Thanks!
Thanks for continuing to share these.
It is very interesting to me to compare the face consistency to my own efforts over the last couple months, and even see the extent of other effects. For example, the reproduction of training set clothing! (And probably setting?)
One thing I have really been wondering about is what your training set looks like. I have found that even small differences in the training set (for example, a small change in distance) have a very large impact. I know that's not the kind of thing people would necessarily want to share so maybe we need some way to describe it, or convert the images into placeholder faces or something. For me, training images at a distance equivalent to image_3.png here are too close and cause premature burn. Even image_2.png is on the edge.
I am impressed at the consistency of the embedding version without noticeable burn (except for with "<s0><s1>", as expected) and I assume that is down to the training set choices (and even training set lighting), but your results sure are different than mine!