Overs

How do you keep AI images on brand?

Checked 8 min read

Short answer

Three things keep AI images on brand: written rules concrete enough for a model to follow, the right reference photos for each shot, and a review whose reasons feed the next round. First define "on brand" as things you can see: palette, light, lens, composition, styling, casting, era, grain and type. Then score every image against those before anyone outside the team sees it.

What "on brand" means in a picture

A brand book says how the brand should feel. An image model needs to know what the camera sees. Before you judge any AI image, turn your look into nine parts you can point at in a photo. A line a reviewer cannot point at in the frame is too vague for a model to follow.

Part of the lookWhat to pin downExample rule
PaletteThe three to six colors that appear most, as names plus hex codes, and how much of the frame each may fillChalk #F2EEE6 for grounds; tea amber #C8893A only on the product
LightDirection, softness, color and time of dayLow morning sun from camera right; long shadows falling left
Lens and distanceWide, standard or long-lens look; camera height; how closeStandard-lens look, camera at table height, never a wide-angle close-up
CompositionWhere the product sits, how much empty space, how you cropProduct on the left third; right third empty for copy
StylingSurfaces, props, fabrics, tidy or lived-inLinen and pale stone; one prop at most
CastingAges, looks, poses, where people lookAdults 25 to 40; nobody looks at the camera
EraWhich years the world in the photo belongs toPresent day; no phones in frame
Grain and finishClean or film grain, contrast, how deep the blacks goFine film grain; blacks lifted to charcoal
TypeWhether words go on the photo, which typeface, whereNo text in the render; headlines set later in the brand serif
The example rules are for a made-up brand of canned tea. Write yours from your own best images.

Rules, references and review: what each one fixes

Each control fixes a different failure. Rules with no references give you the right mood around the wrong product. References with no rules give you your product in a generic scene. Both without review drift a little further with every batch.

ControlWhat it fixesWhat it cannot fixWhere it lives
Written rulesGeneric scenes, wrong light, wrong props, off-palette colorThe exact look of your product or of a faceA one-page rules document, copied into prompts
Reference photosProduct shape, label, true color, the same face in every shotAnything they show that you did not want copiedA folder per shot, in a fixed order
Review with reasonsDrift across a campaign, and faults that repeatRenders already madeA decision log that feeds the rules

Write the rules as things a camera can do: see how to write brand rules an AI can follow. Measure them from your own best images; the measured version is your visual DNA.

Why AI images drift off brand

Each render is drawn fresh from what is in the request. Unless the last image you liked is in that request, as a reference or earlier in the same chat, the model has nothing to match it to. OpenAI lists this as a known limit of its GPT Image models: they may struggle to keep recurring characters or brand elements consistent across generations. Google’s Nano Banana models take up to 14 reference images in one request, but not all 14 get the same care. Nano Banana 2 keeps up to 10 objects and 4 people faithful; Nano Banana Pro keeps up to 6 objects and 5 people.

FailureWhat you seeUsual causeFix
Slow driftFrame 30 is warmer, glossier or busier than frame 1Prompts edited shot by shot; references swapped mid-campaignReuse the light, ground and palette sentences word for word; compare each new frame with the first approved one
The generic "AI look"Poreless skin, plastic sheen, the product dead center every time, light with no sourceGaps in the brief, filled with the model’s most familiar versionName the surface, the imperfection, the grain and the crop
Borrowed styleFrames that look like the mood board or a competitorBoard images sent as referencesSend the board as words; see using a mood board with AI
Palette creepBrand red turns orangeWarm light, grade words, a reference shot under warm bulbsFix the reference and check with an eyedropper; see brand colors in AI images
Face driftThe model in shot 8 looks like a sibling of the one in shot 1No character sheet, or it was not sent firstApprove a character sheet first; see keeping the same model
Product driftLabel, cap color or proportions changeToo many references, or the wrong one for the angleSend only the product view the shot shows

The cheapest guard against drift is to stop rewriting what should not change. On the trouser store photos our agency made in 2026, the backdrop and lighting sentences were reused word for word across the set, and only the framing and pose sentences changed from photo to photo.

An on-brand scorecard

Score each image before it goes to anyone outside the team. Give each line 0, 1 or 2 points. Two gates come first: if the product is wrong (label, shape, color, parts) or anything on your never-do list appears, the image fails whatever it scores.

Line2 points1 point0 points
PaletteOnly brand colors, in their rolesOne color slightly offAn off-palette color, or the brand color wrong
LightDirection, softness and color match the rulesRight direction, wrong softness or colorLight from the wrong side, or from nowhere
Lens and distanceMatches the house lookSlightly wider or closer than usualDistorted, or a look the brand never uses
CompositionPlacement and empty space as plannedPlacement right, crop too tight for the copyNo room for copy, or the product lost in the frame
StylingAllowed surfaces and props onlyOne extra propBanned props or clutter
CastingMatches the casting line and the approved character sheetRight person, pose or gaze offWrong person, or a face that drifted
EraNothing in the frame from the wrong yearsOne detail from the wrong yearsThe world looks like another decade
Grain and finishGrain, contrast and blacks as specifiedClose, but too clean or too crunchyPlastic look or a heavy filter
TypeNo stray letters; copy space where plannedCopy space in the wrong placeGarbled text or letters in the render

The top score is 18. Set a pass mark and write it on the scorecard. A pass at 14 with no zeros is a sensible start; raise it once you have a set of approved images to compare against. The full pre-publish check, from hands to file size, is the QA checklist for AI images.

Which reference photos go with which shot?

Give each shot only the pictures it needs, in a fixed order, and say in the prompt what each one is for. OpenAI’s prompting guide for GPT Image 2.5 asks for the same: number each input and name its job, such as subject, style, clothing or background. In the kept store photos from our trouser job, input image 1 was the approved character sheet, input image 2 was the flat lay of the trousers, and the prompt named both.

ShotSend, in this orderLeave out
Packshot on whiteProduct front; label close-upLifestyle photos, the mood board
Back or side viewThe product photo from that side, with the character sheet first if the product is wornPhotos of the other sides
Detail close-upTexture or label close-upFull-product shots
On a modelApproved character sheet, then the product view for that angleOther people, other outfits
Lifestyle sceneProduct photo; one prop photo if the prop must be exactBoard images, competitor ads
Campaign heroProduct photo; character sheet if a person appearsAnything with another brand’s logo

A working order for one campaign

  1. Update the brand rules and the never-do list.
  2. Photograph the product: front, back, sides, the label straight on, a texture close-up.
  3. If the campaign has a mood board, write it down as words.
  4. If people appear, make and approve a character sheet before any scene.
  5. Plan the shots and choose the references for each one.
  6. Render three to five test frames and score them. When a line fails, fix the rule as well as the prompt.
  7. Render the set and review it in rounds, with a reason for every verdict. See how to review AI images as a team.
  8. After the campaign, move every reason that repeated into the rules or the never-do list.

Questions people also ask

Can a chat assistant like ChatGPT or Gemini make on-brand images?
Yes, one image at a time. Across a set, paste the same rules and attach the same references in every chat, and compare each new image with the first one you approved. OpenAI lists consistency across generations as a known limit of its GPT Image models, so expect to correct some drift.
How many reference images should one shot get?
As few as the shot needs, each with a stated job. On our surf apparel job, the character sheet we kept used 5 references, and earlier attempts with up to 8 were not kept. Check each model’s published caps before you plan a shot; Google, for one, lists separate limits for objects and for people.
Do I need to train a custom model on my brand?
Most brand work does not. Rules, references and review cover it. A model trained on your images helps most when one look or character must repeat across hundreds of images. Adobe’s Firefly Custom Models, which Adobe opened as a public beta in March 2026, train on 10 to 30 images, and you retrain when the brand changes.
Who should own the brand rules?
One person, usually the brand manager, with the creative director signing off the light, lens and casting lines. Date every version and change it through review notes, so the whole team renders from the same page.

Where Overs fits

Overs runs the brand steps before any photo is made: research, brand rules, visual DNA (colors with hex codes, light, lens, grain, styling), art direction, casting and a quality check, with brand rules applied last. Each photo gets only the reference pictures its plan names. Approvals and rejections become brand memory for later photos.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.

  1. 1.Google AI for Developers: Nano Banana image generation (reference image limits)
  2. 2.OpenAI: Image generation guide (limitations)
  3. 3.OpenAI: Image prompting, GPT Image 2.5 prompting guide
  4. 4.Adobe Blog: Adobe Firefly expands video and image creation with new AI capabilities and custom models (March 19, 2026)
  5. 5.Adobe Help Center: How to train Firefly Custom Models