白葱 | 25时初音未来 | 25-ji Miku | ニーゴミク Project_Sekai

#以下为初版简介，关于v1.0B和1.1见下#

#The following is an introduction to the first version. See show more for v1.0B and 1.1#

第一个lora肯定要练白葱啦！Lora是真好用啊（赞叹）

刻意清除了所有数据集中的黑色绑带，因为我想给她换衣服而不是只有那一套，但介于白裙子的图实在太多了出图还是会倾向于白裙，就真的没办法了

I deliberately removed all the black bands from the dataset, because I wanted to change her clothes and not just that one. But since there are too many pictures of white dresses, the pictures will still tend to white dresses, and there is really no way.

推荐使用 ((twintails), very long fluffy grey hair), heterochromia 锁定特征

It is recommended to use "((twintails), very long fluffy grey hair), heterochromia" to lock features

预览图使用Anything V4.5和Counterfeit-V2.5，在Anything v4.5取得较好效果，Counterfeit-V2.5严重依赖高清修复

权重建议在0.4到1，随情况浮动（因为试的时候确实有很多不同情况…），不知道多少先试0.7 #现版本仅推荐0.5#

There is no optimal weight, may fluctuate between 0.4 and 1, can usually be tried from 0.7 #Only 0.5 is recommended for the current version#

(为什么有张没有tag呢，是因为眼睛反了顺手镜像了一下，然后数据就丢了（））

================================================

v1.0B更新说明（为什么这玩意不能放在更显眼的地方，Version changes or notes不注意根本看不到嘛）

（v1.1基本相同）

v1.0B Update description

(v1.1 is basically the same)

事先说明，这版的效果可能不如上一版，所以标的是B而不是1.1

先前那版是我学会怎么练lora后直接用embedding时代的训练集（也就是512x512，仅自动打标）用了50张练出来的，什么都没调看着不错就留了。现在学会了怎么更多知识就想重置一遍，于是经历了漫长的打标签后…就有了这一版

首先，可能是因为训练分辨率拉高了，这个版本需要更高的分辨率以获得更好的效果（如1024x768或更高）（最开始看到低分辨率下的图还以为炸炉了）。

这一版相比之前过拟合有些严重，权重需调至0.5，同时画面会带有一定灰色调（可能是dim调高了？），但优点在于泛化能力大幅提高，对不同模型画风的适应能力提高了非常多，可以从预览图看出来，比如使用Counterfeit-V2.5时不会糊成块了（除了pastel，那真的是糊的要死）。

因为这次处理数据集时尝试清除了所有与头部特征相关的tag，((twintails), very long fluffy grey hair) 现在不需要了（当然你也可以加上，不过效果不大），但 heterochromia 依然需要，大模型对画同色瞳的执念还是太大了。同时，我尝试引入了关键词 25ji_m ，但目前为止我没发现它发挥了什么大的作用。

建议是两个版本都保留，各有一定优缺点（个人觉得新版本好一些，大显存不腰疼（））

================================================

#All the following English is machine translated, if strange is normal and sorry#

To be clear, this version may not be as effective as the previous version, so it's labeled B instead of 1.1. This is just an attempt to use new knowledge relative to the previous version.

First, it may be that the training resolution is high, and this version requires a higher resolution to get a better effect (such as 1024x768 or higher) (the first thing that starts to see a low rate of discrimination is a bad).

Compared with the previous version, the overfitting is a little serious. The weight needs to be adjusted to 0.5, and the screen will have a certain gray tone (maybe dim was too high?). But the advantage is that the generalization ability is greatly improved, and the ability to adapt to different model painting styles is greatly improved. You can see from the preview chart, for example, Counterfeit-V2.5 won't be burned into blocks.

Because this attempt to process the dataset removed all the tags associated with the header features, ((twintails), very long fluffy grey hair) are no longer needed (you can add them too, but not very effective). But the heterochromia is still needed, The big model's obsession with the same pupil is too big. At the same time, I've tried to introduce the keyword 25ji_m, but so far I haven't found it making much of a difference.

It is suggested to keep both versions, each with certain advantages and disadvantages. (Personally, I think the new version is better.)