Graphic Design & Branding with AI

Why design + AI

Professional-looking design used to require a designer and expensive software. That changed: you can generate imagery, get a palette, and produce posts and banners that look considered, without design training. You direct, the tool executes, you select and polish.

Generating an image is not designing

This is the distinction that separates work that looks professional from work that looks like a pile of nice pictures, and it is why so many AI-designed brands feel off without anyone being able to say why.

Generation produces one striking thing. Design is a set of constraints applied repeatedly — the same three colours, the same two typefaces, the same spacing, the same treatment of photography — so that the thirtieth post is recognisably related to the first. Recognition comes from repetition, not from any individual asset being impressive.

AI is extremely good at the first and structurally bad at the second. Ask for a social post and you get a good one. Ask again tomorrow and you get a good one that shares nothing with yesterday's. Do that for a month and you have thirty attractive images and no brand.

So the workflow that works is not "generate assets". It is decide the constraints once, then use AI to fill them. Pick the palette, the typefaces and a layout grid — these are a handful of decisions, made once, written down — and then generate within them. The tool becomes a fast way to produce the next variation of a thing you already defined, which is exactly what a design system is for.

What these tools are reliably bad at

Knowing the failure modes tells you where to draw the line between the generator and a real design tool. These are not bugs awaiting a fix; they follow from how image models work.

• Text inside images. Words come out misspelled, malformed or subtly wrong, and it is worse for any script the model saw less of. Never let a generator produce the words on your asset.
• Exact colours. Ask for your brand hex and you will get something near it. Near is not a brand colour; near is what makes a feed look slightly grubby.
• Precise layout. Alignment, consistent margins, a logo in the same place every time — these are the mechanics of a layout tool, not a prompt.
• Repeating a subject. The same character or product across several images, actually matching, remains difficult and fiddly.
• Vector output. Generators produce pixels. This matters enormously and has its own section below.

The division of labour that follows is clean: generate the illustrative and photographic elements; do type, layout, exact colour and the logo in a design tool. Almost every bad AI-designed asset comes from asking the generator to do one of the second group.

Why a generated logo is a problem

A logo is not an image. It is a mark that has to work at every size, in one colour, embroidered, on a dark background, as a 16-pixel favicon and on a sign. That means it has to be vector — mathematically defined shapes that scale without degrading — and image generators produce raster pixels.

A generated logo will look excellent in the square it was born in, and then you need it in the corner of a document, and it is soft. You need it on a dark background and there is no transparency, or the edges are fringed. The printer asks for vector and you have a PNG.

The workable route is to treat generation as concept exploration, which is genuinely useful: produce forty directions in an hour, find the two that feel right, and then have the chosen mark redrawn as vector — by you in a vector tool if you can, or by a designer, which is a small, cheap, well-defined job. Auto-tracing the raster is possible and usually produces a mess of nodes that looks fine on screen and betrays itself at scale.

Two further cautions specific to logos. Generated marks can resemble existing trademarks, because the model learned from a world full of logos — similarity is a legal risk that no amount of "but the AI made it" resolves, so search the trademark registers for your market before you commit. And the copyright status of purely machine-generated work differs between jurisdictions and is unsettled in several; if the mark matters, the human redrawing step helps here too. Worth a professional opinion for anything you intend to own.

Brand identity

• Logo: generate concepts, choose a direction, redraw it as vector. Simple survives — a good mark works small and in one colour.
• Colour palette: ask for a palette based on the brand's personality with hex codes — then check it, as below.
• Typography: ask for pairings that suit the tone. This is one of the more reliable uses, because it is a recommendation rather than a rendering.
• Brand guide: collect the colours, type and logo into one document. This is the artefact that makes everything else consistent, and it takes an afternoon.

The brand guide is the highest-leverage thing on that list and the one people skip because it produces nothing to post. It does not need to be elegant — a single page with hex codes, font names, logo files and two rules about spacing outperforms anything else you could do in the same hour, because every future asset and every person you hand work to refers to it.

Palettes, and the check the AI will not do

Generated palettes are usually pleasant and frequently unusable, for one specific reason: nothing in the process checks contrast.

A palette can look sophisticated and still put text at a contrast ratio that is unreadable for a substantial share of people — and on a phone in daylight, for everyone. The accessibility guidelines give concrete targets for text against its background, and they are worth meeting whether or not anything obliges you to, because failing them means people cannot read your words.

So: take the palette, put your actual text colour on your actual background colour in a contrast checker, and adjust until it passes. Usually it is a small shift in lightness that nobody will notice except the people who could not read it before.

Two more palette habits. Decide which colour is the accent — the one used sparingly for the thing you want clicked — and do not use it for decoration, or it stops meaning anything. And never let colour be the only signal: a status, a category or a chart series distinguished purely by hue fails for colour-blind viewers and in greyscale print.

Typography, and the licence trap

Asking for font pairings works well; it is a well-trodden question with sensible conventional answers, and two typefaces plus a weight range covers nearly everything a small brand needs.

The trap is licensing, and it catches people after they have built everything. A typeface you downloaded is licensed for particular uses — desktop, web embedding, app, broadcast — and each is priced separately. Free-for-personal-use is very common and is not free for your business. Before you standardise on a face, read what its licence covers for the number of people and the kinds of use you actually have. The open families are genuinely good now and licensed for essentially anything, which is why so much of the web runs on them.

One practical rule: whatever you choose, make sure it has the weights you need and covers the characters you need — including any accented or non-Latin text — before it becomes your brand font. Discovering a missing weight after fifty assets is an annoying afternoon.

Consistency & templates

The professional look is consistency: the same colours, type and structure everywhere. Build templates in a design tool and reuse them. AI then speeds up producing variations inside the template without breaking the visual language.

The practical version is a small set of masters — one post, one story, one banner, one document cover — with the brand applied and the positions fixed. New asset means duplicating a master and changing the content, not starting from a blank canvas and hoping it comes out similar. This is unglamorous and it is the single biggest difference between work that looks like a brand and work that looks like a folder.

Where AI fits: generating the image that goes in the slot, writing the variants of the line that goes in the text box, resizing across placements. The template holds the constraints; the model fills them.

Graphic assets

• Social posts — see AI for social.
• Banners and ads — versions across sizes, quickly.
• Thumbnails — see thumbnails with AI.
• Icons and illustrations — consistent-style assets with prompts.
• Presentations and documents — on-brand layouts generated for you.

For icons specifically, resist generating them. Icon sets exist, they are consistent by construction, most are free, and they are vector. A generated icon set will drift in weight and style across the set, which is the one thing icons must not do.

Prompting for assets that fit a system

Prompting for a brand asset is a different job from prompting for a striking picture, and the difference is that you are trying to produce something repeatable.

Four things make a generated image usable inside a template rather than merely nice:

• Specify the framing and the empty space. "Subject on the left third, plain background on the right" gives you somewhere to put the headline. A perfectly composed image that fills the frame is useless the moment text has to go on it.
• Name the visual treatment and reuse it verbatim. Whatever you settle on — a lighting style, a palette, an illustration approach — becomes a fixed block of text you paste into every prompt. Consistency comes from repeating the words, not from remembering the vibe.
• Ask for the aspect ratio you need, rather than cropping later. Cropping to a square something composed as a landscape is how subjects lose their heads.
• Keep the seed or the reference when a tool offers one. It is the closest thing to "more like that" and it is what makes a set look like a set.

Keep a short file of the prompt blocks that worked. It becomes the visual half of your brand guide, and it means the person who does this after you produces images that match, which is the entire point. See image prompts for the mechanics.

Rights, and what you can actually publish

Before an image goes on anything commercial, three questions have answers worth knowing.

• Does the tool's licence permit commercial use on your plan? Some grant it only on paid tiers; free tiers often reserve rights or require attribution.
• Does it contain something you do not have rights to? Generators reproduce recognisable characters, brands and celebrity likenesses if asked, and sometimes when not asked clearly. Putting one on an ad is the same infringement it would be if you had drawn it.
• Can you prove provenance later? Keep prompts and generation dates for anything important. If a client or a platform asks where an image came from, the answer "I made it in tool X on this date" is worth having.

None of this is exotic. It is the same diligence stock photography always required, applied to a source that makes it easier to forget.

The files you will actually be asked for

A small thing that saves repeated annoyance: assemble the full set once, when you finalise the identity, rather than hunting each time someone asks.

• Logo as vector (SVG for web, and whatever your printer asks for), in full colour, one colour, and white-on-dark.
• PNG with transparency at a few sizes, for the places that will not take vector.
• Favicon — a simplified mark that survives at 16 pixels, which is usually not your full logo.
• Social profile and cover images at each platform's current dimensions, remembering the profile image gets cropped to a circle.
• The brand guide — hex codes, font names and licences, and where the files live.

An afternoon that sets you up for a year

Everything above collapses into one session, and doing it properly once is worth more than any individual tool choice.

1. Write three words describing how the brand should feel. Not aspirational nonsense — words you would actually use, like "plain", "warm", "technical". Every later decision gets checked against them.
2. Pick two typefaces, one for headings and one for body, and confirm the licence covers your use. Two is enough. Two is usually better than three.
3. Pick four colours — a background, a text colour, one accent, one neutral — and run the text-on-background pair through a contrast checker before you fall in love with it.
4. Settle the logo direction, generating freely to explore, and book the vector redraw rather than shipping the raster.
5. Build four templates — square post, story, banner, document cover — with the type and colours applied and the logo in a fixed position.
6. Write the one-page brand guide: the hex codes, the font names and licences, where the logo files live, and the prompt block for imagery.

After that afternoon, producing an asset is filling in a template, and that job is one an AI tool genuinely does well. Before it, every asset is a fresh set of decisions — which is why the output drifts and why it never quite looks like a company.

When to pay a designer

• The logo, at least for the redraw. Cheap as a discrete job, and it is the asset with the longest life and the most technical requirements.
• Anything going to print. Bleed, colour profiles and resolution are where confident amateurs discover an expensive reprint.
• When you cannot say why it looks wrong. That feeling is usually spacing and hierarchy, and it is exactly the thing training teaches and prompting does not.
• The system, not the assets. Paying once for a template set and a brand guide, then producing everything yourself inside it, is the best value in this whole category.

Common mistakes

• Inconsistency. Every post in a different style reads as amateur. Fix the constraints first.
• Overloading. Too many colours, faces and effects. Less is more, and less is also easier to repeat.
• An over-complex logo. Impressive at 1000 pixels, mud at 16. Test small and in one colour before you commit.
• Letting the generator set the words. Text in a generated image is misspelled often enough that it should simply never be how your copy gets on the asset.
• Skipping the contrast check. A beautiful palette nobody can read is a failed design, however good it looks in the swatch.
• Ignoring rights. Check the commercial terms before publishing, not after a takedown.