Starting a shirt brand with AI

Search this and you get a stack. Midjourney for the art, something else for background removal, an upscaler, a mockup tool, a print-prep checklist, then niche research. Eight steps, a subscription at four of them, and a warning that raw output "isn't a finished product."

That advice is accurate about the world it was written in. It's also a description of a job nobody wanted — and most of it has already stopped existing.

What do you actually need to start a shirt brand now?

Something specific to say, and someone specific to say it to.

That's not a motivational answer, it's what's left after subtraction. The old list had two halves: a production half — resolution, cut-outs, color, file prep, minimum orders, stock — and a judgment half: what this is, who it's for, why anyone would wear it rather than a different shirt.

The production half is gone. It's the half every guide is still teaching, because it's the half that's teachable.

Isn't the AI part the hard part?

The AI part is the commodity part, and treating it as the achievement is the most common way to end up with a store nobody visits.

Everyone reading this can produce a competent image. That was true a year ago and it's more true now. So a good-looking graphic is the floor, not the edge — which is why the "funny cat shirt" graveyard exists: those designs aren't badly made, they're addressed to nobody.

What still separates work is what it's about. A wolf is a wolf. The specific ridge your running club calls The Wall, drawn the way people who've run it would recognise, is not available anywhere else and can't be prompted by someone who wasn't there.

What happened to background removal, upscaling, and print prep?

They stopped being steps.

Bringing an image here means bringing the file. Separating the artwork from whatever square it arrived on, sizing it for the press, and getting the colour to behave on the fabric you picked are handled on this side. There's no export preset to choose, no DPI to set, no transparent PNG to produce — and asking a model for a transparent background usually damages the file anyway.

That's one paragraph rather than a chapter, which is the point: half of every competing guide is instructions for work that has already been absorbed.

What's actually scarce, then?

Two things, and neither is production.

Specificity. The narrower the reference, the stronger the object. A shirt legible to four hundred people and meaningless to everyone else is a better product than one that's mildly agreeable to everybody, because the four hundred will actually want it. Generic is the failure mode that AI made cheap to reach.

Distribution. If making costs nothing and producing carries no risk, the scarce step is somebody seeing the thing. That was always true; it used to be hidden behind the inventory problem, because you couldn't get far enough to discover it.

Which means the effort has moved. Time spent on the twelfth variation of a graphic is time not spent on who this is for and where they already gather.

How small can the audience be?

One, and that's a real answer rather than a consolation.

Nothing is held as stock — each piece is produced after it's bought — so a design that sells twice costs the same to carry as one that sells two thousand times: nothing. There's no minimum order and no run to fill.

That removes the question every previous era forced you to answer first: would enough people want this? You can now make the thing for the eleven people it's for, and find out whether it's actually four hundred.

What does a first collection look like?

Three to five pieces that share a treatment and differ in subject.

One design is a product; several related ones read as a point of view, and a point of view is what people follow. Holding the palette and the technique constant while varying what each piece is of is what makes a set legible as a set — and it gives a visitor something to choose between rather than a single yes-or-no.

It also makes the result interpretable. Five pieces that vary along one axis tell you which axis mattered; five unrelated pieces that flop tell you nothing.

What does it cost to be wrong?

An evening.

That number is the whole structural change, and it's worth sitting with. Under the old model, testing an idea meant buying a hundred units, so you tested one or two and committed hard to whichever you picked. The correct strategy was conviction, because the cost of being wrong was enormous.

Now the correct strategy is volume of attempts. Publish several, show them to the people they're for, and let the answer come back. Being wrong cheaply and often beats being right once, and the people who struggle are usually the ones still budgeting for conviction.

How does selling work here?

You publish a design and earn when other people buy it.

Your work becomes something anyone can order in their own size and colour, produced to order and shipped to them. No stock on your side, no fulfilment, no customer service, and no inventory risk — a design that sells three times costs you nothing more than one that sells three hundred.

How do I start?

Bring a design to JustOG — upload it, see it composited on the real garment, and publish it so other people can buy it.

Designs other people have published are in the shop.

The stack was never the brand. It was the toll booth in front of it.