How do you keep AI images on brand?
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 look | What to pin down | Example rule |
|---|---|---|
| Palette | The three to six colors that appear most, as names plus hex codes, and how much of the frame each may fill | Chalk #F2EEE6 for grounds; tea amber #C8893A only on the product |
| Light | Direction, softness, color and time of day | Low morning sun from camera right; long shadows falling left |
| Lens and distance | Wide, standard or long-lens look; camera height; how close | Standard-lens look, camera at table height, never a wide-angle close-up |
| Composition | Where the product sits, how much empty space, how you crop | Product on the left third; right third empty for copy |
| Styling | Surfaces, props, fabrics, tidy or lived-in | Linen and pale stone; one prop at most |
| Casting | Ages, looks, poses, where people look | Adults 25 to 40; nobody looks at the camera |
| Era | Which years the world in the photo belongs to | Present day; no phones in frame |
| Grain and finish | Clean or film grain, contrast, how deep the blacks go | Fine film grain; blacks lifted to charcoal |
| Type | Whether words go on the photo, which typeface, where | No text in the render; headlines set later in the brand serif |
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.
| Control | What it fixes | What it cannot fix | Where it lives |
|---|---|---|---|
| Written rules | Generic scenes, wrong light, wrong props, off-palette color | The exact look of your product or of a face | A one-page rules document, copied into prompts |
| Reference photos | Product shape, label, true color, the same face in every shot | Anything they show that you did not want copied | A folder per shot, in a fixed order |
| Review with reasons | Drift across a campaign, and faults that repeat | Renders already made | A 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.
| Failure | What you see | Usual cause | Fix |
|---|---|---|---|
| Slow drift | Frame 30 is warmer, glossier or busier than frame 1 | Prompts edited shot by shot; references swapped mid-campaign | Reuse 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 source | Gaps in the brief, filled with the model’s most familiar version | Name the surface, the imperfection, the grain and the crop |
| Borrowed style | Frames that look like the mood board or a competitor | Board images sent as references | Send the board as words; see using a mood board with AI |
| Palette creep | Brand red turns orange | Warm light, grade words, a reference shot under warm bulbs | Fix the reference and check with an eyedropper; see brand colors in AI images |
| Face drift | The model in shot 8 looks like a sibling of the one in shot 1 | No character sheet, or it was not sent first | Approve a character sheet first; see keeping the same model |
| Product drift | Label, cap color or proportions change | Too many references, or the wrong one for the angle | Send 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.
| Line | 2 points | 1 point | 0 points |
|---|---|---|---|
| Palette | Only brand colors, in their roles | One color slightly off | An off-palette color, or the brand color wrong |
| Light | Direction, softness and color match the rules | Right direction, wrong softness or color | Light from the wrong side, or from nowhere |
| Lens and distance | Matches the house look | Slightly wider or closer than usual | Distorted, or a look the brand never uses |
| Composition | Placement and empty space as planned | Placement right, crop too tight for the copy | No room for copy, or the product lost in the frame |
| Styling | Allowed surfaces and props only | One extra prop | Banned props or clutter |
| Casting | Matches the casting line and the approved character sheet | Right person, pose or gaze off | Wrong person, or a face that drifted |
| Era | Nothing in the frame from the wrong years | One detail from the wrong years | The world looks like another decade |
| Grain and finish | Grain, contrast and blacks as specified | Close, but too clean or too crunchy | Plastic look or a heavy filter |
| Type | No stray letters; copy space where planned | Copy space in the wrong place | Garbled 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.
| Shot | Send, in this order | Leave out |
|---|---|---|
| Packshot on white | Product front; label close-up | Lifestyle photos, the mood board |
| Back or side view | The product photo from that side, with the character sheet first if the product is worn | Photos of the other sides |
| Detail close-up | Texture or label close-up | Full-product shots |
| On a model | Approved character sheet, then the product view for that angle | Other people, other outfits |
| Lifestyle scene | Product photo; one prop photo if the prop must be exact | Board images, competitor ads |
| Campaign hero | Product photo; character sheet if a person appears | Anything with another brand’s logo |
A working order for one campaign
- Update the brand rules and the never-do list.
- Photograph the product: front, back, sides, the label straight on, a texture close-up.
- If the campaign has a mood board, write it down as words.
- If people appear, make and approve a character sheet before any scene.
- Plan the shots and choose the references for each one.
- Render three to five test frames and score them. When a line fails, fix the rule as well as the prompt.
- Render the set and review it in rounds, with a reason for every verdict. See how to review AI images as a team.
- 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
- Google AI for Developers: Nano Banana image generation (reference image limits)
- OpenAI: Image generation guide (limitations)
- OpenAI: Image prompting, GPT Image 2.5 prompting guide
- Adobe Blog: Adobe Firefly expands video and image creation with new AI capabilities and custom models (March 19, 2026)
- Adobe Help Center: How to train Firefly Custom Models