Updated link in the description while I look for a more reliable file hosting.

Sometimes you'll need to crop/erase/edit your image a bit to help Wan out. https://www.youtube.com/watch?v=S8wKlaoNgc0

If your image contains multiple people, and you only want the action to happen to one of the person in the image, it'll be easier if you crop everyone else out or erase them. That way Wan doesn't have to guess which person you're refering to.

If you want a certain type of action, it'll be easier if that person is already close enough to that action. For example, bj works best when the image/person/female is crop chest up, because Wan can easily just place the man above the female's head. If the image is of the female standing head to toe. Now, Wan has to figure out how to get the man above the female's head when the female is standing tall. If there is like a counter/table near where the woman is standing, Wan might just have the man jump onto the table and stand on top of the table to get the correct height for the action. For the most part, Wan will just get lazy if Wan has to figure out too much to make things happen, and just do what it did in your case, where Wan just have it stick out from the man's nose.

For reverse, the woman's head will have to be tilted back. Wan won't tilt the woman's head back. You'll have to "help" Wan place the woman in position. If the woman is already laying down, all you need to do is specify the direction the man should be from and Wan will take it from there. If the woman is standing up, you'll have to specify the woman laying down or tilting back, or specify the man coming from a high angle from above and behind the woman.