Let’s have a collection of wedding dresses. Everything can be colorful. I found a lot of super beautiful wedding dresses with a palace style while lying on the bed today. How can I let it go? I must make it lora and benefit everyone.
All comments about the current lora are under the description of this lora on the right. There is no need to make random guesses about lora's recommendation weight.
I found that some friends did not read the comments at all, and the weight was 0.5-0.7. Then they couldn't get the costumes, and they started to complain and even give low scores. It was so helpless. Please look at my test results.Description
Florence Dream
Recommended weight is 0.9
FAQ
Comments (13)
更新:感觉脸太暗是增加细节(Add_Detail)之类的Lora的锅,不知道如果用detail_slider的负权重(那个lora的权重可以是-5到+5)能不能改善(体感去掉Add_Detail好像改善了一点点脸太暗的问题,但是没时间过多测试)
的确不把权重拉满到0.85-0.9是出不了模型的效果的,但是加入同类别的提示词(例如这里我加了“beautiful bridal dress“),就可以把权重降到0.6了,甚至还能更低到0.5出一些融合的图(不直接拉到0.9的原因是,好像大部分衣服模型都会影响我想要的脸型,所以被迫用低一些的权重,我是新手菜鸟不懂其他的方法)
降低了lora权重,应该给其他提示词增加权重麽
@ruanyi 我昨天在这最早发的那一组图(现在已经删了,怕误导别人)lora用的0.7的权重,自然出不来任何婚纱的特征(我也不知道为什么是红色的胸罩+红内裤)我试着用0.8或者以上,的确能出婚纱,但是炼出来的脸就变了。。。。后来我加了一个模糊的提示词“beautiful bridal dress“(就这一个),忽然就能画出这个lora的特征婚纱了,一直到0.5的权重都能保留一部分的特征。。。。(0257 castle style这个模型最明显,甚至可以0.5权重出图)
@ruanyi 长话短说,没改其他权重,就只把其他所有的关于衣服的提示词删了(之前我用了lingerie的提示词)替换成和lora同类的(这里是bridal dress)忽然就好用了
@schenad4 哈,这些婚纱没有过多的做泛化测试。只试了几个常用的,然后提高了一些学习率,但是lingerie换掉dress是不是过于离谱了。哈哈
@ruanyi 也可能是巧合。。。。
Having worked with wedding dresses for sixteen years, I think I am competent to write something about it.
A very nice selection of wedding dresses.
Maybe pay attention to the simple princess cut with laces on the back.
与某些画风lora、VAE、二次元模型使用时会导致生成的图片过暗,是因为训练底模的原因吗
确实比较暗,应该是训练集很多是大光比的素材,有些地方很暗,lora把这个画风也学进去了
May I ask if the illustrious version can be made?

