Curious how you handle the aspect-ratio mismatch when the reference is a phone clip. Do you crop first, or let the model re-frame and accept the drift?
We crop first when the subject already fills roughly 60% of the frame — a straight crop keeps framing predictable. Below that we let the model re-frame, because cropping a wide shot into 9:16 tends to slice limbs or on-screen text at the edges. The drift is real though, so we constrain subject position in the prompt and re-roll rather than try to fix it in post.