I tested aesthetic gradients on top of an embed using the same training set
0.0001 learning rate, slerp interpolation on, no text, slerp angle 0.1
Stuff like hair detail gets destroyed almost immediately, but I had some nice results on another embed that doesn't have very intricate identifying features.
Full size 55MB png here:
https://litter.catbox.moe/nsb5vn.png