I asked for no seatbelt. The model gave me two.
How to make AI clips that look real
Realism is not a quality setting — it is a set of constraints. Static phone camera, deliberately imperfect framing, the face carried by reference instead of description, dialogue written out beat by beat, and never telling the model what NOT to show.
Never describe the face — reference it
The face comes from the reference. The prompt describes only the outfit.
Reading locks nothing in. The step this lesson teaches takes minutes in the Studio.
Generate your first clipWhy does my clip look 'too good' and feel fake?
Because the model defaults to cinematic: shallow depth of field, smooth camera moves, perfect framing. Real phone video has none of that. Ask explicitly for the imperfect version.
Can I just add 'no watermark, no blur, no artifacts' to be safe?
No — negations backfire. The model tends to render the nouns it hears, so name only what you want in the scene, never what you want absent.
How long should the clip be relative to the voice line?
Voice duration plus one second. Exact-length clips clip the final word; the extra second lets the take breathe and end naturally.