.5</code> — do not push above ~6, it is v-predSteps28–30Resolution832×1248 or 1024×1024 (both are native training buckets)<h3 id="quality-tags-(noobai-style)">Quality tags (NoobAI style)</h3>Positive<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, bad eyes, asymmetrical eyes, cross-eyed, extra pupils,
blurry eyes, bad anatomy, deformed</code></pre><h3 id="the-chain-behind-the-gallery-images">The chain behind the gallery images</h3>Everything above works on its own. The samples additionally went through this, and the ordering matters:<ol><li>Generate at 832×1248, <code>euler</code> + <code>normal</code>, CFG 4.5, 30 steps</li><li>Upscale with an anime ESRGAN model (<code>4x-AnimeSharp</code>), scale back down to 1.5×</li><li>Hires pass: img2img at <code>denoise 0.40</code>, 24 steps, same prompt</li><li>Face detailer last, <code>denoise 0.45</code></li><li>Final ESRGAN pass, scaled to 2560×3840</li></ol>Do step 4 last. A detailer before the upscale works on a tiny face, and the upscale then smears whatever it produced.<hr /><h2 id="known-quirks">Known quirks</h2>Tested, not guessed.Name colours as weighted tags, not in sentences. The captions are tag-style, so the model reads long natural-language clothing descriptions loosely and lets the colour drift to its own default. <code>a tailored scarlet blazer over a white shirt</code> came out navy; <code>(scarlet blazer:1.4)</code> came out scarlet. Same for everything else — weight the colour.The violet markings on the veil do not come back in colour. The veil carries small four-pointed cross motifs (those render fine on their own) and violet drip lines down the headdress. The drip shapes are there; the colour washes out to white, because nothing in the captions ever named it. <code>purple markings</code> in the prompt restores the violet but tints the whole image; a negative that stops the spread also kills the violet entirely. If you need it, do a masked low-denoise pass over the head region instead.Faces run dark. Dark skin plus dramatic or low-key lighting makes the face read very dark. <code>soft even lighting</code> or <code>side window light</code> in the prompt fixes it; heavy backlighting turns her into a silhouette.Full-body eyes. As with any SDXL model, at full-body framing the eyes degrade. The upscale-then-detail chain above is what fixes it.<hr /><h2 id="training-details">Training details</h2>BaseNoobAI-XL V-Pred 1.0Trainerkohya sd-scripts, <code>sdxl_train_network.py</code>Dataset257 images — 163 signature (×1) + 47 casual twintails (×1) + 47 casual hair-down (×2)CaptionsWD14 EVA02-Large tagger v3, unified and partly written by handSteps2736 (304/epoch × 9 epochs)NetworkLoCon — LoRA dim 32 / alpha 16, conv dim 16 / conv alpha 8OptimizerAdamW8bit, <code>cosine_with_restarts</code> (3 cycles), 100 warmup stepsLRunet <code>1e-4</code>, text encoder <code>5e-5</code>Resolution1024×1024 with bucketing; sources were 1086×1448, 1254×1254 and 1024×1536Extras<code>min_snr_gamma 5</code>, <code>keep_tokens 5</code>, <code>v_parameterization</code>, <code>zero_terminal_snr</code>, <code>scale_v_pred_loss_like_noise_pred</code>Flip augoffOn the captions. Eye colour, hair shape and the halo were removed so the trigger word would absorb them rather than compete with them. That worked — the halo appears from epoch 4 onward even in short prompts with clothing found nowhere in the training set.Hair colour and skin tone were deliberately kept, unified to one value on every image. An earlier project taught this the hard way: strip the hair colour tag and the base model's black-hair prior wins. The same reasoning applies to dark skin.The hairstyle and veil tags were written from the source folders, not taken from the tagger. The tagger only saw <code>twintails</code> on 52–72% of the images and produced essentially nothing for the loose hair (<code>hair down</code> on 1 image out of 47). A switchable trait cannot be learned from a signal that patchy, and the folder layout already knew the answer.Note on the dataset: the training images are AI-generated interpretations of the official design, so small deviations from the in-game outfit are possible.<hr /><h2 id="versions">Versions</h2><ul><li>v1.0 (epoch 9) — recommended. LoCon, 2736 steps. Halo, hairstyle switching and the veil all hold up outside the training distribution.</li><li>Epochs 7 and 8 from the same run are near-identical; 9 is marginally closer on the headdress detail.</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: 4b9025c5fa42107322449892c404d1d62b7268b77ebac813b61edc91f1b5bbd5