The brief
The growth team wanted to test both pitches on the same product, same time, without a second shoot or a second model. That was the brief I gave myself, run through Magnific Spaces, Seedream 5 Pro, and Seedance 2.5.
The risk
If the woman in spot one looked even slightly different from spot two, hair, face, proportions, the split test stops comparing two pitches and starts comparing two different women. The data comes back confused, and nobody knows why. One variable had to hold still while everything else changed.
The split
I shot spot one for daily wear, not kink wear: she ties her shoe, grabs her bag, walks out the door on an ordinary Tuesday, minimalist living room, natural light, handheld energy. For spot two I went the other way on purpose: same face, same product, but the camera orbits her in a dark crimson studio, no daylight anywhere. Her identity holds fixed. Everything around it moves.
What broke, and the fix
3 failure modes showed up across Seedream 5 Pro and Seedance 2.5: her body froze while her head kept moving, a mannequin with a good face; a prop vanished between cuts; and the location itself drifted, a wall that wasn't there last shot. Same fix each time: anchor it. State the weight transfer, the prop's position, and the location's details in every stage of the prompt, then restate it in the consistency lock. Say it once and it drifts back. Say it twice and it holds.
What held
Same reference face through a living room, a city street, a studio, hair and makeup changes, and it never drifted. That's the part I'd trust on a real slate.
Why it matters
20 years directing shoots, identity consistency wasn't mine to solve, it was solved automatically by the same person standing on set. This is the first time I've had to engineer it by hand. Same eye I had at Avon, pointed at a different kind of set.