it's just a bad attempt at making it face-agnostic, a technique for older models like SDXL. It's easier to either wholly cut the face, or better, leave the face, but prompt it in such a way that the face is not the focus of the learning. These krea2 and qwen encoders are capable of understanding the input image well enough to ignore the face. generally things like this is someone trying to do good, just putting that effort in the wrong place for the tool. It's lame to see people be upset with that.