<h2 id="moss-image-loraheal">淡彩萌系画风 · MOSS-Image LoRA（heal 训练）</h2>浅蓝浅紫的淡彩、软软的亮部、圆脸大眼和兽耳小姑娘，整张画面像蒙了一层奶白色的光。用 MOSS-Image-heal 训练，heal 和 SFT 上都能用。<h3>触发词</h3><ul><li>英文 tag / 句子开头写 <code>tyomimas</code></li><li>中文句子开头写 <code>tyomimas画风。</code></li><li>不写触发词也会生效，写上更稳。</li></ul><h3>版本怎么选</h3><ul><li>heal 推荐：s3375（约 33 遍）——在 heal 上出图最完整，淡彩、浅色背景、眼睛画法都到位。</li><li>SFT 推荐：s5033（约 51 遍）——在 SFT 上脸、眼睛、表情还原度很高。</li><li>轻量：s1500（约 14.5 遍）——在 heal 上已经九成像，更听提示词，适合叠别的 LoRA。</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e12b8548-7a55-4242-8f7a-7e209ee9e55b/original=true/e12b8548-7a55-4242-8f7a-7e209ee9e55b.jpeg" /><h3>推荐参数</h3><ul><li>heal：50 步，CFG 3，dpmpp_3m_sde_gpu / sgm_uniform，shift 5</li><li>SFT：40-50 步，CFG 3，euler / simple，shift 5</li><li>两边都建议加 <code>SkipLayerGuidanceDiTSimple</code>（single_layers 22，0～0.5），画面更干净</li><li>LoRA 强度 1.0</li></ul>想要原汁原味就在 heal 上出图。SFT 自带的光影很强，在 SFT 上脸会很像，但淡彩冷调和浅色背景会被 SFT 往它自己的方向拉。<h3 id="heal-heal">为什么用 heal 训，也推荐你用 heal 训</h3>MOSS-Image-heal 是 KohakuBlueleaf在 Base 上做的全量微调：Base 的残差流里有几个通道异常大（最大值约 5.5 万，接近 fp16 上限），heal 把它们压平了，出图和 Base 很接近。我用同一份数据在 Base 和 heal 上各训了一次做对比，再加上另外两个画师的结果：<ul><li>在 SFT 上出图：heal 训的 LoRA 起效约快一倍、同遍数更像，固定姿势这类过拟合也晚一倍多。</li><li>在 heal 上出图：十几遍就很像，15～30 遍基本够用。</li></ul>以前在 Base 上训画风，常常要练到一百多遍、或者得加 DMD 才像；换成 heal 训轻松很多。所以训 MOSS 画风 LoRA，推荐直接用 heal 当训练底模。说明：只测了 3 个画师、一条主测提示词。「LoRA 的小改动被 Base 的大通道淹没」是我的推测，不是官方结论。<h3>训练信息</h3><ul><li>底模：<a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/MOSS-Image-heal">MOSS-Image-heal</a>（int8 convrot）</li><li>171 张；WD14 打 tag，KohakuGBC 生成自然句；中英 tag + 中英自然句四种 caption 混训</li><li>rank 32 / alpha 128，LR 1e-4 恒定</li><li>训练器：<a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/KohakuMossTrainer">KohakuMossTrainer</a>；本机 4070 Ti SUPER 训到 36 遍，RTX 5090 续训到 51 遍</li></ul>感谢 KohakuBlueleaf 的 heal、GBC 和训练器。 Pastel Soft Kemonomimi Style · MOSS-Image LoRA (trained on heal)Pale blue and lavender pastels, soft milky highlights, round faces, big eyes and cute animal-eared girls. Trained on MOSS-Image-heal; works on both heal and SFT.<h3 id="trigger-words">Trigger words</h3><ul><li>English tags / sentences: start with <code>tyomimas</code></li><li>Chinese sentences: start with <code>tyomimas画风。</code></li><li>It works without the trigger too, but the trigger makes it more stable.</li></ul><h3 id="which-version">Which version</h3><ul><li>For heal: s3375 (~33 epochs) – the most complete look on heal: pastel palette, light backgrounds, eye style.</li><li>For SFT: s5033 (~51 epochs) – face, eyes and expressions come through very well on SFT.</li><li>Light: s1500 (~14.5 epochs) – already about 90% there on heal, follows prompts better, good for stacking.</li></ul><h3 id="recommended-settings">Recommended settings</h3><ul><li>heal: 50 steps, CFG 3, dpmpp_3m_sde_gpu / sgm_uniform, shift 5</li><li>SFT: 40 steps, CFG 3, euler / simple, shift 5</li><li>Add <code>SkipLayerGuidanceDiTSimple</code> (single_layers 22, 0–0.5) on both for cleaner images</li><li>LoRA strength 1.0</li></ul>For the most faithful look, generate on heal. SFT has strong built-in lighting: faces match well, but the cool pastel tones and light backgrounds get pulled toward SFT's own style.<h3 id="why-heal-and-why-i-recommend-training-on-heal">Why heal – and why I recommend training on heal</h3>MOSS-Image-heal by KohakuBlueleaf is a full finetune of Base that flattens a few huge channels in the residual stream (Base peaks around 55k, close to the fp16 limit) while keeping outputs close to Base.I trained the same dataset on Base and on heal for a side-by-side comparison, plus two other artists on heal:<ul><li>Generating on SFT: the heal-trained LoRA takes effect about twice as fast and looks closer at the same epoch count; overfitting (e.g. frozen poses) also shows up more than twice as late.</li><li>Generating on heal: it looks right after a dozen or so epochs; 15–30 epochs is usually enough.</li></ul>Style LoRAs trained on Base often needed 100+ epochs or a DMD LoRA to look right. Training on heal is much easier, so I recommend using heal as the training base for MOSS style LoRAs.Note: tested on 3 artists with one main test prompt. "Small LoRA updates get drowned out by Base's huge channels" is my guess, not an official statement.<h3 id="known-issues">Known issues</h3><ul><li>Some checkpoints add a white outline around characters (sticker look); the dataset includes white-background versions.</li><li>On snowy-night scenes, checkpoints past ~36 epochs tend to add a handbag.</li></ul><h3 id="training-info">Training info</h3><ul><li>Base: <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/MOSS-Image-heal">MOSS-Image-heal</a> (int8 convrot)</li><li>171 images; WD14 tags + KohakuGBC natural-language captions; English/Chinese tags and sentences, four caption types mixed</li><li>rank 32 / alpha 128, LR 1e-4 constant</li><li>Trainer: <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/KohakuMossTrainer">KohakuMossTrainer</a>; 36 epochs on a local 4070 Ti SUPER, resumed to 51 epochs on an RTX 5090</li></ul>Thanks to KohakuBlueleaf for heal, GBC and the trainer.

Files with SHA256: d0003868ba71253a6c6ca06cfdb781eb3b29d62a6e09cda473e88439f4c5f29c