How to Keep AI Imagery Brand-Consistent Across a Campaign

Everyone can spot inconsistent brand imagery, but few teams can say what is actually wrong with it. The images are the right subject, the right product, the right price of polish — and still they look like they came from five different companies. The cause is almost never the model. It is that four variables changed between generations and nobody wrote them down.

Consistency is a short list, applied every time

Brand recognition in imagery comes from a small number of highly visible signals. Lock these five and a set will read as one family even when the subjects differ completely:

• Palette — three to five approved colours, including how they appear in materials
• Light — direction, softness and temperature
• Framing — where subjects sit, how much space is left, how tight the crop is
• Treatment — how surfaces read: matte or glossy, grainy or clean, warm or clinical
• Restraint — what is deliberately absent

Notice that style adjectives are missing from the list. Style is the output of these decisions, not a separate instruction. When you write "cinematic" you are asking the model to guess which of the five you meant.

Build a reference set before writing rules

Collect six to ten images you already consider on-brand: real photography, previous approved generations, packaging shots, campaign work. Spread them out and write down what repeats. You will usually find that the palette is smaller than you thought, the light almost always comes from the same side, and the best images share one framing habit — a centred subject, or a consistent amount of negative space.

That written pattern is your rulecard. Derive it from evidence rather than aspiration, and it will survive contact with production.

One page, five lines, no jargon

A usable rulecard looks like this:

Palette: oat, off-white, slate grey, faded terracotta (accent only)

Light: soft directional daylight, always from the left, gentle shadow

Framing: subject centred or right of centre, generous negative space, nothing within 8% of the edge

Treatment: matte surfaces, visible material grain, no reflections

Never: text, logos, chrome, neon, plastic sheen, busy backgrounds

Five lines is not a simplification. It is the amount a team will actually apply under deadline. If a rule cannot be seen in a finished image, it does not belong on the page — move it into a longer internal document instead.

Review in sets, not one image at a time

Single-image review is how drift enters a campaign. Any one image can be defended; the problem only becomes visible in a row. Generate in batches, then review as a contact sheet of six to twelve candidates at thumbnail size.

At thumbnail size, three defects become obvious that a full-screen view hides:

1. Temperature drift — one image cooler or warmer than the rest
2. Light drift — shadows suddenly falling the other way
3. Density drift — one frame noticeably busier than its neighbours

Fix drift at the batch level. Approving nine individual images and hoping they match is what produces a folder of near-misses.

Diagnose drift by field, not by taste

When an image feels wrong, name the field before giving feedback. A short diagnosis keeps revisions productive:

• Palette: an unapproved colour is carrying too much area
• Light: direction or softness changed between shots
• Framing: the subject moved, or the negative space closed up
• Treatment: the material reads glossy instead of matte
• Restraint: decorative elements were added that no one asked for

This is the same review discipline described in the AI visual brief template. Criteria first, opinion second — it changes revisions from preference battles into corrections.

Turn the rules into a style pack

Hand-written rules still need to be re-entered into every prompt. A reusable style pack converts the rulecard into a saved set of constraints you can apply to any new subject. The palette, light behaviour, treatment and negative list travel with the pack, so a new product shot starts from the brand instead of from zero.

Packs also make consistency testable. Generate the same three test subjects with a pack: a product, a person and a textural detail. If the pack holds across all three, it will hold across a campaign.

Keep exceptions deliberate

Consistency rules should not be broken by accident, and they should not be so rigid that seasonal or campaign work cannot breathe. The workable middle ground is to decide, in advance, which variables may change:

• Always fixed: palette, light direction, treatment
• Sometimes varied: framing and background environment
• Fully open: subject, story, props and scene

Written that way, a bold campaign still reads as your brand because it moved inside known boundaries. Anyone reviewing it can tell the difference between a deliberate change and a mistake.

Check consistency across formats before publishing

The final test happens after cropping. A square social still, a wide hero and a vertical video frame from the same set should still look related. Crops are where palettes shift, because a tight crop can remove the one accent colour that tied an image to the family.

Do the format pass on one screen, side by side. If the set holds, publish. If it does not, adjust the crop rather than the generation — it is faster, cheaper and keeps the approved asset intact.

Then run the whole set through the AI image generator with the pack applied, and keep the approved batch plus its rulecard together. The rulecard is what makes the next campaign a fifteen-minute job instead of a renegotiation from first principles.