我上班都没这么认真过，我是真喜欢她。 I invested significant effort into creating her model. The training dataset totals 236 manually curated high-quality images, which I further categorized into four types: 1、56 lower-quality images (full-body shots, side profiles, dark/blurry scenes, multiple girls) 2、61 medium-quality images 3、59 high-quality images (emphasizing lighting/shadow and facial details) 4、60 instrument-focused images (with musical instruments or complex elements) Tagging Strategy: For high-quality images: Increased threshold during [tag recognition] to minimize tags, allowing the model to learn default styles more freely. For medium-quality images: Applied balanced thresholds for tagging. For lower-quality images: Lowered thresholds to maximize tags, suppressing interference from "blur, instruments, or extraneous elements." Training Approach: Increased training iterations for high-quality images Reduced training iterations for lower-quality images (Why keep lower-quality data? To preserve diversity: full-body poses, alternate outfits, and contextual elements.) Two Training Versions: 1、Adafactor-trained version Strengths: Better preserves female facial features and original art style Weaknesses: Less accurate male character depictions Base triggers: 1、,hibike kuroe mayu,1girl, long hair,blue eyes, 2、,hibike kuroe mayu,1girl, ponytail, short hair,blue eyes, Outfit: 1、kitauji high school uniform, school uniform, brown skirt, white sailor collar,neckerchief, shirt, skirt, pleated skirt, sailor collar, brown shirt, serafuku, long sleeves, brown serafuku 2、school uniform, solo,serafuku,skirt, neckerchief,white neckerchief, sailor collar, long sleeves, pleated skirt, blue serafuku, blue eyes, shirt, black serafuku,black pantyhose,seira uniform,pantyhose, 3、kitauji high school uniform, school uniform, skirt, serafuku, neckerchief, blue skirt, pink neckerchief, pleated skirt, short sleeves, sailor collar,shirt, white shirt, watch, blue sailor collar,black pantyhose,pantyhose, 4、shirt, white shirt, short sleeves,towel, t-shirt,