Sign In

明日方舟 风丸 初始服装 Arknights Kazemaru (original uniform)

102
495
5
Type
LoRA
Stats
495
Reviews
Published
Apr 15, 2023
Base Model
SD 1.5
Trigger Words
kazemaru
Hash
AutoV2
2EDDF6126B
default creator card background decoration
linch_'s Avatar
linch_

Kazemaru from Arknights.

Because there're not many high quality images for training, this model is still kinda overfitting. Though most features are preserved pertty well, sometimes the face may have a little distortion. Try rolling the random seed another time while praying to the RNG god.

I recommend using weight 0.5~0.65.

Other than using low weight, I personally recommend using LoRA block weight extension to reduce overfitting while applying with higher weight. Normally you can achieve good results with weight around 0.7.

About this extension, check this GitHub repo: https://github.com/hako-mikan/sd-webui-lora-block-weight

Here's my block weight preset, you can simply save it through the extension:

De-overfitting:1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1

明日方舟 风丸,初始立绘。

因为风丸的高质量图实在太少,我也比较挑剔,这个模型炼出来还是有些过拟合,泛化性能一般。不过大部分风丸的特征都保留得不错,基本不需要额外的tag就能还原,当然如果没还原了肯定还是需要加上的。模型生成的面部有时候会有点崩,不一定是你的权重高了,可能只是种子的问题,建议多试试。

建议权重:0.5~0.65。

除了用低一点的权重防止崩坏以外,我个人推荐使用 lora block weight这个插件,可以有效减弱模型的过拟合,链接和权重预设贴在上面了。教程在B站就有,搜“分层控制”就能看到。基本上加了减弱过拟合的分层权重后,可以用更高的权重实现更好的效果,我一般是用0.7配合这个分层控制。