How to separate style and content in the reference image?

[Sutongtong233]

If one use a textual inversion to describe the style image, how to separate style and content in the reference image? I understand that "content" is part of the "style"(as the paper mentioned, semantics elements are part of style). However, in Fig.4, 2rd row, InST result is totally a robot as the reference image, while the content is a human. So I am confused about the evaluation: which parts of elements of the style image are we really want to transfer to content image...