The subquery returns all potential media id's. Typically, we have a couple of faces and thousands of media files.
A join is much faster.
With about 50k images and a face that was present in 2'000 images the query was unusable slow, it took about 60s.
This is a very important performance fix, as I think many users will run into it.
Maybe there are other areas where the same improvement is possible, I didn't check.