Why your AI character's face keeps changing between shots

Why does this happen at all?

A video model generates frames from noise, guided by your prompt. When you generate shot four, it has no idea what shot three looked like. There's no persistent character stored anywhere — only the constraints you provide each time.

So the question isn't "why did it forget." It never knew. The question is how tightly you constrained the guess.

The more you constrain, the less it invents. That's the whole principle, and everything below follows from it.

Which shots break identity worst?

Not all shots are equally hard. In rough order of difficulty:

Profile. The model has to infer a side view it may never have been shown. The single most common failure point.

Backlit. Facial detail falls into shadow, so there's less signal to anchor to.

Turning head. Mid-turn frames are the hardest to keep consistent, because the face is in a position no reference image shows.

Extreme angles. Low angle, high angle — same problem as profile.

Front-on close-ups, by contrast, are usually reliable. Which is why creators often think consistency is solved right up until they cut to a reverse.

What actually fixes it?

Three approaches, in ascending order of effort and reliability.

Reference locking

Hand the model an image at generation time. Immediate, no setup, works everywhere.

Where it falls short: less reliable across high shot counts and extreme angles. Fine for a single clip, strained across a twelve-shot scene.

Reference propagation

Generate your first clip from the original reference, export a clean frame from it, and use that frame as the reference for the next shot. Repeat down the scene.

Free, needs no extra tooling, and noticeably better than a single static reference.

Where it falls short: errors compound. If shot three drifts slightly, shot four inherits that drift and adds its own.

Trained identity

Build a persistent model of the character first — typically from twenty or more photos across different angles — then apply it to everything after. Setup takes a few minutes.

This holds substantially better on high shot counts and extreme angles, which is exactly where the other two fail. For a recurring lead across a series, it's the only approach that really works.

Where it falls short: the setup. For a one-off clip, it's a bad trade.

Does it help to specify the face in the prompt?

Somewhat, and there's a specific mistake worth avoiding.

Write the identity description once and paste it into every prompt unchanged. Never paraphrase it. "Dark wavy hair" and "black wavy hair" are different constraints to a model, and rewriting the description each time is a common self-inflicted source of drift.

A worked example:

MEERA — 32, Mumbai. Shoulder-length black hair, centre part, tucked behind the right ear. Oval face, sharp jawline, small mole below the left eye. Medium-brown skin. Straight eyebrows. No jewellery except thin gold studs.

Specific, unambiguous, and identical every single time.

What about two characters in one shot?

This is the hard case, and worth being honest about: two locked characters in the same frame degrade on every platform. Features bleed between them — one face borrows the other's bone structure.

The model isn't maintaining two identities. It's producing one image that has to satisfy two constraints simultaneously, and they interfere.

The practical workaround is avoidance. Short drama doesn't need many true two-shots — over-shoulders and clean singles cut together perfectly well, and each of those has only one identity to hold.