MICCHAN — SDXL Character LoRA ============================================================ Micchan by Ohisashiburi — the feminine half of the pair. Short white hair, that blue-yellow bow at her collar, grey pleated skirt swaying as she walks. Sometimes cat ears peek through her hair. She's the kind of character who makes you want to protect her. Until you realize she doesn't need it. Nacchan's partner. These two are made for each other — literally. Check my other LoRA if you want the full yuri set. 🏷️ TRIGGER + ANCHOR ============================================================ micchan, 1girl, white hair, short hair, purple eyes, school uniform, cardigan, bow ⚙️ RECOMMENDED SETTINGS ============================================================ LoRA weight: 0.7 to 1.0 I run her at 0.85. At 1.0 she gets extra detailed — the bow, the cardigan folds, the hair highlights all pop. CFG: 5 to 7 Lower CFG (5-5.5) gives her a softer, dreamier look. Higher (6-7) sharpens the details. Sampler: Euler Ancestral or DPM++ 2M Karras Steps: 25 to 35 Resolution: 832×1216 (2:3 portrait) or 1024×1024 ❌ NEGATIVE PROMPT ============================================================ lowres, bad anatomy, bad hands, extra fingers, missing fingers, blurry, monochrome, greyscale, blue skin, grey skin, colored skin, dark skin, black hair, speech bubble, text, watermark, censored ⚠️ KNOWN ISSUES ============================================================ WD14 sometimes tagged her with "colored skin" or "blue skin" — these were removed during caption cleanup, but the LoRA may still pick up slight skin tone shifts at extreme weights. Keep "colored skin, blue skin, grey skin" in negative. Her eyes are pink but WD14 consistently tagged them "purple eyes" — this is what the LoRA learned. They'll render as a soft purple-pink. If you want pure pink, you may need to prompt-engineer with "pink eyes" at higher weight. Not all Illustrious merges work equally well. She's solid on base Illustrious models and bismuthIllustrious (my top recommendation — best results at weight 0.7-1.0). Some merges may produce inconsistent colors or weak character recognition. If a merge isn't cooperating, switch to bismuthIllustrious. At very high weights (above 1.2), the cardigan and bow can dominate outfit changes. Drop to 0.8 for outfit flexibility. 📊 TRAINING DETAILS ============================================================ Base Model: Illustrious XL Network: LoRA — Dim 32 / Alpha 16 Epochs: 10 Batch Size: 2 Optimizer: AdamW with Cosine LR Resolution: 1024 (bucket training, min 256 max 2048) Total Steps: 2,600 Final Loss: 0.0603 Captions: WD14 Tagger ConvNeXt v2 ✅ COMPATIBILITY ============================================================ Tested and consistent on Illustrious XL 2.0. Should work on most Illustrious-based checkpoints (NoobAI, Hassaku, WAI). Not tested on Pony or SD 1.5. When stacking with other character LoRAs, bump micchan's weight to 1.0-1.2 — at equal weights other LoRAs can wash out her white hair and purple eyes. 💡 TIPS ============================================================ She's the expressive one. "smile", "blush", "embarrassed", "happy", "looking at viewer" all work naturally. She's not the brooding type — let her be cute. The cardigan + bow combination is her signature. If you swap outfits, keep the vibe soft and feminine — frilly dresses, sundresses, pastels. At CFG 5 with DPM++ 2M Karras she gets a beautiful soft painterly quality that suits her character perfectly. For cleaner anime lines, bump CFG to 6-7. The animal ears are in about half the training images — they'll appear naturally sometimes. If you don't want them, add "animal ears, cat ears" to negative. 🎨 COMMISSIONS &amp; SUGGESTIONS ============================================================ Got a character you want turned into a LoRA? Drop a comment or DM. I do commissions — character, style, outfit, whatever you need. I'll work with you on dataset and training to get the result you want. Check my profile for Nacchan — her tomboy girlfriend. They're meant to be together.

Files with SHA256: 6f3c4e38c5f50cbc5c220fc27e76fcd7931279478cfeca4960af9b3f4980a5c8