How to keep AI-generated design on brand

Image generation removed the cost of producing a good-looking image and, in doing so, exposed the thing that actually mattered all along: whether the images look like they came from the same company. A feed of fifty individually attractive, mutually unrelated visuals reads as a stock library. A feed of fifty simpler images that share a palette, a typeface and a way of framing things reads as a brand.

Write the rules down as constraints

A brand guide for this purpose is not a 40-page document. It is a short list of constraints that every asset obeys, written concretely enough that two different people applying them produce compatible results:

• Colours: two brand colours, one neutral, one accent used sparingly. Exact values, not names.
• Type: one typeface for headlines, one for everything else. Two weights each.
• Composition: where the subject sits, how much empty space, whether text goes left, centred or bottom.
• Subject treatment: photographic or illustrated, warm or cool, busy or spare, people or objects.
• Light: direction and hardness. This is the single most under-specified rule and the one that most affects whether images look related.
• What never appears: stock-photo handshakes, fake dashboards, clip-art icons, gradients you do not use elsewhere.

Six lines. Every generated asset is checked against them, and anything that violates one is regenerated rather than published because it looked nice.

Constrain the palette harder than feels comfortable

Colour is what people recognise before they read anything, and generation tools default to whatever is most photogenic — which is a different palette every time. Two brand colours plus a neutral is enough for a year of assets. The discipline is refusing the fifth colour, and the test is simple: shrink three of your images to thumbnail size, put them side by side, and see whether they look like a set.

Text on an image is design, not decoration

Most social visuals carry words, and this is where generated assets most often fall apart. The rules are unglamorous:

1. One message per image. If it needs two, it needs two images.
2. Readable at thumbnail size — which is how nearly everyone sees it first. If you cannot read it at 20% scale, it is too small.
3. Contrast that survives a busy background. A darkened band behind the text beats a clever transparency every time.
4. Your typeface, not the tool's default. Type is a stronger brand signal than the logo on a small screen.
5. Leave the edges alone. Platforms crop, and text near a corner disappears on someone's phone.

Non-Latin scripts deserve particular care: Arabic, for example, must be rendered with correct joining and shaping, and text that is not connected properly is instantly recognisable as broken to every reader of the language — while being invisible to a reviewer who does not read it. If you publish in a script you do not read, have someone who does check it before it ships.

Build formats, not one-offs

Recognition comes from repetition. Define five or six recurring formats and reuse them rather than inventing a new look per post:

• Quote or statement card.
• Product in a real context.
• Before / after, split the same way every time.
• A single number, large.
• A step-by-step carousel with a fixed layout.
• Announcement, with the same badge in the same corner.

Each of these maps neatly onto the recurring formats in a content calendar, which is what makes batching visuals possible: six of the same format take a fraction of the time of six different ones, and they come out looking related.

Sizes and where things get cropped

• Produce square as the default; it survives the most placements.
• Vertical 4:5 or 9:16 for feeds and stories that reward height.
• Wide 1200×630 for link previews — the ratio every platform crops to.
• Keep anything essential inside the middle 80%. Crops are unpredictable and unforgiving.
• Check the thumbnail, always. It is the version most people will ever see.

What to keep human

Generation is excellent at volume within a defined style and poor at defining the style. So:

• The logo is drawn once, by a person, and used as a file. Never regenerated — a mark that changes shape is not a mark.
• The brand rules are decided by a person, once, and then applied mechanically.
• Anything appearing in an ad or a physical product gets a human review before it ships.
• Product photography that has to be accurate — colour, texture, what is included — is photography, not generation. A misleading product image is a return and a refund.

This is the same boundary that applies to content automation: humans set the constraints and approve the output, machines do the repetitive production in between.

A five-minute audit

Take your last nine images. Put them in a three-by-three grid at thumbnail size and answer three questions:

1. Do they look like they came from one company? If not, the palette or the light is drifting.
2. Can you read every one at thumbnail size? If not, the type is too small or the contrast too low.
3. Could you tell they were yours with the logo removed? If not, nothing in the system is distinctive yet — usually because the palette is too broad or the composition changes every time.

Fix whichever one fails, tighten the rule that allowed it, and run the same audit next month. Brand consistency is not achieved once; it is maintained by refusing exceptions.

Frequently asked questions