Per-render drift: the real reason AI ads break brand consistency

Three statics for our own brand, generated separately

Three worlds, one language

Bedroom, lamplight, cool greenTravertine, hard sun, warm stoneBedside, candlelight, near dark

A bedroom, a travertine ledge in hard sun, and a bedside table by candlelight. Nothing about the subject or the world repeats. The wordmark sits in the same top-left corner at the same size in all three, and the display serif, the plate geometry and the grain repeat with it. None of those four came out of a generator.

01Why do AI-generated ads look inconsistent?

Because each render is drawn on its own, from scratch, with no knowledge of the last one. Everything you specified is honored approximately. Everything you did not specify is filled in from the model's prior, which is the average of an enormous amount of other people's advertising. On brand, to a generator, means near the middle of everything it has ever seen.

That is why the prompt is the wrong place to look. You can write a perfect prompt and get twelve outputs that share a subject and disagree about everything else: the warmth of the light, how contrasty the shadows are, how heavy the type sits, whether the edges are crisp or soft. Each of those is a decision. Nobody made them, so the sampler did.

02Your designer was the consistency. The guide never was.

Hand a good designer a brand guide and they fill the gaps with taste, and their taste is consistent because it belongs to one person. Hand the same guide to a sampler and the gaps get filled with the average. The guide was never the thing keeping your brand together. The designer was.

A guide describes a look. It does not decide a shot.

The line we put at the top of every brand system we write

So the work is deciding, in advance, what a guide leaves open. Not more adjectives. Values in a file that a build reads. Here is the shape of one.

GRADE ....... one file, applied last, identical on every build
GRAIN ....... one scanned plate, tiled, never scaled, re-seated
              per frame, overlaid at 0.10
EDGE ........ 1.6px gaussian feather on every cutout.
              no keyline, no torn edge, nothing else
EASING ...... smoothstep only. no overshoot, no spring, no bounce
SHADOW ...... an offset blurred copy of the element's own alpha,
              offset as a fraction of the element's width,
              never a fixed number of pixels
BANNED ...... chromatic aberration, halation, scanlines, tape,
              pins, staples, full-frame brand color fields

The shadow line is the one worth stealing. Offsets expressed as a fraction of the element's own width scale correctly when the element does. Offsets in fixed pixels do not, which is precisely why composited AI elements read as pasted on: a big object and a small object in the same frame end up wearing the same shadow.

The decided parts of one of our own statics

Tap the numbers

12345

Four of these five are fixed for the whole brand and were never generated. One of them, the world, is free to change on every asset. That split is what makes a set look like a family instead of a folder.

03What actually drifts when the prompt does not change?

Grade and white balance drift first, and they drift most. Then contrast and how deep the blacks sit. Then edge quality, texture and grain, which is the one nobody thinks to specify. Then type weight and the space around it. Then the small stuff that turns out to be identity: how heavy a shadow is, whether an object is centered or offset, how much air sits above a headline.

The fix is structural rather than verbal. Let the model make the subject and nothing else, then build everything around the subject in code, the same way every time. It is the same underlying problem as holding one person across six shots, moved up a level from a face to a whole brand.

04The color field that belonged to somebody else's brand

We spent weeks admiring a reference system in which a solid brand-color frame fires between shots, ported it into a client engine as craft, and had the owner catch it in the delivered films. In that brand nine products each own a signature color, so a field cutting to a color is naming a SKU; under any other brand it names nothing and reads as a slide transition, which is why every film in the set was recut to remove it and the engine now refuses that shot type with an error naming the reason.

The same disease in a second organ

The reverse failure is worse and harder to see. We once retired eight finished ads at once because they differed on every mechanical axis and matched on all six of the ones a viewer actually reads, and what makes an ad resist fatigue is that story at full length.

05How do you keep AI-generated ads on brand?

Decide the visual language once, write it as values rather than adjectives, and apply it in code after generation instead of asking for it in the prompt. The generator's job shrinks to producing subjects. Grade, grain, type, plate geometry, shadow rule and easing never touch a model at all, which is why they cannot drift.

06Enforce it at creation time, not in QC

We audited two weeks of our own work against our own rulebook and the pattern was unambiguous. The rules that never broke were the ones enforced by a gate in the only path to the effect. The rules that broke repeatedly were the ones living as prose in a document that people were expected to remember.

In practice: text never escapes the plate it belongs to, because the box fits the words or the plate grows, decided before render rather than found after. Text escaping its box stops being a review note and becomes a build failure. That is the argument of our piece on deterministic pipelines plus AI generation, and it is the least glamorous thing we do.

07What this is worth, honestly

A system like this is a hygiene floor, not a growth engine. CreativeX put the nearest measurement on it: about ten percent more Creative Quality Score buys around two percent off CPM across roughly 822,000 observations. Real and small. We clear the floor every time, at speed. We will not tell you the grain plate is why the ad won.

What the system genuinely buys is cadence. Motion's benchmark set, 550,000 ads across more than 6,000 advertisers, found top accounts and average accounts running the same rough hit rate, with the top large account producing 5.99 winners a month against 1.75 on identical spend, purely by shipping around 31 new creatives a week instead of 11. Nobody sustains 31 a week if every one of them needs a person to protect the brand by hand.

What it costs you

One honest limit. A decided language is a constraint, and constraints cost you shots. We have lost frames we liked because a rule refused them, and kept the rule anyway. That trade is also why our work for two brands in one category should never look like it came from the same director, which is the subject of one visual language per brand.

Most brand inconsistency in generated work was caused by decisions nobody had made, on a document that read as though they had.