How to Keep the Same Face Across AI Video Scenes
Identity drifts between shots because each shot is generated independently — nothing carries your face from one to the next except the references you supply. The controls that hold identity across a multi-scene video, and the point where none of them do.
Generate every scene from one frozen reference set, keep each shot under about eight seconds, and never feed one shot's output in as the next shot's input. Faces drift between scenes because the model is not forgetting anything — there is nothing to forget. Most video models treat each generation as an independent sample conditioned only on what you passed in. Two scenes are two separate draws, and everything below is about making them land in the same place.
Why scene-to-scene drift is a different problem
Two mechanisms get called "drift", and they need different fixes.
Inside a clip, identity degrades over time as duration multiplied by motion amplitude. That is why a fifteen-second dance drifts where a fifteen-second talking head often does not.
Between clips, nothing degrades — each one is a fresh sample. Two scenes drift apart when the conditioning differs: a different prompt, different described lighting, a different angle, a different subset of reference photos. Any of those draws from a different part of the distribution, and it lands as a slightly different person.
The controls that actually hold
1. One reference set, reused unchanged. Not "similar photos" — the same files, in the same order, for every scene. Supply the maximum the model accepts: reference-driven models take up to nine, and most people supply one frontal selfie. Frontal, both three-quarters and a profile, same lighting, whole head in frame, no sunglasses. If wardrobe has to match across scenes, shoot the references in that outfit — the model reads clothing from them too.
2. Character binding, where the model has it. On Kling 3.0 Motion Control, character_orientation must be video for the element that binds your face to the character to do anything at all. On image the binding is a silent no-op: the request succeeds, you are billed, the identity drifts. This single setting accounts for more "the model drifts faces" reports than every other cause.
3. Constants written verbatim in every prompt. Hair, wardrobe, lighting, lens distance — copy the identical clause into each scene's prompt rather than paraphrasing it. Paraphrase is a change in conditioning.
4. A shot-length budget. Five to eight seconds holds reliably; twelve to fifteen is where most models start sliding. A long beat is two shots with a cut, not one long generation. The cut is invisible; the drift is not.
5. A face pass at the end, if you need it. An identity pass over the finished cut can rescue appearance, but it cannot fix a body that moved wrong, and it softens texture. A repair, not a plan.
Where the ceiling actually is
No current consumer pipeline guarantees one recognisable person across a dozen shots. Three to five is the range where the controls above hold without babysitting; beyond that you are re-rolling scenes until they match, and each re-roll costs.
Extreme profiles, heavy occlusion and fast turns stay hard at any shot count — they push the face through states the model has to invent. Re-rolling an unchanged request never helps: a structural cause drifts the same way for the same price. Change an input, then re-run. The drift diagnostic covers the single-clip version in order.
What it costs to hold identity
On Kling Motion Control, Pro at $0.168 per second held identity visibly better than Standard at $0.126 in our comparisons — checked 10 August 2026, about a third of a dollar on an eight-second shot. HappyHorse 1.1 reference-to-video sits at $0.14 per second and takes up to nine character references, which is the cheapest identity insurance here.
The credible models sit in a narrow $0.12–0.17 band — see the model catalogue. Switching model is a small lever next to the reference set and the shot length, and those two are free. If you would rather not chain scenes yourself, our recreate a Reel flow reuses the reference set across shots for you.