Brand Identity in AI Ads: Scene Bibles, Outfit Locks, Product Fidelity

Brand consistency in AI ads comes from production discipline rather than a model setting. The method that works is a Scene Bible: before you generate anything, lock one location, one outfit, and a short list of look constants, then feed real product reference images into every shot so the model never gets to imagine the product. Skip this and you get eight shots that are individually fine and collectively incoherent.

Generative video models have no memory between shots. Every generation starts from zero: a new kitchen, a new sweater, a new color grade, and, most dangerously, a new version of your product. Each shot passes review on its own. The failure appears when you cut them together and the ad reads like stock footage from six different libraries.

We learned to describe the failure mode with one phrase: “four drinks on four different tables”. You ask for a six-shot ad about one beverage, and you get six technically competent shots in which the cup changes shape, the lid changes color, the table changes material, and the daylight changes season. No single shot is wrong. The set is.

A 30–40 second ad runs 6–12 shots in our pipeline, which gives the model 6–12 independent chances to reinvent your world. Better prompts shot by shot will not produce consistency. You impose it from above, through constants that every shot inherits.

A Scene Bible is a short document, often under a page, that fixes everything the campaign may not vary. Ours has three sections: one location, one outfit, and the look constants. Every prompt for every keyframe carries these constants verbatim. You write the Scene Bible once, before the first generation, and you do not edit it mid-campaign.

One location

Pick a single, specifically described setting and keep the whole ad in it: “small sunlit café counter, pale oak surfaces, white tile behind the espresso machine” rather than “a café”. Specificity makes regeneration converge. When shot 4 fails QA and you regenerate it, a vague location description produces a different room every time, while a specific one produces variations of the same room.

One outfit lock

If a presenter or creator appears in more than one shot, lock their outfit to a written description with color, garment type, and one identifying detail, then ship that description in every prompt. Outfits drift faster than faces. A model will happily move your presenter from a black crewneck to a grey hoodie between shots 2 and 3, and viewers read that as a cut to a different person. Face-level identity is its own problem with its own mitigations, covered in how we keep AI characters consistent. The outfit lock is the cheap half of the fix, and teams skip it constantly.

Look constants

• Palette: 2–3 named colors that should dominate every frame, usually the brand colors plus one neutral.
• Lens feel: one phrase like “35mm handheld, shallow depth of field” repeated in every prompt. Mixing a phone-camera look with a cinema look across shots is one of the loudest continuity breaks.
• Light: time of day and quality (“soft morning window light”) locked once. Light direction changing between shots is the single most common giveaway that an ad was generated piecemeal.
• Energy: a one-line note on motion (“slow push-ins, no whip pans”) so the edit cuts together at one tempo.

1. 01Collect real product reference images first: clean shots of the actual product from 2–3 angles, with labels readable. Feed these into every generation, and treat them as required input rather than optional inspiration.
2. 02Write the location in one sentence with at least three concrete physical details (surfaces, fixtures, what is behind the subject).
3. 03Lock the outfit: garment, color, fit, one identifying detail. If you run a persistent brand character, this becomes part of its permanent definition, following the same logic behind virtual influencers.
4. 04Pick the look constants: palette, lens feel, light, motion energy. Four lines, no more.
5. 05Generate 3–4 test keyframes of different moments using the full constant block, side by side. If they look like frames from one film, the bible holds. If not, tighten the vaguest line and retest.
6. 06Freeze it. Paste the constant block into every shot prompt for the rest of the campaign, unedited.

This works because our process is keyframe-first: we generate still keyframes, grade them, and animate only the survivors via image-to-video. We enforce the Scene Bible at the keyframe stage, where a rejected image costs cents and seconds, rather than at the video stage, where a rejected clip costs a render queue slot and 1–5 minutes of waiting. Our AI ad production pipeline describes the full flow.

The product is the one element where “close enough” is a defect. Three rules cover most of it.

First: always generate from real product reference images. Reference-aware image models hold a product’s shape, color, and label well, but only when you give them the real thing to hold onto. A text description of your packaging invites the model to invent.

Second: never let the model imagine the product. If a shot does not need the product visible, fine. But the moment the product is in frame, the reference images go into the generation. There is no shot where “roughly our bottle” is acceptable.

Third: check labels and logos in QA, every time. Models love to redraw type. A label that reads correctly at a glance turns out on inspection to have a mangled letterform, a doubled word, or a logo that is 90% right, which is worse than 0% right because 90% right ships. Image-to-video adds a second layer of risk, since even a perfect keyframe label drifts over the course of a clip. That is one reason we keep identity-critical clips short at 4–8 seconds and re-anchor every shot on a fresh keyframe.

Because the model will do it, cheerfully and plausibly. Ask for “a customer enjoying a drink from the brand” without constraints and you may get a frappé with whipped cream, an iced matcha, or a smoothie in a branded cup, none of which the brand sells. The model pattern-matches to “beverage brand ad” rather than to your actual menu. The same applies to flavors, sizes, colorways, and bundle packs in any category.

The fix is explicit negative scope in the Scene Bible: list what the brand actually sells, name the hero SKU for the campaign, and state that no other products may appear. This sounds pedantic until the first time a client asks why the ad features a drink they have never made.

We grade every keyframe and clip against a checklist before it enters the edit, machine-graded first and human spot-checked after. We generate roughly 2–4 candidates per shot to keep one. Here is the checklist; copy it.

• Product matches the reference images: shape, proportions, cap/lid, material.
• Label and logo are legible and correct. Read the actual letters instead of glancing.
• No invented SKUs, flavors, or packaging variants anywhere in frame, including the background.
• Location matches the bible: same surfaces, same fixtures, same room.
• Outfit matches the lock: garment, color, identifying detail.
• Light direction and quality match the bible and the neighboring shots.
• Palette holds: brand colors present, no stray dominant hue.
• Lens feel consistent: same focal-length character and depth of field as the rest of the set.
• No artifacts: extra fingers, warped text in background signage, melted edges on the product.
• For video: identity and label hold for the full clip rather than only the first frame.

The cost stays invisible at the shot level, which is the trap. Each generation looks fine, so nothing flags it. The bill lands at assembly, when the edit refuses to read as one piece and the only honest fix is regenerating half the shots under constants you should have written on day one. We have watched the no-bible path roughly double generation volume on a campaign: the same 2–4 candidates per shot, run twice, once before the bible existed and once after.

A softer cost follows: trust. Viewers cannot articulate “the light direction flipped between shots”, but they feel it, and what they feel is cheapness. For AI UGC ads, where the format trades on reading as real and native, incoherence becomes a credibility problem rather than a style problem.

Before generating a single frame, write the bible: one location sentence with three concrete details, one outfit lock, four look constants, the real SKU list, and a folder of clean product reference photos. It takes under an hour, and that hour pays back more than any other in the production. Enforce it at the keyframe stage, where rejection is cheap, and run the continuity checklist on everything that survives.

Or hand it to a pipeline that already works this way. Scene Bibles, reference-fed product fidelity, and machine-graded continuity QA are built into how we produce AI UGC ads, so every campaign gets one world instead of eight shots from eight different ones.