AI images in marketing: on brand rather than arbitrary
The difference between arbitrary and on-brand AI images lies not in the model but in a fixed recipe: the same colours, the same lighting, the same visual language — and a short list of what never appears in the picture.
In short
- A fixed image recipe — colours, lighting, visual language, exclusion list — works better than any change of model.
- Five parts make a usable image prompt: the subject core, colour values, lighting, composition, exclusion list.
- The exclusion list is the most important part: without it, text, symbols and company logos that are none appear.
- For professional articles, abstract subjects are nearly always the better choice than generated people.
The image recipe
A recipe is half a page and gets passed along unchanged with every image. It contains five items:
1. Colour values, written out
Not "violet" but the exact colour value — models hit named values considerably more reliably than colour names. Two to three main colours plus one accent.
2. Background
One fixed description, identical in every image. The background does more for recognisability than the subject.
3. Lighting
Soft glow, hard edges, deep shadows, cinematic? One description, reused word for word.
4. Composition
Plenty of empty space, a centred subject, depth — and a fixed aspect ratio. For social preview images 16:9 or 1200×630.
5. Exclusion list
No text, no letters, no numbers, no logos, no people, no icons. The most important part — and the one missing most often.
Worth knowing
The exclusion list decides the result more strongly than the subject description. Without it, generated images regularly show sequences of characters that look like writing and are none, as well as shapes reminiscent of company logos.
Both are unusable on a company website: false writing in the image looks unprofessional, and logo-like shapes can be confused with other people's brands. The list belongs at the end of every image prompt — that is where it is taken into account most reliably.
How an image prompt is built
[the subject core in one sentence – abstract, no objects] [background, word for word from the recipe] [colour values, written out] [lighting, word for word from the recipe] [composition: space, depth, arrangement] [exclusion list, word for word from the recipe]
The only part changing per image is the subject core. Everything else stays word for word — and that is exactly where the recognisability comes from.
Subjects for professional articles
Abstract subjects are nearly always the better choice. They do not age, infringe no personality rights, and do not look like stock photography from a database.
| What the piece says | Abstract subject |
|---|---|
| Something becomes simpler | Many elements order themselves into a few |
| Something is missing at one point | A continuous line breaks at one point |
| Two routes are on offer | One light source splits into two streams |
| A little becomes a lot | A dense core radiates into many shapes |
| Something hidden becomes visible | A surface opens and reveals structure |
Before publishing
- Convert. Generated images are often several megabytes. For preview images, 1200 pixels wide as JPEG; for images in the piece, 1400 pixels as WebP — that typically gives 20 to 80 kilobytes instead of several megabytes.
- Write alternative text. It describes what can be seen and what it means in context. Not a repetition of the caption.
- Set the dimensions in the HTML, plus
height: autoin the CSS — otherwise the layout shifts while loading. - Check for unwanted characters. Despite the exclusion list, writing-like shapes occasionally appear. Take a look at the finished image at full size.
- Check the usage rights. Read the provider's terms: commercial use permitted, labelling required, redistribution allowed?
The most effective single move for recognisability is the background. If every image has the same deep dark base tone with the same lighting, even very different subjects read as a series.
Conversely, a single image with a different background tears apart the effect of the whole series. That is why the background comes second in the recipe — ahead of the lighting and well ahead of the subject.
Help me develop an image recipe for our brand and derive image prompts from it. Our specifications: - Main colours as colour values: [values] - Accent colour: [value] - How our brand should feel: [three adjectives] - Where the images get used: [preview images, in the piece, presentations] - Images we like, and why: [description] Pieces I need images for: [list with title and core message] Tasks: 1. Phrase an image recipe with the five parts: subject description as a pattern, background, colour values, lighting, composition. Keep it to half a page. 2. Phrase an exclusion list that stops text, letters, numbers, logos, people or icons appearing. 3. Derive a subject core for each piece in my list – abstract, fitting the core message, without objects. 4. Assemble the complete image prompt for each piece: subject core, then the recipe parts word for word, then the exclusion list at the end. 5. Phrase alternative text per image describing what can be seen. Use only the colour values I have stated.
In closing
On-brand AI images come from repetition, not from better models: the same background, the same colour values, the same lighting, word for word in every image. Only the subject core changes.
And the exclusion list at the end of every prompt is not an extra but the part that decides usability — without it there are letters in the image that are none.
Common questions
How do AI images become on brand?
Through a fixed recipe passed along word for word with every image: the same background, the same colour values, the same lighting, the same composition. Only the subject core changes per piece. Recognisability comes from repetition, not from a better model.
Why does the exclusion list matter so much?
Because without it, sequences of characters regularly appear that look like writing and are none, as well as shapes reminiscent of other companies' logos. Both are unusable on a company website. The list belongs at the end of the prompt, because instructions there are taken into account most reliably.
Should you use generated people?
Rarely, for professional articles. They quickly look like stock imagery, age conspicuously, and depending on the execution, labelling obligations apply for machine-generated content that could depict real people. Abstract subjects are usually the better choice.
Which part of the recipe works most strongly?
The background. If every image has the same base tone with the same lighting, even very different subjects read as a series — and a single image with a diverging background tears apart the effect of the whole series.
What has to be done before publishing?
Five things: convert the image to delivery size — 1200 pixels as JPEG for preview images, 1400 pixels as WebP in the piece; write alternative text; set the dimensions in the HTML plus height: auto in the CSS; check for unwanted writing-like characters; and read the provider's terms on commercial use.
Marketing that sets itself up
The Studio Engine beta is live. Claim your spot and help shape it from the start.
Join the beta →