The consistency problem is the whole problem

Anyone who has used a video model has had the same experience. The first generation is astonishing. The second, from a nearly identical prompt, is astonishing in a completely different way — different face, different lens, different world.

For a single social post that is fine. For an advertisement it is fatal, because an advertisement is not one shot. It is four or five shots that have to read as one piece of work.

What actually breaks

Three things drift, and they drift independently.

Identity. The person in shot two is not the person in shot one. This is the most obvious failure and the one viewers notice first.

Object continuity. The product changes. Hardware moves, proportions shift, a strap becomes a handle. Viewers may not consciously notice, but the film stops feeling like it is about a real thing you can buy.

Grammar. Shot one is a slow dolly on a long lens. Shot two is a handheld wide with a different colour temperature. Individually fine; together, incoherent.

Why more prompt does not help

The instinct is to write more. Add the hair colour, the age, the jacket, the lighting.

This fails for a structural reason: prompts describe categories, and you need an instance. “A woman in her thirties with dark hair” has millions of valid answers, and the model picks a new one each time. Piling on adjectives narrows the category without ever selecting a member of it. Past a point, the extra description competes with the parts of the prompt that describe the shot, and the framing gets worse while the identity stays random.

What does work

Three things, in order of impact.

Send images, not names

Modern video models accept reference images. This is the difference between describing someone and showing the model who they are. A registered character means every generation featuring them carries the same approved photographs — not a re-description, the actual images.

The written description still matters, but its job changes: it keeps the prompt coherent alongside the references rather than doing the identification itself.

Fix the grammar centrally

Shot-level consistency comes from the reference system. Film-level consistency comes from creative direction that lives above the individual scene: one statement of style, lighting and pacing applied to every prompt the system writes.

When you say “make everything more premium”, that should change one thing that then propagates — not five prompts you edit by hand and inevitably make inconsistent.

Carry forward only what was established

Naive continuity does more harm than good. If you assert “same lighting as the previous shot” when the previous shot never specified lighting, you have told the model to match something that does not exist, and it will invent one.

Continuity should be conditional: forward the location when a location was set, the wardrobe when wardrobe was set, the time of day when it was established. Silence should stay silence.

The cost angle

There is a practical reason this matters beyond craft.

If shot three is wrong, you should be able to regenerate shot three. That requires each scene to be an independent unit with its own references, its own prompt and its own version history — not a frame range inside a single monolithic generation.

Get that right and a failed shot costs you one shot. Get it wrong and every fix costs a full re-render, which is where AI video budgets actually go.

The short version

The models are good enough. The gap between a striking clip and a usable advertisement is not model quality — it is whether the system around the model remembers who your character is, what your product looks like, and what the last shot established.

That is a software problem, and it is a solvable one.