为什么25年了，换脸软件还是2021的deepfacelab？

whatsall

Cxsmo 发表于 2025-5-15 23:17
I heard somewhere he stopped because dfldata.cc creator stopped funding him to make it

dfldata.cc creator? 那不就是滚石么？你是开玩笑的对么？滚石如果是资助者的话，他干嘛整天转发一些其它竞争对手的换脸工具啊？

Cxsmo

whatsall Posted at 2025-5-17 14:20
dfldata.cc creator? Isn't that the Rolling Stones? You're kidding, right? If Rolling Stone is a funder, why would he be retweeting some other ...

because he stopped funding it a while ago from what I was told

Cxsmo

htyy 发表于 2025-5-17 04:43
也就ice在更了吧，虽然也是一年多前的版本了，效果真不赖，我玩不透而已
...

ice is not as good nowadays when there main thing which was higher res xseg model's other forks have and also, I have an obstruction pack trainer for xseg training that uses a pack and non obstruction images and trains it randomly for obstruction and also xseg big models that retains more of the trained data but not releasing it public unless someone adapts it to my fork DeepAscension

幽雅猫猫

主要还是dfl的配置要求相对友好，操作简单，虽然上限没sd高，但是只要掌握好训练方法，还是很像的，效率也高，完全碾压现阶段的所有类似的软件！楼主想玩还是dfl吧！另外我怀疑dfl的原作者是不是去打仗啦！

Cxsmo

幽雅猫猫 发表于 2025-5-18 05:15
主要还是dfl的配置要求相对友好，操作简单，虽然上限没sd高，但是只要掌握好训练方法，还是很像的，效率也 ...

SD can't do as good live or videos swap with the efficiency of dflab or live

lordor

应用场景的问题，现在都是放live里才用这玩意儿

bigmac

fengjunwei 发表于 2025-5-16 00:54
训练人物视频只用hunyuan，甚至win2.1都不用，因为hunyuan的lora是多注意力模型，SDXL或者FLUX这些更差的 ...

请教一下这个多注意力的意思？ 我的脑子还停留在SD对像素猜测的层面。就感觉现在的AI其实并不能理解一个三维物体在不同角度的投射准确度。比如一个人可能1米8，但出的图可能只有1米6。