- Generators output for screens. A common 1024 by 1024 result has nowhere near the pixels a full front apparel print needs, and no amount of resizing invents them.
- One large print on demand network tells sellers to target 4500 by 5400 pixels for apparel and says most of its products need 150 to 300 DPI depending on the product and print area.
- Generators produce sRGB, made for backlit screens. Printing uses ink, so the brightest greens, oranges and blues in your design will come back duller than the mockup showed.
- The United States Copyright Office has said since March 2023 that human authorship is required, and prompts alone are not enough. You can sell the design. Registering it is a different question.
- Generated images cheerfully invent text and logo shapes. Anything word like on a product is a trademark question, and the free search at the patent office answers it in minutes.
- The single highest value habit is ordering a sample of every new design before a customer does. It catches resolution, colour and placement in one go.
The design took ninety seconds. The mockup looked better than anything you have made by hand. Then the first customer photograph arrives and the shirt in it has soft edges, a background that is not quite white, and a green that has gone the colour of an old road sign.
Nothing went wrong at the printer. The file was always like that, and the mockup was a picture of your file laid over a photograph, which is a rendering rather than a prediction.
Why does a generated image fail at print size?
Because it has too few pixels and there is no way to add real ones. Print quality is decided by how many pixels land inside each inch of the finished product, so the only question that matters is your pixel count divided by the physical size you are printing at.
Work it the way a printer does. A full front apparel print is commonly around 12 by 14 inches. At 300 pixels per inch that needs roughly 3,600 by 4,200 pixels. A square generation at 1024 by 1024 gives you, at that print size, somewhere in the region of 85 pixels per inch. It will look like a photocopy of a photocopy.
The numbers the platforms publish line up with that arithmetic. Printify's guidance on preparing low resolution AI art for print tells sellers to configure upscaling to output 4500 by 5400 pixels for apparel, and notes that most of its products need at least 150 to 300 DPI depending on the product and the print area. It warns plainly that an insufficient file produces blurry, pixelated prints.
| Print area | Pixels needed at 300 DPI | What a 1024 square gives you |
|---|---|---|
| Sticker, 3 by 3 inches | 900 by 900 | Comfortable, nothing to do |
| Mug wrap, 8 by 3.5 inches | 2400 by 1050 | Short, needs upscaling |
| Pocket print, 4 by 4 inches | 1200 by 1200 | Marginal, visibly soft |
| Full front apparel, 12 by 14 inches | 3600 by 4200 | About a quarter of the width needed |
| Poster, 18 by 24 inches | 5400 by 7200 | Not close |
Two practical notes on that table. The DPI figure is a means, not an end: what matters is the pixel count against the finished size, which is why the same file is fine on a sticker and unusable on a poster. And some products tolerate less than 300, because a fabric texture hides softness that a smooth mug surface reveals. Check the specification for the actual product rather than applying one number everywhere.
Does upscaling solve it?
Partly, and it is worth understanding what it is doing. A simple resize stretches the pixels you have and produces a larger blurry image. A generative upscaler invents plausible detail: it rebuilds textures and sharpens edges by guessing what should be there. Printify's guidance points at that second category for exactly this reason.
Guessing is the operative word. On a texture like fabric or foliage, invented detail is convincing and nobody will ever know. On anything with a defined shape, especially lettering, small faces or a logo, the upscaler will confidently produce something that is subtly wrong, and it is wrong at the size where a customer holds it eight inches from their eyes.
The better habit is to generate at the largest size your tool offers, at the aspect ratio of the print area, rather than generating a square and rescuing it afterwards. Upscaling is a repair. Starting at the right shape and size is a design decision, and it costs the same.
What happens to the colours?
They get quieter. Every mainstream generator outputs in sRGB, a colour space built for screens that emit light. A printed product reflects light instead, using a limited set of inks, and a meaningful part of what your screen can show simply has no ink equivalent. The conversion happens whether or not you manage it, and if you do not manage it, the printer's software decides on your behalf.
The colours that suffer most are the ones generators love: electric blues, neon greens, hot oranges, deep saturated purples. Those come back noticeably flatter. Meanwhile dark navy and near black areas tend to lose the separation between them, so a design with a black element on a very dark background can arrive looking like a solid blob.
You cannot fix this by making the image brighter. Pushing saturation on a colour the printer cannot reproduce produces the same in gamut result with worse contrast elsewhere. The fix is to design within what ink can do: fewer saturated colours, more contrast between the elements that need to stay separate, and a check of the converted version before you list.
There is a second colour problem specific to apparel that catches every new seller once. A generated image usually has a background, and even when it looks white it is often a very pale grey. Printed on a white shirt with direct to garment ink, that pale grey becomes a visible rectangle around your design. The background has to be genuinely removed, not merely made light, and the edge left behind has to be clean enough to survive being printed at full size.
Can you own a design you generated?
You can sell it. Whether you can register it, and therefore whether you can act against a copy, is a different question with a clear answer. The United States Copyright Office has taken the position since its policy statement of 16 March 2023 that copyright requires human authorship, that prompts alone are not sufficient creative input, and that where a work combines human and machine contributions, protection extends only to the human authored parts.
For a print on demand seller this has a specific and unromantic consequence. Your catalogue of generated designs is not a protected asset in the way a catalogue of drawings would be. If a competitor lifts your bestseller, your remedies are thinner than you assume, and the thing you actually own is the shop, the listings, the reviews and the customer list around it. We worked through the broader ownership position in the piece on selling what the AI made without owning it, and nothing in this year's guidance changes that shape.
The practical response is not to abandon generated designs. It is to stop treating them as the moat. Sellers who do well with print on demand at small scale are winning on niche, on listing quality and on the speed with which they iterate, none of which a copy of one image takes from them.
What about words and logos in the image?
This is the one that produces legal letters rather than refunds. Image generators produce text like shapes constantly, sometimes as gibberish and sometimes as recognisable words, and they reproduce logo like forms because logos were abundant in what they learned from. A product carrying a phrase or a mark is a trademark question, and trademark law does not care that a machine drew it.
The check is free and fast. The United States Patent and Trademark Office runs a public trademark search system and publishes guidance on clearance searching and on likelihood of confusion, which is the standard that decides whether two marks can coexist. Search the exact phrase, then search the obvious variants, then look at what goods the existing registrations cover, because a mark registered for software does not necessarily block the same word on a mug.
Two habits keep sellers out of this entirely. Read every word your generator put in the image, including ones that look like decoration, and delete anything you did not deliberately write. And treat common slogans with suspicion rather than affection: the short, punchy phrases that sell well on merchandise are exactly the ones somebody has already registered for apparel.
The check before you list
Four things, in order, on every design.
Pixels against the real print size. Not the DPI tag in the file, which can say anything. Open the image, note the pixel dimensions, divide by the print area in inches, and see whether the answer clears what the product needs.
Every letter and symbol in the design. Zoom to full size. Machine drawn text at thumbnail size looks like text and at print size looks like a mistake, and machine drawn marks that resemble something registered are a takedown waiting to be filed.
The converted colours. Look at the design in a form that reflects what ink will do, not at the glowing version on your screen. If your tool cannot show you that, assume the brightest areas will lose energy and design so it does not matter.
A physical sample. This is the check that subsumes the other three and the one most sellers skip because it costs a unit and a week. Order the first copy of every new design yourself. You will catch placement that sits too high, a print that runs too small on a size medium, and a colour that reads differently on grey marl than the mockup implied.
The returns arithmetic that decides whether this works
Print on demand hides its economics behind a low barrier to entry. There is no stock, so a failed design costs nothing to hold. What it costs instead arrives as returns, refunds and the reviews attached to them, and those land months after the design was published, which is why sellers rarely connect the two.
Work the numbers on a single unit rather than on a catalogue. On a typical apparel item the seller's margin is a few dollars after the base cost and the platform fee. A single return on that product wipes out the margin on several sales, because you refund the customer while the base cost has already been paid to the printer and the item cannot be resold. Quality problems in the print file are therefore not a cosmetic issue. They are the difference between a product line that makes money and one that runs a loss while looking busy.
This is what makes the sample habit pay for itself immediately. One sample costs you the base price of a unit. One avoidable return costs you the margin on several. If ordering a sample stops even a tenth of the designs you would otherwise publish, it has paid for itself many times over, and it also gives you a photograph of the real product, which converts better than any mockup.
The second lever is the listing. A large share of print on demand returns are not quality complaints at all, they are expectation mismatches: the colour looked different, the print was smaller than imagined, the fit was not what the size chart implied. Every one of those is fixed by information rather than by production. State the print dimensions in inches or centimetres, show the design on more than one garment colour, and photograph a real person wearing it if you possibly can.
How many designs should you actually publish?
Fewer than the tooling encourages, and this is where generation quietly works against a seller. When each attempt costs a minute, the obvious move is to publish hundreds and let the market choose. The trouble is that every published design carries an ongoing cost you do not see at publication: it can be reported, it can attract a takedown, it consumes your review time when a customer asks about it, and at volume it makes your shop look like a scraped catalogue rather than a maker's, which affects both search placement and conversion.
A workable discipline is to generate widely and publish narrowly. Produce fifty candidates, hold them, and publish the five you would order for yourself. Then judge those five in the market for a month before generating the next batch. The information you get from five listings you understand is worth more than the noise from a hundred you do not, and it keeps the trademark and quality checks in this article achievable rather than theoretical.
The sellers who get into trouble are almost always the ones who published faster than they could check. That is not a moral point, it is an operational one: the checks described here take about ten minutes per design, and ten minutes multiplied by two hundred designs is a working week nobody scheduled.
Where does AI genuinely earn its place here?
In the volume of attempts, not in the quality of any one attempt. The economics of print on demand have always been about how many designs you can test cheaply, and generation collapses the cost of an attempt from an afternoon to a minute. That is a real change and it is the whole benefit.
It also helps with everything around the design. Listing titles, variant descriptions, the alt text for each image, translations for a second market, the reply to a review about sizing. Those are text jobs on a surface you control, with a human reading the output before it ships, which is the profile of AI work that reliably pays. The same reasoning applies to your photography and mockups, which we covered in the piece on where AI product photos are allowed and where they must be labelled.
What it does not do is remove the craft at the edges: knowing that this print will crack on a ribbed fabric, that this placement will be cut by a seam, that this colour combination looks cheap in daylight. That knowledge comes from samples and from returns, and it is the part of the business that compounds.
Does the customer care that it was generated?
Some do, and disclosure rules increasingly decide it for you rather than leaving it to preference. Marketplaces have been tightening their requirements on declaring how a product or its imagery was made, and the direction of travel is towards more disclosure rather than less. The mechanics of how these images are produced, and what that means for what you are actually selling, are set out in the explanation of how an image generator works and what you own.
Commercially, the sellers who handle this best are straightforward about it and compete on something else: the fit, the fabric weight, the packaging, the fact that a real person answers an email about a return. A generated design is not a secret worth keeping and it is not a selling point either. It is a cheaper way to find out what people want, which is only valuable if you then do something good with the answer.
All of that lives or dies on a storefront that shows the product accurately and states the specification plainly, which is why the boring fields matter more here than in most retail. Print area, fabric, sizing, delivery estimate, return terms. Those are the fields a shop should carry by default rather than as an afterthought, and it is why we generate a storefront from a description of what you actually sell instead of dressing a template that was designed for something else.
The mockup is a rendering of your file, not a prediction of the product. The only honest preview is the one that arrives in a parcel.MaShop, 4 September 2026