Upscaling AI Art for Print: The Arithmetic Nobody Shows You

Every image generator hands you roughly the same thing: a square or near-square image somewhere between 1024 and 2048 pixels on a side. Every wall-art listing wants something on a completely different scale — a 24 by 36 inch poster at 300 DPI is 7200 by 10800 pixels, which is seventy-four times the pixel count of a 1024 square.

That gap is the whole problem, and it is not really an upscaling problem. Most of it is decided before you type the prompt, by the aspect ratio you pick and the size you intend to sell.

The arithmetic, in one line

pixels needed = print size in inches × target DPI

That is the entire calculation, and it runs in both directions. Divide instead of multiplying and any file tells you the largest size it can carry: 4096 ÷ 300 = 13.7 inches.

Here is what the common poster sizes actually demand.

Read that table as a target rather than a hurdle. The number in the right-hand column is what the file has to be at the moment you upload it, and every earlier decision either moves you towards it or quietly rules it out.

How far short a raw generation falls

Put a typical generation next to those numbers.

A 2048 square is a coaster. Reaching 7200 × 10800 from it takes about 18 times the pixel count. Measured the way upscalers are sold — per edge, shortest side to shortest side — that is a little over 3.5×, and from a 1024 square it is 7×. Both numbers get worse once the aspect ratio takes its cut, which is the next section.

Two numbers are worth keeping separate in your head, because people conflate them and then argue past each other. A 4× upscale means 4× on each edge, which is 16× the pixel count. When a tool says 4×, it means the edge figure.

The aspect-ratio problem

This is the one that gets missed, because it does not look like a resolution problem and it never produces an error message.

A square does not crop into a poster shape for free. Cropping 1:1 down to 2:3 keeps the full height and only two thirds of the width, so you throw away a third of the image — and you throw it away from the left and right edges, which is exactly where a centred composition puts its breathing room.

The pixel loss compounds the upscaling problem, because you are now enlarging from a smaller starting image. A 1024 square cropped to 2:3 is 683 × 1024. Reaching a 24 × 36 inch poster from that is a 10.5× enlargement per edge rather than 7×.

The compositional loss is worse than the pixel loss. A generator composes for the frame you asked for. Ask for a square and it centres the subject in a square, fills the corners and balances the sides — then you cut both sides off and the balance is gone. You cannot crop your way back to a composition that was never made.

So the ratio is a prompt-time decision. Generate at 2:3 if you are selling 12×18 and 24×36; at 1:1.414 if you are selling A-sizes; and accept that those two families genuinely need separate generations, because 1:1.5 and 1:1.414 are about 6 per cent apart and that shows up as a visible crop at A2 and above. Midjourney for print-on-demand has the exact ratio flags for matching each size family.

When 150 DPI is genuinely fine

300 DPI is a convention with a reason behind it: it is comfortably past what the eye resolves at reading distance, about 12 inches. It is not a physical law, and it does not scale with print size the way people assume.

Nobody inspects a 24 × 36 inch poster from 12 inches away. It hangs on a wall and gets looked at from four to six feet, at which distance 150 DPI is indistinguishable from 300. That halves each edge and quarters the file: 3600 × 5400 instead of 7200 × 10800, which is a 4.3× upscale from a one-megapixel 2:3 generation rather than 8.7×. For AI art specifically, that is often the difference between an upscale that holds and one that visibly invents.

The honest rule: anything held in the hand needs the full 300; large wall pieces do not. The caveat is that your printer's minimum overrides your reasoning — many specify 150 DPI and reject below it regardless of viewing distance. DPI, resolution and colour for print-on-demand works through the viewing-distance table and the colour side of this in full.

Resampling versus model-based upscaling

These are different operations and they fail in different ways.

What naive resampling does

Bicubic, bilinear and Lanczos add pixels by interpolating between the ones already there. They have no idea what they are looking at. Every new pixel is a weighted average of its neighbours, so the output carries exactly as much information as the input, spread over more area. Edges become gradients. Texture becomes a smooth field. The result is not wrong, it is simply soft, and it gets softer in proportion to the factor. Above about 2× it is obviously mushy on paper.

Lanczos and Photoshop's Preserve Details variants are smarter interpolation, not different in kind. They hold an edge better and they still cannot add anything that was not captured.

What model-based upscaling does

A model trained on pairs of small and large images guesses what the missing detail should be. It does not interpolate, it reconstructs. Given a soft edge it produces a hard one; given a smudge of foliage it produces leaves. On photographic and painterly AI output this genuinely looks better, because the model was trained on broadly the same kind of content.

The catch is in the word guess. The model produces plausible detail, not correct detail, and plausibility breaks down exactly where the eye is most sensitive: faces, hands, text, and any structure that has to be regular. It also has no idea what your image means, so it will confidently sharpen a generation artefact into something that looks deliberate.

A practical compromise for AI art: upscale once with a model, at a factor between 2× and 4×, on the raw generation before any design work. Chaining two upscalers compounds both models' inventions, and upscaling a finished layout means the upscaler is also redrawing your typography.

Artefacts to look for at 100 per cent

Fit-to-screen hides all of this. A 7200 pixel wide file on a laptop is displayed at roughly a fifth of its size. Zoom to 100 per cent — approximately what the paper shows — and pan the whole image, not just the middle.

• Over-sharpened edges. A faint bright rim on the light side of a high-contrast boundary and a dark rim on the other. Halos are the most common tell of an over-aggressive upscale, and they print as visible outlines.
• Smoothed-out texture. Skin, fabric weave, paper grain and brush marks flattened into plastic. This is the opposite failure to halos and often appears in the same file — hard edges over-sharpened, flat areas scrubbed clean.
• Invented detail in faces. Eyes that acquire lashes that were not there, teeth that resolve into a regular grid, hair that turns into fine parallel strands. At 100 per cent it reads as uncanny; at poster size it reads as a mistake.
• Mangled lettering. Any text the model touches gets its letterforms rounded off into near-letters. This is why type should be set by hand after upscaling, never prompted into the image and never enlarged with it.
• Banding in skies and gradients. Smooth tonal ramps break into visible steps, because the same number of tonal levels is now spread across four times the area. Work in 16-bit through the edit where you can, add a very slight grain to gradient areas, and export at high quality rather than the default.
• Repeating texture. Some upscalers tile their invented detail, producing a faint regular pattern across large flat regions. Easiest to spot on walls, water and skies.

If any of these are visible at 100 per cent, they will be visible on the wall. Nothing later in the process removes them.

When to regenerate instead

Upscaling is the wrong tool more often than people admit. Regenerate when:

• The aspect ratio is wrong. Always. Cropping and stretching both cost more than a regeneration does.
• The required factor is above roughly 4× per edge. Beyond that the model is producing most of the image rather than enlarging yours.
• The subject is a face, a hand, or anything with regular structure. Small errors get magnified into large ones, and upscaling will not fix an anatomy problem — it will render it at higher resolution.
• There is text in the image. Regenerate without it and set the type yourself.
• The generation is already soft or noisy. An upscaler amplifies whatever it is given, including flaws. It has no concept of a mistake.

The general principle: pixels you never captured cannot be recovered, only fabricated. Given the choice between fabricating them and generating properly the first time, generate properly. Mastering artwork at the largest size you sell is the same argument applied to the file you keep afterwards.

A workflow order that holds up

Order matters more here than tool choice.

1. Decide the largest size you will sell, and work out its pixel dimensions at the DPI your printer requires. That is the target, and every earlier step serves it.
2. Generate at the matching aspect ratio. Not the default square, and not something you plan to crop.
3. Pick the frame, then upscale once, on the raw image, at the smallest factor that reaches the target.
4. Check at 100 per cent across the whole image, looking for the six artefacts above.
5. Then compose — borders, type, signature — at full pixel dimensions, so everything you add is genuinely sharp rather than enlarged along with the rest.
6. Export sRGB at high quality, and read the exported file's dimensions rather than assuming they survived.
7. Verify against every size you enable, not just the one you expect to sell first.

PSDmate's print-file checker handles step seven: it reads a file's real pixel dimensions and tells you which print sizes it can honestly support, which is faster than doing the division by hand for nine sizes.

The short version

Pixels equal inches times DPI, and a 24 by 36 inch poster at 300 DPI wants 7200 by 10800 of them — roughly seventy times what a square generation gives you. No upscaler closes that gap on its own.

What closes it is the decision made before generating: the right aspect ratio, the largest size fixed in advance, and an honest choice between 300 and 150 DPI. Upscale once after that, check at 100 per cent, and set the type yourself. If the art comes out of Midjourney, PSDmate can pull those generations straight into a mockup run, which removes the manual step between the image and the listing.