Print on Demand with AI: Designs, DPI and Real Costs

Most guides to AI for print on demand spend their time on prompt ideas and skip the thing that actually gets designs rejected: resolution.

A design that looks perfect on screen can be unusable on a t-shirt, because print platforms want 300 DPI at print size and a standard AI generation is nowhere near that for a full-front placement. Get that wrong and you either get rejected at upload or ship something that looks soft on the garment.

So this guide starts with the technical gate, then covers commercial-use rules, which models suit apparel work, how to produce designs at volume without destroying your margin, and where AI is genuinely the wrong tool.

Can you use AI images for print on demand?

Generally yes — you own the output of the models you run, subject to each provider's terms. The real constraints are platform rules rather than model licences: most print-on-demand marketplaces now require AI-generated content to be disclosed, and all of them prohibit infringing material regardless of how it was made.

Three separate things get conflated here, and they have different answers.

• The model's terms. Commercial use is normally permitted on paid access, but terms differ between providers and between free and paid tiers. Check the terms of the specific model you used, not a general article.
• The platform's rules. Marketplaces set their own policies on AI content, and several now require disclosure at upload. These change — read the current policy of the platform you sell on before listing.
• Infringement. Prompting for a copyrighted character, a trademarked phrase or a living artist's style does not become safe because a model made it. This is where most takedowns actually come from, and it has nothing to do with AI licensing.

What resolution do print-on-demand designs need?

300 DPI at the final print size. That makes pixel dimensions the thing that matters: 2,048 pixels is only 6.8 inches at 300 DPI, which is fine for a sticker or pocket print but too small for a full-front apparel design. Full front generally needs 4,000–4,500 pixels on the long edge.

The arithmetic is simple and worth internalising: pixels ÷ 300 = inches of print.

Individual platforms publish their own required dimensions — commonly in the region of 4,500 × 5,400 pixels for apparel — and those specs change. Treat the table above as the physics and the platform's own page as the requirement.

Two practical consequences. First, generate at 4K for anything going on a garment front, or generate at 2K and upscale before upload. Second, upscaling is a legitimate route — a dedicated upscaler adds pixels far more cleanly than exporting at a larger size. The print upscaling guide covers the DPI maths in more detail.

How do you get a transparent background for apparel?

Two routes: generate the design with a model that outputs a native alpha channel, or generate on a flat unused colour and key it out. Generating transparent is cleaner because there is no cutout step, and for apparel it matters — a white box behind your design prints as a white box.

This is where a lot of first-time POD designs fail. The design looks fine as a JPEG and then arrives on the garment sitting inside a visible rectangle.

• Generate with native transparency. GPT Image 1.5 accepts a transparent background setting and returns a real alpha PNG. No cutout, no halo, no edge cleanup.
• Or key out a flat colour. Generate on bright magenta or another colour that appears nowhere in the artwork, then chroma-key it. This is also how you get several designs from one generation — see below.
• Avoid cutting out a photographic background. If the design was generated on a textured or gradient background, removing it cleanly is much harder than generating it right in the first place.

The full comparison of removal methods and their costs is in the background removal guide.

How do you design at volume without destroying your margin?

Generate several designs on one sheet and cut them apart, rather than generating each separately. Six designs on one sheet costs one generation instead of six, and they come out visually consistent because they were rendered together — which matters when you are building a themed collection.

Print on demand is a volume business with thin margins per unit, so cost per design compounds in a way it does not for most creative work. Two habits do most of the work.

Sheets instead of singles

Space several designs apart on one flat unused colour, generate once, then chroma-key each into its own trimmed transparent PNG. In Rangy this extraction is free with the chroma engine, so you pay for the sheet and not for the objects. Six designs from one 4K generation costs about eight cents rather than roughly fifty.

The caveat: each design on a shared sheet is smaller than a full-frame generation, so this suits sticker packs, badge sets and small placements. For a full-front apparel print, generate that design alone at 4K.

Draft cheap, finish expensive

Explore concepts on a cheap fast model at low resolution — two cents a go — and only re-generate the ones you actually intend to list at 4K on a better model. Most designs you generate will never be listed; there is no reason to pay flagship rates to find that out.

Which AI model is best for print-on-demand designs?

It depends on whether the design has text. GPT Image 2 for anything with readable words, which is most apparel. Seedream for illustrative and painterly concepts. A cheap fast model for exploring directions. Nano Banana for editing a design you already have.

Rates via Kie.ai at 2K, August 2026, checked against Rangy's live pricing tables. 4K costs more; rates vary by provider and change over time.

Can AI put readable text on a design?

Yes, but only some models, and only for short strings. GPT Image 2 is the strongest at in-image text and handles a few words reliably when you put the exact wording in straight quotes. For anything longer, generate the artwork without text and set the type yourself in a real design tool.

This matters more for print on demand than almost any other use case, because slogan and typographic designs are a large share of what actually sells.

The honest limits:

• Short phrases work. Three to five words, quoted exactly in the prompt, on a clean background. Check every character before listing — a subtly misspelled shirt is a refund.
• Long text does not. Paragraphs, stacked multi-line quotes and small type break in every current model.
• Fonts are approximations. Models imitate typefaces; they do not use your licensed font file. If a specific typeface matters, set the type yourself.

The reliable professional workflow is a hybrid: generate the illustration with AI, then add the type in a vector or design tool where you control the font, kerning and exact spelling. You also get real vector text, which prints crisper than any raster approximation. The text-in-images guide covers the prompting side.

When is AI the wrong tool for print on demand?

Skip it when the design must be true vector for large-format or cut-vinyl production, when you need an exact licensed typeface, when you are chasing a trademarked or trending phrase, and when the market is already saturated with the same AI-default aesthetic.

I build a tool in this space, so weigh this accordingly — but four cases genuinely call for something else.

• You need real vector. AI outputs raster. Cut vinyl, embroidery digitising and some large-format work need true vector paths. Auto-tracing a raster design is a downgrade, not a conversion.
• Exact typography. Covered above — if the typeface is the design, set the type yourself.
• Trademark-adjacent ideas. Trending phrases are frequently trademarked. A model will happily render one, and the platform will happily remove your listing. This risk is unrelated to AI and entirely about what you ask for.
• Saturated aesthetics. If a category is already flooded with the same AI look, adding more of it is a volume game you are unlikely to win. Differentiation there comes from a distinct style or a niche, not from generating faster.

That last point deserves emphasis. AI removes the production constraint, which means it removes the constraint for everyone. The advantage moves to whoever has a genuine niche, a recognisable style, or a real understanding of a specific audience — not to whoever generates the most.

Frequently asked questions

The bottom line

Solve resolution first, because it is the only part that gets your design rejected outright. Generate at 4K for garment fronts or upscale before upload, output transparent PNG, keep text short or set it yourself, and batch designs onto sheets to keep cost per listing near nothing.

The interesting shift is that AI has removed production cost as a barrier to entry in print on demand — for you and for everyone else simultaneously. Generating a hundred designs is now trivial and therefore no longer an advantage.

What still is: knowing a specific audience, having a style people recognise, and getting the boring technical details right so your listings do not get rejected. The first two are the actual work. This guide only covers the third.