AI branding without the slop: how I art-direct AI images so they look like your brand

AI branding stops looking generic when a person directs the tool, the way an art director directs a designer. I start with a written brief and a real reference, and I get a direction approved before anything is generated. After that I use real photos where the subject is real, check the work against a list of banned clichés, and reject whatever still looks generated.

What AI slop is, and why it happens

AI slop is generic output from an image tool: the same glossy faces, purple gradients, melted text, and an overall look that nobody chose. I think most people can tell something is off, even when they cannot say what.

It happens when nobody directs the tool. Image tools tend to fill every gap in your instruction with the most common answer they know. Ask for “a logo for a consulting firm” and you have left every decision open: the idea, the shapes, the colours, the mood. The tool makes all of them, and each time it leans towards what it has seen most often.

This matters for a brand, because a brand is how people recognise you. If your images look like the images from every other account that used the same tool, a customer scrolling past cannot tell you apart from them.

AI does make design faster, mostly because it can explore many directions quickly, but deciding what is right for your brand still takes a person.

Before anything is generated

I work in a fixed order, because each step stops a different kind of slop. Four of the seven steps happen before the tool runs.

1. A written brief comes first. It says what the piece is for and who it must reach. For an existing brand, it also holds your colours, your tone and the looks you have ruled out. A poster that sells tickets and a poster that builds prestige are different designs, even for the same event.
2. Then I find a real reference to aim at, such as an existing poster or photo with the mood and structure I want. I borrow the composition and the feeling, never the artwork.
3. Next I write a direction, give it a name and a few sentences, and you approve it before I generate a single image. “Editorial: two colours, one real photo, a heavy headline, a lot of empty space” is a direction. “Clean and modern” is not. I also try to finish one sentence: “This is unmistakably yours because …” If I cannot, the direction is still generic and I go back.
4. Wherever the subject is a real person, place or event, the picture is a real photo. Generated images are for ideas and moods. When I overhauled the Feed Philly Coalition website in August 2026, I replaced the generated imagery on the main pages with real photos. That site is about real people in Philadelphia, so the pictures should be real too.

If you use AI tools yourself, try the first three steps before you write a single prompt. They need nothing but some thinking, and I think they do more than any clever wording in the prompt itself.

After it is generated

The last three steps are about checking what comes out.

5. Every brand gets a written list of banned clichés, and every image is checked against it. The general part of the list covers purple-to-blue gradients, a soft glow around the subject, blurred backgrounds used as decoration, everything centred, a headline floating in the middle, and the literal picture of the topic, like a lightbulb for an idea or a handshake for a deal. Each brand adds its own entries, and when someone rejects a piece, the reason goes onto the list so the mistake does not come back. The list is for checking the result. Writing “no lightbulbs” into a prompt can invite the lightbulb, so I brief better ideas instead.
6. Checking happens at real size. I look at each image at full size, where melted letters and strange details show up, and again at the size people see it: a phone feed, a printed page, a 16-pixel icon. I read all the text in the image letter by letter.
7. Last, I reject anything that looks generated. A face that is too smooth goes back, and so does a shadow with no light source. So does a look I keep seeing on other accounts. The first render is a draft, and I stop when the work looks right, not when it matches the words I typed.

How I handle logos

AI is good at one part of a logo job: it sketches many directions quickly. For the Gemini Strategic Consultants mark, I asked for sheets of nine different ideas at a time. Gemini means twins, and the mark I chose is two hands closing into a G.

What I got depended on the brief. In my tests, asking for an object, like “a book” or “a link”, gave me the stock icon of that object. A vague brief of a sentence or two gave me stock ideas such as a compass, a gear, a puzzle piece, a handshake and a chart. What worked was one idea with a second meaning built into the shape. Even the best sheet had stock icons among its nine, so picking is not a formality.

Picking is the human part. I choose one mark and redraw it as a clean vector, because a generated picture is never the file you hand to a printer or a developer. Then I try to break it. It has to pass each of these:

• A 16-pixel version, the size of a browser tab icon.
• One colour only, and then reversed on a dark background.
• Turned and mirrored, in case it reads as something I never meant.
• A one-sentence description that a stranger could sketch from, with no picture to help.
• A swap test: it should look wrong on a bank or on your biggest rival. If it would fit either one unchanged, it is generic.

AI people get a label

Some jobs need a person who does not exist. For app ads I use two fictional creators, Mila and Theo, and I keep each of them looking the same from ad to ad. They are AI, and every place they appear says so. A face in an ad suggests that a real person stands behind the words. Mila and Theo are not real people, so the ad should say that.

Where this shows up in real work

The Velvet Riot Wig Co. needed images that read as one brand. I made a hero image, a product shot, a lookbook frame, a brush wordmark and a monogram. The check for a set like this is to put the pieces side by side. If they look like one brand, the direction held. If each looks like its own separate generation, it did not.

I also designed a Bijaya Dashami greeting poster (2083 in the Nepali calendar) for a neighbourhood committee. Bijaya Dashami is the tenth day of Dashain, one of Nepal’s biggest festivals, and the people who see the poster know it well. A real event needs a real reference before the first image is made.

SameShot is the progress-photo camera app we build at Appynox. For its store listing we made two complete looks, one light and one dark, to test against each other. Both had to be fully finished, because if one look were rougher than the other, the result would tell you about the polish and little about the idea.

Then there is volume. I built a system that makes social graphics for many brands at once, each to that brand’s own rules for colour, voice and banned looks. That only works because the rules are written down. I cannot hold that many brands in my head, and a tool only knows what it has been given. A person still approves every post, and the same approach is behind my AI social media automation service.

Start with a brief

If you run a small business or a nonprofit and want logos, brand images or social graphics that look like your own, the details are on my AI graphic design and branding page. Tell me what you sell and who you need to reach, and book a 30-minute call to start the brief.