***未经允许,禁止搬运
点击右侧的 "about this version" 阅读版本说明
***脸部用adetailer修复
插件地址:https://github.com/Bing-su/adetailer
修复模型地址:https://huggingface.co/Bingsu/adetailer/tree/main
喜欢用二次元底模的兄弟姐妹可以试试我的手绘LORA,精心制作的,绝对美观好用,只影响画风,可画任何角色:
https://civarchive.com/models/160858/oror-hand-drawn-style-monochromelight-color
————————————————————————————————————————
Click "about this version" on the right to read the instruction of the model you're about to download
***Better use adetailer to fix the faces
script's page:https://github.com/Bing-su/adetailer
model's page:https://huggingface.co/Bingsu/adetailer/tree/main
If you like using anime checkpoints, then try my hand-drawn LORA, which you can use to paint any character: https://civarchive.com/models/160858/oror-hand-drawn-style-monochromelight-color
Description
(不知道为啥,我一共上传了6张图片,没有任何NSFW元素,但是只有两张过审了,可能是因为这个水手服泳衣,xdm先凑活着看吧)
本期是蒙德第一偶像芭芭拉小姐~
关键词:((1girl)), babara_in_swimsuit, cowboy shot, blonde hair, blue eyes, twintails, sailor hat, sailor collar, babara's swimsuit, detached sleeves, bare legs, thigh strap, babara's strap, babara's sandals
LORA权重:0.65~0.85
***关于模型权重的问题,我想再提一嘴,刚入门的萌新们可以看一下,LORA的最佳权重跟你用的底模、tag权重、tag数量都有关系,所以我给的范围只是一个参考,最好的权重值还需要你自己去试,不过一定要遵循一个原则:在保证还原度的前提下,LORA权重的值越低越好
(注:本模型不含任何特定的、可能引发争议的真人面部,数据集的脸部由CyberBeautyV20和majicMix sombreV2.0融合而来)
————————————————————————————————————————
(I uploaded a total of 6 images, but for some reason only two of them went through the moderation, all images have no NSFW elements)
Babara in her swimsuit.
Trigger Words:((1girl)), babara_in_swimsuit, cowboy shot, blonde hair, blue eyes, twintails, sailor hat, sailor collar, babara's swimsuit, detached sleeves, bare legs, thigh strap, babara's strap
LORA's weight:0.65~0.85
***The face of this model doesn't come from any specific real person, I fused the faces of CyberBeautyV20 and majicMix sombreV2.0 and got this face.
FAQ
Comments (5)
大佬,神子的低束长发(low-tied long hair),有没有其他合适的提示词可以辅助触发呢?
我权重调到了0.8,但生成的还是普通散发,无法在底端束起来。
(我用的是majicmix的Lux版本)
这是训练集的原因了,训练集里带这个束发的基本没有,所以没有炼进去。当然也有我自己的原因,我当时没有注意到这个细节,否则专门对其特化训练一下还是有很好效果的⸝⸝· ᴥ ·⸝⸝
@SettingsNoProblem 原来是这样
说到这个我想问下大佬,2D插图可以用来训练majic这种真人模型吗?
比如我想在majic模型上训练一个姿势,但只找到了2D插图,没有真人图片,这种情况也可以用于训练真人模型吗?
@SAY_AI 没试过,但感觉应该练不出来,我猜效果会变成2.5D,效果最好的情况应该也就是排名最靠前的那个神子的模型。不过我可以给你说个方法你可以试试,你如果只有2D图片,可以先用真人模型练出一个2.5D效果的模型,然后再用这个模型产出的图片再去练,反复几次大概就能得到比较拟真的模型了,炼制的时候还可以多选取一些真人图片作正则化,标就打“realistic”或“photorealistic”之类的,这样出来的丹应该会更接近真人效果
@SettingsNoProblem 好的,感谢大佬👍
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.

