While the results are stunning for closeup face portraits, this embedding tends to force all generated images into that direction. Even (closeup, facial portrait, profile picture) in the negative prompt is mostly ignored and only helps sometimes, same with (full body shot, standing upright) etc. in the regular prompt.

What's interesting about that is that you can see that strong forcing during the generation in the preview, the noise vaguely looks like the torso and quite like what you're describing in your prompt, but then about halfway through the steps - BAM! - here comes the closeup face. I tried grabbing the noise before that point and using it as a depth controlnet, but even then, in most cases it just switches to the - then sometimes distorted - face abruptly at one point.

Any chance this could be refined further with an upcoming update?

I don't see any difference using epiCRealism as positive embedding. It produces the same images.