Do AI edits hold up in print?

What actually falls apart when the image gets big

Synthesised texture is the first failure. Where a model has reconstructed skin, fabric, foliage or stone, the result tends to be locally plausible and globally repetitive, with a softness that has no grain structure behind it. At phone size the eye reads it as detail. At 24 inches the eye reads a smooth region that should have had texture, and the picture looks slightly out of focus in a way that no sharpening fixes because there is nothing underneath to sharpen.

Hair and fine edges are the second. Anywhere a subject meets a changed background, the boundary is a reconstruction, and hair strands are the hardest thing in photography to reconstruct convincingly. Look for strands that terminate abruptly, clumps that merge into a soft mass, and a faint halo of background colour running along the shoulder line. All three are close to invisible at 800 pixels wide and unmistakable on a print.

Then the smaller ones that still ruin a sale: skin that has lost pore structure and reads as plastic under gallery lighting; jewellery and glasses frames subtly warped where the model resolved an edge; and fine text on signage, packaging or clothing that has become approximate lettering. Text is the fastest tell of all, because a viewer does not need any technical knowledge to see that a word is not a word.

• Soft synthesised texture with no underlying grain
• Hair strands merging into clumps, plus a colour halo at the subject edge
• Pore structure erased, skin reading as matte plastic
• Warped jewellery, spectacle frames and repeated patterns
• Fine text turned into approximate letterforms
• Background geometry that does not stay consistent under scrutiny

Resolution, interpolation and what upscaling really gives you

Print resolution requirements depend on viewing distance rather than on a universal number, which is why a billboard works at very low pixel density and a 12-inch album spread does not. For a print someone will hold or stand close to, most labs want substantially more data than a web-sized file, and the sensible move is to check your specific lab's current recommendations rather than relying on a figure you memorised years ago, because lab guidance and printer capabilities change.

AI upscaling does help, and it helps most on subjects with forgiving texture: landscapes, architecture, anything where invented detail is not compared against a known reality. It helps least on faces, where a viewer holds a lifetime of expectations about what skin and eyelashes look like. An upscaler generates a larger file with synthesised detail; it does not recover information the sensor never captured, and enlarging an already-generative edit compounds two rounds of invention.

The practical sequence matters. Do generative editing before upscaling, not after, and evaluate the result at 100 per cent at final output size rather than fit-to-screen. Most photographers who get caught out never looked at the file at the size they sold.

Colour: screen to paper is a separate problem

Even a technically perfect edit shifts when it moves to paper. Prints are reflective and cannot reach the luminance of a backlit display, so deep shadows tend to block up and bright, saturated regions compress. Blues and vivid greens are the usual casualties, and skin tones drift in ways that are subtle on a screen and obvious next to a person.

None of this is AI-specific, but generative edits make it worse in one particular way: the synthesised regions often carry a slightly different tonal character from the captured ones, and paper's narrower range exaggerates that difference. A background that blended fine on screen can read as a separate layer once the tonal range compresses.

The unglamorous answers still work. Calibrate the display you edit on, soft proof to the lab's profile for the paper you are actually using, and ask the lab for their current profile rather than assuming last year's. Where a print matters, order a small test print on the same paper before committing to size.

Test print before you sell it

For anything going into a saleable print, an album, a gallery show or a client's wall, print a test. You do not need the full size: an 8x10 crop of the most demanding region of the image at final output scale tells you almost everything, and costs a few pounds. Choose the region with hair against background, skin at the edge of light, or fine texture, not the region you like.

Look at it in the light it will hang in, not at your desk. Gallery spotlights, domestic tungsten and north-facing daylight each reveal different problems, and reflective paper under a spot is where plastic skin becomes undeniable. Look at it from arm's length and from six inches, because a buyer will do both.

If it fails, you have three options in order of preference: use the original unedited frame, redo the edit more conservatively, or reshoot. Selling a print you have doubts about is the only genuinely bad option, because prints are permanent, they hang in rooms, and they are the version of your work that people see for years.

Choosing the right tool for the size

The honest division is by output size and scrutiny. For social, web, small prints up to modest sizes, and client previews, generative editing is a reasonable choice and often the fastest one available. Foxi AI handles a single frame in under a minute for a few credits, which is genuinely useful for a hero image that needs a cleaner background before it goes on a site or a card.

For large prints, albums, exhibition work and anything a buyer will examine at close range, hand techniques remain the right answer for the parts that matter. Manual masking, frequency separation, dodge and burn, and careful compositing are slow precisely because they operate on real captured pixels, and real pixels are what survives enlargement. There is no shortcut here worth taking, and a photographer selling wall prints should plan their time accordingly.

A useful middle path is to keep two versions of a strong frame: a generatively cleaned version for screen use, and a hand-finished version for print, both stored with the original. It costs a little extra work on your best images only, and it means the answer to 'can I get this at a metre wide?' is yes rather than a wince.

Setting print expectations with clients

If a client has seen an AI-edited version on screen and asks for it large, say what the constraint is before taking the order. 'That version was finished for screen. For a print at that size I would want to redo the finish by hand, which is additional work, or we can use the original.' This is a normal professional conversation and clients accept it readily when it comes before the purchase.

Be careful about which files are in the gallery a client orders prints from. If your gallery service sells prints directly, a screen-finished edit sitting in there can be ordered at 30 inches with no conversation at all. Either keep those files out of print-enabled galleries or cap the available sizes.

And never promise that an edit will look identical in print to how it looked on their phone, because it will not, and that is true of every photograph regardless of how it was made. Explaining the reflective-versus-backlit difference once, at the point of sale, prevents a return conversation that is much harder to have after the print is on the wall.

Frequently asked