Keeping the Same Product, Person and Scene Consistent Across a Whole Campaign of AI Images
The first image is great. The founder in the office, the product on the desk, the light coming in from the left, the brand blue on the wall. The second is close. By the third, the product’s cap is a different shape, the founder’s jaw has changed, the office has gained a plant and lost a window, and the blue has gone teal. Image six looks like a different company’s campaign shot in a different building with a lookalike.
This is the campaign consistency problem, and it is different from the single-image brand consistency problem that reference images now solve reasonably well. One image can be held to a reference. Ten images, across four channels, with two rounds of edits each, generated over three days by a person who is also doing everything else, cannot be held by prompting. This post explains why drift happens and compounds, then gives the two artifacts that fix it: a continuity sheet and a drift budget.
Every generation is a fresh draw
An image model does not remember the last image. Each generation reconstructs the subject, the setting and the light from learned patterns plus whatever you gave it as reference. The product in image one and the product in image two are two independent drawings that happen to be aimed at the same target. They will differ in the way two sketches by the same artist differ: recognizably the same thing, not the same drawing.
Three things make this worse in real campaigns.
Detail is where drift lives. The model resolves composition first and fine detail last. Large shapes hold; caps, seams, labels, fingers, jawlines and material textures wander. This is the same mechanism that makes logos unreliable, applied to everything small.
Edits re-render. Ask for one change and the model produces a new image that resembles the old one, not the old image with one change. Users on OpenAI’s developer forum documented this pattern in detail: objects, backgrounds and themes they had not mentioned changed with the edit. Two edit rounds on each of ten images is thirty fresh draws.
Chaining compounds it. The most natural workflow, feeding image three’s output in as the reference for image four so they match, is the worst one. Image four now inherits every error in image three, plus its own. By image eight the campaign has drifted from the original the way a photocopy of a photocopy drifts. Google’s own developer forum has threads from users doing exactly this with product photos and watching the logo and product change regardless of the prompt.
Vendors have named the phenomenon. Product-photo tools call it visual drift or SKU drift, and one of them put the scale of it well: at fifty variants, the drift shows up dozens of times before anyone catches it.
Consistency is a reference and asset problem
Prompting cannot fix drift because the prompt is words and drift is pixels. “The same product as before” means nothing to a model that does not have a before. What it can hold is a file: an image of the product, from this angle, in this light. The fix is to stop describing the constant elements and start supplying them, the same files, every time, from a fixed source that never changes.
Film sets solved this problem a century ago with a continuity binder: photographs of every costume, prop and set dressing, so a scene shot on Tuesday matches the one shot on Friday. The same idea applied to a generator is the first artifact.
The continuity sheet
Before image one, lock four things as files, not as prompt words.
- The hero object, from two or three angles. The product, the packaging, the device, whatever must be identical in every image. Photographed or generated once, approved, and never regenerated for the rest of the campaign. If there is no real photo, generate the object alone on a plain background first, pick the one you can live with, and that becomes the reference for everything else.
- The person, if there is one. One approved image of the founder, the customer, the technician. Same rule: generated or photographed once, then used as a reference, never redrawn from a description.
- The environment. The office, the van, the kitchen, the site. One establishing image that fixes the architecture, the furniture and the wall color.
- The lighting and color recipe. A short written spec plus a swatch: light from camera left, warm, late afternoon; the palette with hex values; the film look if there is one. This is the one item that is words, because light and color are things a model follows well from text when the text is specific.
Then two rules. Every image is generated from the sheet, never from the previous image. Image seven’s references are the same four files image one used. Drift cannot compound because nothing inherits from an output. Each channel variant inherits a defined subset. The email hero gets the object and the environment. The LinkedIn tile gets the person and the object. The blog cover gets the environment and the light. Deciding this up front stops the “I just need a quick square version” edit that turns into a fresh draw of everything.
Where the logo goes. Not on the sheet. The logo is never generated; it is composited from the real file after generation, on every image, in the same corner at the same relative size. The logo post explains why this is the only method that works, and the placement guide has the numbers.
The drift budget
The continuity sheet controls generation. The drift budget controls review. It is a list, written before the campaign, of what may vary and what never may.
| May vary | May never vary |
|---|---|
| Crop and aspect ratio per channel | The product’s shape, proportions, color, label |
| The person’s pose, expression, clothing details | The person’s identity (face, build, hair) |
| Background props, time of day, weather | The environment’s architecture and palette |
| Camera angle within a range | The logo (composited, never drawn) |
| Supporting objects | Brand colors (locked as hex, applied in the composite step) |
Two things happen when this exists. The review step turns from “does this feel off” into a check against a list, which is faster and catches more. And the generator’s reference settings have something to protect: the right-hand column gets high-fidelity references and strong weights; the left-hand column is left to the model. Without the budget, people either lock everything (and get ten near-identical images) or lock nothing (and get the campaign from the opening paragraph).
Edits operate on the original
The last leak is editing. A campaign image gets a round of changes: reframe for a different ratio, remove a stray object, brighten the corner. Each edit re-renders, and each re-render drifts, and if the edit is run on an already-edited image the drift compounds exactly like chaining.
The rule: every edit runs on the clean, approved original from the sheet-based generation, never on a previous edit’s output, and the logo is applied after the edit. Keep the originals somewhere the whole team can find, named so nobody edits the edited version by mistake. This is tedious in a folder and trivial in a system that keeps the original linked to every derivative.
Marqeable’s image workflow is built around these rules: reference images live with your brand look and are attached to every generation in the campaign, so nothing derives from a previous output; brand colors, fonts and the logo are applied after generation as locked assets; AI edits (reframe, inpaint, remix, upscale) operate on the stored original and the logo is re-applied to the result; and the review gate sits before anything ships. The continuity sheet and the drift budget are the workflow; the tool just stops you from breaking them at 6pm.
Frequently asked questions
Why do my AI images stop matching after a few generations?
Each generation is a fresh reconstruction, not an edit of the last, so small differences accumulate in the details. Chaining, where one output becomes the next reference, compounds the drift because each image inherits the previous one’s errors.
What is a continuity sheet?
Four things locked as files before image one: the hero object from two or three angles, the person, the environment, and a lighting and color recipe. Every image derives from the sheet, never from a previous image, and each channel variant inherits a defined subset.
What is a drift budget?
A written list of what may vary between images (crop, pose, props, time of day) and what never may (product, person’s identity, environment, logo, palette). It makes review a checklist and tells reference weights what to protect.
Do reference images in GPT Image 2.5 or Nano Banana Pro solve this?
They hold a subject within a run and are the right tool for a product or a person. They do not manage a campaign: they will not stop chaining, lock a palette, or carry a logo. Campaign consistency comes from the workflow around the references.
The bottom line
Drift is not a prompting failure; it is what generation is. Every image is a fresh draw, every edit is a fresh draw, and chaining from outputs makes each draw inherit the last one’s errors. Lock the constants as files on a continuity sheet, generate every image from the sheet, write a drift budget so review checks a list instead of a feeling, edit only the originals, and composite the logo last. Ten images that look like one campaign is a workflow, and it holds up at 6pm on the third day only if the tool enforces it for you.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
