Can AI-generated images be printed? Pixels and print size

Pixels decide the print size

A PNG or JPG has no physical size of its own, only pixels and an optional DPI field that tells software how large to show them. Set that field to 300 and the same pixels simply claim a smaller size. What prints is the effective resolution: the pixels across the image divided by the inches it prints at, with 25.4 mm to the inch.

Pixels      1024 × 1024
At 300 ppi  87 × 87 mm
At 150 ppi  173 × 173 mm
Across A4   210 mm → 124 ppi

How many pixels a page needs

The minimum depends on the job and on the printer. Work seen up close usually asks for 300 ppi; 150 ppi is the minimum Prepta’s free check uses; a banner read from metres away can need less, and Prepta’s built-in Large Format profile asks for 120. If the image runs to the edge, add the bleed as well, usually 3 mm on every side: an A4 page with bleed measures 216 × 303 mm.

The pixels each A size needs, before bleed:

Paper  at 150 ppi   at 300 ppi
A4     1240 × 1754  2480 × 3508
A3     1754 × 2480  3508 × 4961
A2     2480 × 3508  4961 × 7016

What the generators deliver

Generators document a default size and a larger option, and both change often, so check the pixel size of the file you downloaded rather than a table. Midjourney’s own guide does the arithmetic the right way round: 2048 pixels at 300 dpi make 6.8 inches. Its advice to set 300 dpi with resampling off changes the label, not the pixels; the print size still follows from the pixels. If a download arrives as WebP, which the free check does not read, save it again as PNG: that changes the container, not the pixels.

As each vendor documented it on 16 October 2026:

OpenAI API   1024 × 1024,
             1536 × 1024 or
             1024 × 1536; custom
             sizes up to 3840 px a
             side
ChatGPT app  sizes not stated
Midjourney   1024 × 1024 (SD);
             2048 × 2048 (HD or
             upscaled)
Gemini       1K by default
             (1024 × 1024); 2K and
             4K on request

A PDF, or a new DPI value, adds no detail

Placing an image in a PDF, or saving it as one, keeps its pixels; the page size then decides the effective resolution. Converting a PNG to a PDF is a change of container, not of quality, and the same holds for typing 300 into the DPI field.

Prepta’s test library holds the case: a fictional flyer built around a photograph made with an image model, exported at phone size and placed on an A4 page with 3 mm of bleed.

salt-market-flyer-A4.pdf
Image      612 × 866 px, filling the
           page
Page       A4 with 3 mm bleed
Effective  72 ppi
Required   150 ppi
Verdict    Blocked
Upscale    not offered: the photo is
           CMYK in the PDF, the
           upscaler works in RGB

Upscaling: more pixels, not more detail

An upscaler, including a model-based one, redraws edges at a larger pixel count. It can make a soft image look cleaner, but what it adds is a plausible guess: fine textures repeat, small text deforms and faces can change. Whether that is acceptable depends on the job and the viewing distance, which makes it a decision for a person, not a setting. Prepta’s own description of the operation says it raises the pixel count by redrawing edges and cannot recover detail the file never had, and its check after an upscale measures the resolution, not how the image looks.

Another library file, a fictional 2-metre roll-up banner whose picture was also made with an image model, shows a case where the arithmetic leaves little room:

kilnyard-rollup-850x2000.pdf
Banner     850 × 2000 mm (roll-up)
Image      1700 px across 860 mm
           (850 mm and 5 mm bleed a
           side)
Effective  50 ppi
Required   120 ppi (Large Format)
Verdict    Blocked
Offered    Upscale image, only once
           a person confirms it

Colour, lettering and the shape of the page

A PNG holds only RGB or greyscale, so a generated PNG reaches the printer as RGB. What happens next is the printer’s decision: a desktop printer usually takes RGB as it is, while a press that prints CMYK converts it through its own production condition, and saturated colours can shift. Lettering an image model draws is part of the picture: nobody can correct a misspelt word in it, and at a low effective resolution it prints soft, so set text that matters as live type in a layout application. Generated images also come in a set of fixed proportions, while A-series paper is about 1 : 1.41, so filling an A4 page means cropping the image or extending it, bleed included.