<h1 id="rosy-haze-or-moss-image-lora">Rosy Haze 柔粉雾光画风 | MOSS-Image LoRA</h1>粉紫色调的柔光二次元插画风：整体蒙着一层淡淡的粉色光雾，线条细、发丝轻，脸颊腮红明显，适合校园、冬日、夜景这类安静可爱的单人题材。本 LoRA 为个人爱好训练，与原画师无关，非官方。请勿用于冒充画师本人、伪造约稿或商业用途。<h2>触发词</h2><ul><li><code>@tokkyu</code>（放在提示词最前面）</li></ul>训练标注是英文 tag（WD14），推荐用英文 tag 写提示词。MOSS 的画风 LoRA 不写触发词也会生效，写上更稳。<h2>推荐设置</h2><ul><li>不加 DMD：40-50步，CFG 3，euler / simple，shift 5</li><li>加 DMD：叠加 8 步 DMD 加速 LoRA，12 步，CFG 1，euler / simple</li><li>LoRA 强度：1.0</li><li>分辨率：约 1MP，例如 1024×1536、832×1216</li></ul>底模说明：在 MOSS-Image SFT（int8）和 Base（int8）上都测过。SFT 自带画风比较重，会把效果往它那边拉；想要画风更明显，用 Base，或在 SFT 上加 DMD（颜色更浅更亮，最接近原画风）。<h2>提示词示例</h2><code>@tokkyu, 1girl, solo, silver hair, long hair, wolf ears, wolf tail, winter coat, scarf, standing in the snow, night, street lights, looking at viewer, smile</code><code>@tokkyu, 1girl, solo, blonde hair, short hair, school uniform, blazer, sitting on a desk, classroom, sunset, window, holding a book, looking at viewer</code><code>@tokkyu, 1girl, solo, black hair, twin braids, yukata, holding a paper fan, summer festival, fireworks, night sky, upper body, looking back</code>负面：<code>worst quality, bad anatomy, watermark, jpeg artifact, blurry</code>小提示：<ul><li>团扇、招牌上出现乱码字，是提示词没指定文字造成的，不是 LoRA 的问题；想要文字就把内容写清楚。</li><li>训练图都是单人图，写单人题材效果最好。</li></ul><h2>训练信息</h2><ul><li>底模：MOSS-Image Base</li><li>数据：36 张，2023 年以后的作品，剔除了线稿和带大字的海报</li><li>训练器：MonadForge-Kohaku（本地 RTX 4070 Ti SUPER 16G）</li><li>rank 32 / alpha 128，学习率 1e-4（前 5% 预热），每步 1 张；e112 = 4032 步，e140 = 5040 步</li><li>标注：WD14 英文 tag + 触发词</li></ul><hr /><h2 id="english">English</h2>Rosy Haze anime style for MOSS-Image. Soft pink-purple tones with a light hazy glow, fine linework, airy hair strands and noticeable blush. Works well for quiet, cute single-character scenes: school, winter, night streets.Fan-made and unofficial; not affiliated with the original artist. Do not use it to impersonate the artist or for commercial purposes.Trigger: <code>@tokkyu</code> at the start of the prompt. Trained on English WD14 tags, so tag-style prompts are recommended.Versions: e112 = recommended, brighter and most balanced / e140 = stronger style but darker and duller.Settings: 40 steps, CFG 3, euler / simple, shift 5; or with an 8-step DMD LoRA at 12 steps, CFG 1. LoRA weight 1.0. Around 1MP, e.g. 1024×1536. Tested on MOSS-Image SFT and Base (int8). SFT has a strong built-in style that pulls the result toward it; for a stronger effect use Base, or add DMD on SFT.Training: 36 images, MOSS-Image Base, MonadForge-Kohaku, rank 32 / alpha 128, lr 1e-4, 1 image per step; e112 = 4032 steps, e140 = 5040 steps.

Files with SHA256: 8497fa498e7982a1710f2cdb8950ec7c0e8afadcd1d36d4bb16a75f64a5194a0