How to keep the same face, and the same product, in every shot

Why do AI characters change between shots?

Because most tools treat every image as the first image. There is no standing idea of who your character is — only whatever the description happened to say this time.

Change the lighting words and the jawline moves. Change the angle and the jacket invents a pocket. Nothing is wrong with any single frame. They are simply not the same person, and a viewer notices that long before they could tell you what changed.

Is character consistency the same problem as product consistency?

It is the same problem wearing two hats, and the product one is worse because it is quieter. A face going wrong is obvious to everyone in the room. A label going wrong is obvious only to you and to your legal team.

Across a few seconds of motion a logo can creep a degree at a time and no single frame ever looks bad. That is precisely how it reaches a client: not by being obviously broken, but by never being obviously broken enough for anyone to stop.

What makes a good reference?

Range beats resolution, every time. One glamorous hero shot tells a system almost nothing — it is one lighting setup, one angle, one expression. Five ordinary frames from different angles in different light tell it everything.

Include the details you would notice if they went missing. The buckle. The pocket square. The exact curve of the label where it wraps. The way the sleeve sits. If you would flag it in a review, put it in the reference — and if it is not in the reference, do not be surprised when it is not in the work.

How many references is enough?

For a person, three to five frames covering front, three-quarter and profile, in at least two different lighting conditions. For a product, add whatever angle shows the thing that identifies it — usually the label, sometimes the silhouette.

Beyond that you get diminishing returns fast. What helps is variety, not volume: ten near-identical shots from the same shoot are worth about as much as one.

What about references that are not photographs?

They count and most people never think to attach them. A brand guideline document holds the rules and the exact colours. A voice clip holds a voice. A previous ad holds everything you already agreed was correct.

Attach them. A document is read for its rules rather than summarised, which is the difference between a system knowing your palette and a system having been told you have one.

How does KIT ENGINE keep a cast consistent?

You build a cast once. A character, a product, a prop, a location, a brand, a voice — each gets a name, the references you upload, and an approval from a person before it can ever be used.

Then you name it in the brief. The same amber bottle, the same founder, the same studio, shot after shot, week after week, without you re-describing any of it. The description stops being something you write and becomes something the workspace holds.

Why does approving a likeness need a person?

Because generating somebody's likeness is a decision, not a setting. It should not happen as a side effect of typing a name into a box.

Approval is a deliberate act, it records who made it, and until it exists the work will not run. We would rather refuse a job you wanted than deliver one you cannot legally use — and a picture of somebody who did not consent to be there is not a quality problem, it is a different kind of problem entirely.

Can you pin an exact version of a reference?

You can, and you should when it matters. A brand that re-shoots its product photography would otherwise silently change what every earlier brief would now produce.

Pinning a version means a campaign made in March still makes March's product in September. If you ask for a specific version and it is not approved, you are told — never quietly given a different one. Somebody who asked for version three and got version seven has been handed the exact thing they said no to.

What happens when a shot drifts anyway?

It is flagged before you see it, not quietly patched. This is a deliberate choice and it costs us something.

A quiet fix is a second opinion nobody asked for. The label comes back legible and slightly wrong. The character comes back consistent and slightly generic. Pass rates go up and the work gets worse, which is the exact trade this was built to avoid. So it stops, puts the frame in front of you, and you decide.

Does consistency limit what you can make?

No, and the assumption behind that question is worth naming. Consistency is about identity, not about repetition. The same person can be in a kitchen, a warehouse and a studio; the same bottle can be lit six different ways.

What is fixed is who and what. Everything else — light, framing, mood, motion, format — is still entirely yours to move.

What should you do first?

Take the one product or one person that appears in most of your work and build that first. Give it good references, get it approved, and use it for a week.

You will feel the difference immediately, because that single element is the thing you have been re-describing more than anything else. Everything after that is the same move, repeated.

What are the most common reference mistakes?

Uploading only the hero shot. Uploading ten frames from one moment of one shoot. Uploading images that have already been retouched so heavily that the real object is gone.

And the quiet one: uploading a reference that contains something you do not want carried forward — a competitor's product on the shelf behind, a seasonal backdrop, another person's arm. References are read closely, which cuts both ways.

Crop to what you mean.

How should you name a cast?

Plainly and consistently, the way you would name a folder somebody else has to find in a year.

'Amber pouch 200ml' beats 'Product 1'. 'Dana, founder' beats 'Woman A'. You will be typing these into briefs, other people on your team will be guessing at them, and a name that describes the thing is the difference between binding the right one and binding the nearest one.

Names are unique within a workspace on purpose, so there is never a question of which pouch you meant.

What about a character who changes on purpose?

Different outfits, ages or settings for the same person are directing decisions and belong in the brief. The identity stays in the cast; the circumstances go in the sentence.

If a character changes so completely that they are functionally a different person — a mascot in two eras, say — make them two cast members. Trying to hold two identities under one name is how both get blurred into an average.

How does consistency work across formats?

The same cast is bound whether the output is a still, a video or a cutdown, which is the point of holding identity in one place rather than in each tool.

This is where multi-tool workflows break hardest. The stills tool and the video tool each have their own idea of your product, and no amount of careful prompting reconciles two systems that have never seen the same reference.