LIAE的优点是融合度高，缺点是易受DST影响。但是不论DF还是LIAE，都没办法把与SRC差异太大的DST完美替换。也就是说，目标脸是长脸，你怎么也不可能换成圆脸。作者给出的相对解决方案是DF里的true face power选项，以及LIAE架构类型后面的T，这也是为什么官方教程的RTM（万用模型）训练法里给的示范架构是LIAE-UDT，这个T就相当于DF里的true face power打开。"The recommended type of models for use with RTM workklow are SAEHD LIAE models, LIAE-UD or LIAE-UDT thanks to their superior color and lighting matching capabilities as well as being able to adapt better to different face shapes than DF architecture."（摘自官方指南）
除此以外，官方指南还提到了解决脸型不像SRC的方案，基本上有两种（1）打开blur mask选项，训练background style（2）关闭masked training选项，我自己试过的结果是，第二种方法效果更好，但是耗时耗力。最完美的解决方案是用AE或者达芬奇来后期修整。“If you suspect some parts of DST may stick out from underneath your predicted (final face) during merging (long nose on profile shots, puckered lips, eyelashes) you can also enable background style power to make the background more like DST background and blur out mask to make it more blurry and smooth so that you can expand the mask during merging and esentially cover those parts of DST with the trained face background. You can additionally disable masked training after step 6, that will train entire training area, improving quality of the background but you can also try to run it before before GAN - keep in mind all of these are optional, if you are beginnger skip this step, pretend it doesn't exist.”（摘自官方指南）
|Archiver|手机版|deepfacelab中文网
|网站地图
楼主
发表于 2022-8-18 10:31:25