It actually only does anything of significance to faces. Bodies, backgrounds, sky etc are mostly just blurred or unsharpened. So it’s not looking to recognize a tree then improve it with a much clearer tree. But if it finds a face it can make assumptions about the eyes and replace them with supposedly exactly the same eyes that are much sharper.

Unfortunately the combination of “improving” the eyes, nose, ears, chin, and hair results in something that really isn’t the same person. But to a stranger, it would look like an improvement as they’d have no baseline expectations on what the person actually looks like.