本页面包含 MOSS 和 Anima 两个版本，请选择对应的使用说明。
This page includes both MOSS and Anima versions. Please follow the instructions for your model.
推荐参数/Recommended Settings
FOR ANIMA:
Trigger word:
V2 dar:@skp_d
V2 nor:@skp_n
V1 nor:@skpstyle
V1 dar:@skp_darkstyle
LoRA weight: 0.7~1
Sampler: er_sde + beta
V2 TIPS：你可以一起使用两个风格，获得平衡的效果，NOR权重高会偏清透，DAR权高重会偏油腻，你可以按照自己的感觉来玩出最好看的比例。 You can combine both styles to achieve a balanced look. A higher NOR weight gives you a cleaner, clearer appearance, while a higher DAR weight creates a richer, oilier look. Feel free to experiment with different ratios and find the balance that looks best to you!
FOR MOSS：
no Trigger word！！！
LoRA weight: 1-1.3
Sampler: er_sde + beta OR euler + normal
cfg:3
steps:30-50
____________________________________________________________________________________
Normal & Dark
同一位画师，两种画风。
Two styles from the same artist.
**Normal：**明亮柔和，色彩清透。
Normal: Brighter and softer, with a lighter palette.**Dark：**色彩浓郁，阴影厚重，光影对比更鲜明。
Dark: Richer colors, deeper shadows, and more dramatic lightin
MOSS
V1 DAR:
新参数，整体比nor版好。Updated training parameters. Overall, this version performs better than the NOR version.
V2 NOR:
解决了不能使用质量提示词的问题，更换了配方和训练参数，整体效果更好。但注意质量词用不用，表现上仍有分别，原因不明，加上质量词，整体表现更规整、更细腻，不用会更有多笔触。我推测，在MOSS上，会将一些笔触、凌乱的线条当成是低质量的关系，所以一旦加上质量词，一些笔触会消失。
I've fixed the problem where quality prompt tags couldn't be used, and switched to a different recipe and training parameters — the overall results are better. Note, though, that whether the quality terms are used still makes a visible difference, for reasons I don't yet understand: with them the output is more orderly and more refined, without them it shows more brushwork. My hypothesis is that MOSS treats certain brushwork — loose, messy line work — as low quality, so once quality tags are added, some of that brushwork disappears.
V1.1 NOR：Deprecated
_____________________________________________________________________________________________
Anima
V2：
新增了些许数据集。
TARGET_RES:1536，修改了配方和使用了新的训练器，整体表现更好，指令遵从度改善，光影、画面、阴影也有一定改善。
我也试着训练了2048的版本，但我认为过强，纹理过深。
added a small amount of new training data.
TARGET_RES:1536.Updated the training settings and switched to a new trainer, resulting in better overall performance, improved prompt adherence, and noticeable improvements in lighting, image quality, and shadows.
I also experimented with training a 2048-resolution version, but I personally find the effect too strong, with overly pronounced textures.