淡彩萌系画风 · MOSS-Image LoRA（heal 训练）

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。

触发词

• 英文 tag / 句子开头写 tyomimas

• 中文句子开头写 tyomimas画风。

• 不写触发词也会生效，写上更稳。

版本怎么选

• 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。

• SFT 推荐：s5033（约 51 遍）——在 SFT 上脸、眼睛、表情还原度很高。

• 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。

推荐参数

• heal：50 步，CFG 3，dpmpp_3m_sde_gpu / sgm_uniform，shift 5

• SFT：40-50 步，CFG 3，euler / simple，shift 5

• 两边都建议加 SkipLayerGuidanceDiTSimple（single_layers 22，0～0.5），画面更干净

• LoRA 强度 1.0

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。

为什么用 heal 训，也推荐你用 heal 训

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：

• 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。

• 在 heal 上出图：十几遍就很像，15～30 遍基本够用。

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。

训练信息

• 底模：MOSS-Image-heal（int8 convrot）

• 171 张；WD14 打 tag，KohakuGBC 生成自然句；中英 tag + 中英自然句四种 caption 混训

• rank 32 / alpha 128，LR 1e-4 恒定

• 训练器：KohakuMossTrainer；本机 4070 Ti SUPER 训到 36 遍，RTX 5090 续训到 51 遍

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.

Trigger words

• English tags / sentences: start with tyomimas

• Chinese sentences: start with tyomimas画风。

• It works without the trigger too, but the trigger makes it more stable.

Which version

• For heal: s3375 (~33 epochs) – the most complete look on heal: pastel palette, light backgrounds, eye style.

• For SFT: s5033 (~51 epochs) – face, eyes and expressions come through very well on SFT.

• Light: s1500 (~14.5 epochs) – already about 90% there on heal, follows prompts better, good for stacking.

Recommended settings

• heal: 50 steps, CFG 3, dpmpp_3m_sde_gpu / sgm_uniform, shift 5

• SFT: 40 steps, CFG 3, euler / simple, shift 5

• Add SkipLayerGuidanceDiTSimple (single_layers 22, 0–0.5) on both for cleaner images

• LoRA strength 1.0

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.

Why heal – and why I recommend training on heal

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:

• 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.

• Generating on heal: it looks right after a dozen or so epochs; 15–30 epochs is usually enough.

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.

Known issues

• Some checkpoints add a white outline around characters (sticker look); the dataset includes white-background versions.

• On snowy-night scenes, checkpoints past ~36 epochs tend to add a handbag.

Training info

• Base: MOSS-Image-heal (int8 convrot)

• 171 images; WD14 tags + KohakuGBC natural-language captions; English/Chinese tags and sentences, four caption types mixed

• rank 32 / alpha 128, LR 1e-4 constant

• Trainer: KohakuMossTrainer; 36 epochs on a local 4070 Ti SUPER, resumed to 51 epochs on an RTX 5090

Thanks to KohakuBlueleaf for heal, GBC and the trainer.