dio:dioxth
kujo:joxsp,stand \(jojo\)
weight:0.8-1
通过分层缓解了画风过拟合的问题
The problem of overfitting of painting style is alleviated through lora block
Description
dio:dioxth
kujo:joxsp,stand \(jojo\)
weight:0.8-1
通过分层缓解了画风过拟合的问题
The problem of overfitting of painting style is alleviated through lora block
FAQ
Comments (11)
hi, this is something amazing LORA, if you please, can you explain or tell me about the "B15" model that you are using? or where can i get it? because I've been looking for days about the model reference, and the results are nil, thank you
https://civitai.com/models/13034/lucky-strike-mix,bl5 is a self-use hybrid model, if you want to get a similar effect, you can try the previous model
看到例图已经不是单纯的人物Lora了阿喂！已经是某种特殊的替身Lora了hhh
单人也能摆出jojo立那feel了
绝了，这个lora看起来不但能还原人物还能很好兼容其他lora，真是太厉害了
emmmmmm，其实不太好用，替换人物我是用了分层，可以看我图片信息，挺麻烦的……不过你这么一说我倒是可以直接改一下lora改到大家都方便替换……
@kozue 哈哈哈，抱歉我有点激动了，主要是因为我自己练lora一直有问题...小体量的对细节还原力不佳，大体量的容易过拟合或产生某些“偏见”，您的lora我使用感很好，而且本身就喜欢半现实画风。希望我没有在无意之间捧杀您
@_Qing_ 不不谢谢谢谢，因为某些原因我一直没更新训练包，使用的最老版本，很多新功能都用不了，导致一种处于一种努力尝试用螺丝刀造核弹的状态……
@kozue 哇，真的厉害了！听起来就很辛苦...或许您可以试试bilibili上up说的云端训练lora？就是工作日期间比较难抢到机位...
@_Qing_ 没事 orz啥时候电脑出问题了强制更新再说吧……
least confusing lora