2023,04,23 v3.1
提升了一点泛用性,其余的和下面一样。
Google Translate:
A little generality has been improved, and the rest is the same as below.
2023,04,20 v3.0
三个衣服能正常画出来了,和服最容易(素材给多了),另外两个难
和服:Renge-ki, kimono
私服:Renge-pc, black dress
水手服:Renge-se, serafuku
这三个衣服nsfw有点费劲,低强度描述一下人物在画。
(模型练的很强,建议小于0.5)
Google Translate:
The three clothes can be drawn normally, The kimono is the easiest (more materials are given), the other two are difficult.
Kimono: Renge-ki, kimono
Private Clothes: Renge-pc, black dress
serafuku: Renge-se, serafuku
The nsfw of these three clothes is a bit difficult, the intensity is low, describe the characters again, and then go to draw.
(The model is very strong, it is recommended to be less than 0.5)
OLD
v1.0 更像 不好换衣服,v2.0不像 更好换衣服, 强度0.6-0.8,例图用的 Cetus-Mix V3
两件衣服:kimono, dress
Google Translate:
v1.0 is more similar, but it is not easy to change clothes, v2.0 is easy to change clothes, and the strength is controlled at 0.6-0.8, the example picture is used Cetus-Mix V3
two clothes: kimono, dress
Description
更新能画好三件衣服
但是对应衣服的裸体我没练,所以写对应衣服+nude他指不定怎么给你脱
FAQ
Comments (9)
Great work! Can you update 篝ノ 霧枝lora?
过拟合有点严重,换衣服太难了,lora block都不一定好使,希望能练一个泛化好一点的,不一定要多皮肤,先关注着~加油加油~
OK,我的疏忽,还有一个v2我来试试,但还是期待一个更像的容易换衣服的版本(单皮肤好练一点)
我还弄不明白怎么在游戏立绘只有半身和三个动作的情况下提高泛化度 XD
0.4强度去用就行了,我练得lora都是没几个同人图的,泛用没几个高的
@Nicaine 如果只追求还原形貌,把多个衣服一块练,20张图应该够的,打标直接用tagger处理(0.6~0.7阈值差不多),只屏蔽人物形貌的tag(black hair等等),衣服相关的tag不处理,炼出来的基本能保证泛化,衣服的按照对应衣服的描述其实基本也能还原,当然如果追求形貌还原,衣服还原也是次要的。
@tanger 三个衣服提示词一个都不写,只描述头发眼睛就能换衣服了,刚好反过来。
另外其实没20张那么多,一个和服三个角度三张。,黑裙子一张还没有全身,水手服三张也没全身。,都是切块切出来的数量
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.


