Write a one-page image section for your brand guidelines with three parts: colours (hex codes plus plain colour words, because image models don't reliably hit exact codes), one style (described and shown with five to eight approved reference images), and limits (what AI must never depict). Store the references in your tools and check every image against the page.
Without that page, each person prompts differently and your feed starts to look like five different companies. Worse, it starts to look like everyone else's, because every image tool has the same defaults: glossy stock-photo lighting, teal-and-orange grading, smiling people in suits shaking hands. Guidelines are how you get images that could only be yours, and how you stop the ones that could cause trouble.
What the image page needs to cover
Keep it to one page. Longer guidelines get skimmed, and image tools can't read them anyway: what reaches the model is a prompt and a few reference images. The page has seven parts:
- Palette: four to six colours, each with hex code, plain words and where it's used.
- Style block: two or three sentences describing the look, pasted into every prompt.
- Reference images: five to eight approved images that show the style, stored where everyone can reach them.
- People and composition: who appears, how, and how the frame is arranged.
- Limits: what is never generated, with the reason for each.
- Tools and settings: which tools are approved, with the style codes or reference settings to use.
- Before posting: the check and labelling rules, linked rather than repeated.
The written equivalent for your words, a brand voice guide, is a separate document with its own rules.
Colours: hex codes for designers, words for the model
Designers and design tools work in hex codes. Image models work from words and pictures, and asking for "#1F3A5F" in a prompt rarely produces that exact shade. So each palette colour needs both, plus a note on where it belongs. A filled-in palette for an illustrative financial planning firm:
| Colour | Hex | Words for prompts | Where it appears |
|---|---|---|---|
| Deep navy | #1F3A5F | deep navy blue | Dominant tone: clothing, furnishings, shadows |
| Sage | #8FA998 | muted sage green | Plants, walls, secondary areas |
| Warm sand | #E8DCC8 | warm sand, soft cream | Backgrounds and light surfaces |
| Soft coral | #E07A5F | soft coral, used sparingly | One small accent per image at most |
| Avoid | n/a | no neon, no saturated red, no teal-and-orange grading | Nowhere |
The "avoid" row matters as much as the palette. Saturated red reads as warning or loss, the wrong message for a firm talking about people's savings, and the teal-and-orange look is the most common default grade in AI imagery.
Then accept that generation gets you close, not exact, and plan a correction step. An illustrative prompt using the palette words:
A couple in their sixties reviewing papers at a kitchen table,
morning light, deep navy blue and muted sage green tones,
warm sand walls, one small soft coral accent, calm and
unhurried, natural photograph, no text, no logosA typical result: the navy came out closer to a mid blue, the walls leaned yellow rather than sand, and the coral accent appeared three times (a mug, a cushion and a flower). The correction: pull the blues and the wall tone towards the palette with your editor's colour adjustment, and remove two of the three coral accents. Five minutes' correction brings a generated image in line; for flat illustrations and vector graphics, recolouring tools such as Adobe's generative recolour can apply an exact palette.
Photographic, illustrated or real photos only?
Decide the kind of image before the details, because it changes every other part of the page. Many small firms end up with a mix: real photographs for anything that shows who they are, and one AI style for everything else.
| Choice | Suits | Watch out for |
|---|---|---|
| AI photographic | Warm, people-centred services where scenes of everyday life support the message (homes, families, small offices) | Anatomy and text errors; implied real clients; labelling of realistic images |
| AI illustration | Abstract services, audiences who notice detail errors, firms that want no invented people | Drifting into generic "corporate flat" art without a tight style block |
| Real photos only | Your team, your premises, your products, client stories (with consent) | Cost and time; still worth doing for anything customers will compare with reality |
The planning firm below chose AI photography for concept images and real photographs for its team page and office. The bookkeeping firm later in this tutorial chose illustration for everything AI makes. Both are coherent; what doesn't work is switching between them post by post.
Style: one look, described and shown
Pick one look and commit to it: photographic or illustrated, and within that, a consistent treatment of light, setting and mood. Mixing styles is the fastest way to make a feed look assembled from stock sites.
Write it as a style block of two or three sentences that goes at the end of every prompt:
STYLE: natural documentary photograph, soft window light,
real homes and small offices rather than corporate towers,
people mid-task rather than posing, shallow depth of field,
muted palette of deep navy, muted sage and warm sand.Words alone drift. Pair the block with reference images, because both main image tools now take them:
- Midjourney: a Style Reference captures an image's overall style (colours, medium, textures, lighting) without copying its objects or people, and works with versions 6 and later, including the current default V8.2. You add the image in the Style Reference section of the web app, or with
--srefand an image URL. For a broader house look, a Moodboard lets you collect several images into a style you apply with a code (--pplus the moodboard's ID). - Adobe Firefly: with Firefly's own Image models, the Styles section has a Reference option where you upload an image, and a Strength slider sets how closely the results follow it.
- Canva: a Brand Kit (on paid plans such as Pro and Teams) holds your logos, colours and fonts, which is where the palette and typography get applied when you lay out the final post.
Choose five to eight reference images that show the style at its best, ideally your own photography or images you have already approved, and keep them in one shared folder named for the purpose. Only use reference images you have the rights to use. Whether Midjourney is the right tool for your marketing at all is weighed up in whether Midjourney is worth it for small business marketing.
People and composition
Most AI images for service businesses contain people, and the defaults are where brand guidelines earn their keep. Rules for the planning firm:
- Show the clients you actually serve. The firm's clients are mostly in their fifties to seventies, with a real mix of backgrounds. The model's default "client" is a thirty-something couple in a show home.
- No handshakes, suits or pointing at charts. The three most generic images in financial marketing.
- One or two people per image. Fewer people means fewer anatomy errors and a calmer frame.
- Leave room for text. Keep one third of the frame quiet (a wall, a window, a table surface) so a headline can go on in the layout tool, never in the generation.
- Generate at the final shape. 4:5 portrait for feed posts, 16:9 for website banners and video thumbnails, so nothing important gets cropped.
Check the people in every batch for stereotypes as well as errors: who is shown as the expert, who as the client, and whether the range reflects your real clientele. The checks for that are in spotting bias and stereotypes in AI content.
Limits: what your AI images must never show
This is the part of the page that prevents real damage, so each limit comes with its reason. Staff follow rules they understand.
| Never generate | Why |
|---|---|
| Real people: staff, clients, public figures | Consent and likeness rights; a generated "client" implies a real endorsement |
| Your own premises, presented as real | Clients visit and compare; a flattering fake office misleads |
| Your logo or name inside the image | Models misspell and distort; add them afterwards from the Brand Kit |
| Text, figures, charts or screens | Garbled text looks careless; a rising chart implies returns |
| Wealth imagery: piles of cash, yachts, luxury cars | Implies outcomes and can clash with financial promotion rules |
| Distress: illness, bereavement, disasters | Exploits sensitive situations; use real, consented stories or none |
| Other companies' products, logos or buildings | Trademark and misrepresentation risk |
| "In the style of" a named living artist or photographer | Imitating someone's recognisable style invites rights disputes |
| Children | Not needed for this firm's work; avoids consent questions entirely |
Add one line on tools and rights. Adobe says Firefly's own models are trained on licensed content, but partner models offered inside the Firefly app aren't covered by that, and Adobe's IP indemnity is for enterprise customers only. Whatever you use, record which tool and model made each image, so you can answer questions later. Several of these limits have a legal basis, explained in the tutorial on the law and AI marketing images.
How the limits change for a law firm
The categories stay the same across sectors; the specifics don't. For an illustrative small law firm, the limits list gains its own entries. No gavels or scales of justice, the two most overused legal images, and gavels aren't even used in many courts, so they signal a stock image rather than a real practice. No courtroom, arrest or police scenes, which imply case types and outcomes. No documents with legible wording, because a generated "contract" with nonsense clauses is exactly what a lawyer's audience will read closely. And no images implying a client's situation, such as a couple arguing for a family law page, since those put a face on something clients find painful. The firm's style block leaned on architecture and materials instead: stone, timber, daylight through tall windows, desks with blank paper, and people shown from behind or at a distance.
A filled-in image page for a financial planning firm
Pulling the pieces together, here is the complete page for the illustrative four-planner firm, the version pinned in its shared drive:
AI IMAGE GUIDELINES v1.0 (September 2026) - owner: marketing lead
PALETTE Deep navy #1F3A5F (dominant) | Muted sage #8FA998 |
Warm sand #E8DCC8 (backgrounds) | Soft coral #E07A5F
(one small accent max). Never neon, saturated red,
teal-and-orange grading.
STYLE Paste at the end of every prompt:
"Natural documentary photograph, soft window light, real
homes and small offices, people mid-task not posing,
shallow depth of field, muted navy, sage and sand."
References: /Brand/AI-image-references (6 images).
Midjourney: use moodboard [ID] with --p.
Firefly: Styles > Reference, Strength about two-thirds.
PEOPLE Clients 50-75, mixed backgrounds. 1-2 people per image.
No suits, handshakes, pointing at charts.
FRAME 4:5 for feed, 16:9 for web. Keep a third of the frame
quiet for text. No text, logos or screens generated.
NEVER Real people | our office as if real | our logo | text,
charts, figures | wealth imagery | distress | other
brands | "in the style of" living artists | children.
BEFORE Run the image error checklist. Label per platform rules.
POSTING Log tool, model and date for every image.It fits on one screen, which is the point. A new member of staff can read it in two minutes and produce an on-brand image the same afternoon.
Testing the page: ten images, one scorecard
Before the page becomes the rule, test it. The planning firm's marketing lead generated ten images for real upcoming posts, following the page exactly, and scored each on three questions: palette right, style consistent with the references, no limit broken. Illustrative results:
| Outcome | Images | What happened |
|---|---|---|
| Passed as generated | 6 | Palette close enough, style consistent |
| Passed after colour correction | 3 | Blues too bright; fixed in about five minutes each |
| Broke a limit | 1 | A "planning meeting" image put a laptop with a rising chart on the table |
The broken limit was instructive. The prompt said nothing about charts; the model added one because "financial planning meeting" and "rising chart" appear together in countless stock images. The fix went into the page as a standing prompt ending: "no screens, charts, graphs or paperwork with visible numbers". One test round of ten images turned a sensible page into one that holds up in use.
Repeat the test whenever a tool changes its default model. Midjourney, for example, made V8.2 its default on 24 July 2026, and a new default can shift colour, lighting and texture enough that the same prompt and references no longer match last month's posts. Ten images and a scorecard, run the week an update lands, tell you whether the style block or references need adjusting before the feed drifts. Keep the scorecards from each round: they show which parts of the page do the most work.
A bookkeeping firm that chose illustration instead
Photographic style isn't the only option, and for some businesses it's the wrong one. An illustrative bookkeeping firm decided early that AI photographs of people would always feel slightly false for a firm whose pitch is accuracy, and that realistic images of offices and paperwork invited exactly the errors (garbled spreadsheets, impossible calculators) its audience would notice first.
It chose flat illustration instead: simple shapes, no faces, two brand colours plus a neutral. Its style block reads: "Flat vector illustration, simple geometric shapes, no faces, no text or numbers, forest green and pale gold on off-white, generous empty space." The before-and-after shows why it works. Before, a photographic prompt for "small business owner doing their accounts" produced a smiling stranger at a desk with a nonsense spreadsheet on screen. After, the illustration prompt produced an abstract desk, a plant and a stack of shapes suggesting paperwork: nothing to misspell, nobody to misrepresent, and unmistakably the firm's colours. Illustration also sidesteps most labelling questions, since clearly non-realistic images rarely need a platform label.
Keeping it consistent across people and tools
A realistic way guidelines fail: a new assistant joined the planning firm, used a different image tool she knew from a previous job, and pasted the style block without the reference images. Within a month the feed had a visibly different look: brighter, glossier, more stock-like. Nobody had broken a rule on paper; the references just hadn't travelled with the words. Four habits prevent that:
- Approved tools only, listed on the page with their settings. A new tool gets tested with the ten-image scorecard before anyone uses it for posts.
- A shared prompt library of five or six proven prompts for common post types (client meeting, home, office, seasonal), each ending with the style block.
- A quarterly grid review. Screenshot the last 30 posts as a grid. Inconsistencies invisible post by post are obvious side by side.
- One owner who approves changes and keeps the version number current.
Every image still goes through the same pre-posting routine: the checklist for AI image errors, then the rules for labelling AI images on each platform. For turning these guidelines into finished social graphics with text and layout, creating on-brand social media graphics with AI picks up where this page stops.
AI image guidelines: follow-up questions
Can I train an image model on our own brand images?
Some tools offer custom models or style training, usually on business or enterprise plans, and they can help large brands with many assets. Most small businesses get enough consistency from a written style block plus a style reference or moodboard built from five to eight approved images. Try that first, and only look at custom training if the reference approach clearly fails.
Should our AI images match our existing photography?
They should sit comfortably next to it, which usually means the same palette, similar lighting and the same kind of setting. They don't need to be indistinguishable, and they shouldn't pretend to be your real premises or people. Many firms use real photos for anything that shows who they are and AI images only for concepts and backgrounds.
Who should own the image guidelines in a small business?
Whoever already owns the brand or marketing, even if that's the owner. One person approves changes to the palette, style block and limits, keeps the reference folder tidy and reviews the feed each quarter. Everyone else follows the page. Shared ownership sounds democratic but usually means the style drifts and nobody notices.
Further reads
- Canva AI for Small Business: Which Features Save Real Time? — Which Canva AI features save time for a small business.
- How to Keep AI-Made Videos On-Brand: Captions, Voice and Intros — Extend the same guidelines to AI-made video.
- How to Build a Brand Voice Guide That AI Can Follow — The written-voice counterpart to your image guidelines.
- How to Create Product Photos With AI on a Small Budget — Apply the palette and limits to product photos.
- How to Write a House Style Guide Your Team and AI Both Follow — The written house style that sits beside the image page.
- How to Keep Your Brand Safe When AI Places Your Ads — Brand safety once AI starts placing your ads.
- Can AI Mock Up a Flower Arrangement Before You Make It? — What AI mock-ups of arrangements get right and wrong, three ways to make one, a prompt that respects real flowers, and wording that sets expectations.
- How Interior Designers Use AI for Mood Boards and Presentations — How interior designers use AI for concept boards, real-product selection boards and client decks, with prompts, sample outputs and the traps to avoid.
- AI Room Visualisation From Client Photos: A Designer's Workflow — Turn a client's room photo into an honest concept visual: shooting rules, a lock-list prompt, drift checks and the caption every image needs.
- AI for Freelance Designers: Faster Concepts Without Rights Issues — Where AI speeds up concept work, how to pick generators by their terms, a process log that proves authorship, and the checks before you hand over files.
- AI Product Photos vs a Photographer: Cost and Quality Compared — Per-image costs of AI product photo tools against a photographer's quote, a shot-by-shot quality verdict and a 40-product worked example.
- How to Design Flyers and Printed Marketing With AI — Let AI draft copy and layout options, keep it away from small text, and do the print checks yourself: size, bleed, resolution, colour and QR code.
- Should You Design Your Logo With AI? Copyright and Quality Risks — Decide which parts of logo design AI can help with and what needs checking before you order signs, packaging or uniforms.
- Do You Own the AI-Generated Content Your Business Publishes? — What AI vendors' terms really hand you, why copyright can still be thin, and the edits, checks and contract wording that make published AI content safer.
- How to Write an AI Content Policy for Your Marketing — How to write a marketing-specific AI content policy, with red lines, before-and-after clause wording, labelling rules and a filled-in one-page policy.
- AI Alt Text: How to Generate and Check Image Descriptions — Generate alt text with AI for product photos, banners and charts, then check every description with a five-point test before it goes live.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Midjourney documentation (Style Reference, Moodboards); Adobe Help (Firefly reference images for styling, set styles for image generation); Canva Help (Brand Kits). Checked September 2026.