Ask any AI video model for the same person twice and you will get two cousins. Same rough age, same hair, a slightly different nose. In a single image nobody notices. Across a 30 second ad, viewers feel it before they can name it, and the ad starts to look cheap.
Start with a model sheet
Before we generate a single shot, we build the character: a front view, two three-quarters, both profiles and the back, all in the same light and the same clothes. That 360° sheet becomes the reference for every generation that follows. It is the same idea animators have used for a century, pointed at a new kind of camera.
Lock the world, not just the face
Faces drift less when everything around them is fixed. We write the location, lens, time of day and wardrobe into a short bible and reuse it word for word. When a shot comes back wrong we change one variable at a time, never the whole prompt.
Fix it in the edit
Even with good references, some takes slip. That is where years of compositing earn their keep: a frame-level face fix, a re-timed cut, a grade that pulls every shot onto one palette. AI produces the footage. The edit makes it one film.