An upscaler does not recover detail. It invents plausible detail. Knowing the difference is what separates a usable enlargement from one that falls apart in print.
The short answer: an upscaler does not recover detail, it invents plausible detail. That is exactly what you want on skin, foliage and fabric, and exactly what you cannot use on text, logos, small faces or anything with legal weight. Upscale last, keep the factor at 4x or below, and check the result at 100%.
Every upscaler makes the same promise and it is worth being precise about what the promise is. It does not recover detail that was never captured. It produces plausible detail consistent with what was captured.
Most of the time that distinction is academic — plausible is exactly what you want. Occasionally it is the whole problem, because plausible skin texture on a portrait is fine and plausible text on a product label is a legal problem.
What is actually happening
Traditional resampling — bicubic, Lanczos — interpolates. It calculates new pixels from their neighbours. Enlarge far enough and everything goes soft, because there is no new information anywhere in the file.
A machine-learning upscaler was trained on a very large number of image pairs, each a high-resolution photograph and a downscaled version of it. It learned what kinds of detail typically correspond to what kinds of low-resolution pattern. Given a new small image it produces the detail that is statistically likely to have been there.
That is why an upscaled photograph can look genuinely sharp rather than merely large. It is also why it can be confidently wrong, and why the wrongness is hardest to spot in exactly the places you would most want it to be right.
What upscales well
Clean, well-exposed, in-focus originals. The model has more signal to work from and its guesses are better constrained.
Natural texture. Skin, foliage, fabric, stone, hair. If the reconstruction is slightly different from reality, nobody can tell, because nobody knows what that particular pore was doing.
Moderate factors. 2× is comfortable. 4× is often fine. 8× and beyond is where results start looking like a very sharp painting of your photograph rather than your photograph.
What upscales badly
Text and logos. The model produces letter-shaped marks that are not your letters. On a product shot, packaging or signage this is not a quality issue, it is an accuracy issue.
Faces at small pixel sizes. A face occupying 60 pixels does not contain a specific person. Upscaled, it becomes a plausible face — sometimes a noticeably different one. For a crowd in the background this is fine. For anybody the client can name it is not.
Fine repeating patterns. Fabric weave, mesh, brickwork, grilles. The model tends to regularise them into a cleaner pattern than the real one, which reads as artificial the moment you look for it.
Anything already damaged. Heavy compression artefacts, noise and motion blur all get reconstructed as if they were detail. Upscaling a bad file gives you a large bad file with more conviction.
How to test an upscaler properly
The honest test is easy and almost nobody does it:
- Take a high-resolution photograph you already have, of the kind you actually shoot.
- Downscale it by the factor you care about — say 4×.
- Upscale it back with the tool.
- Compare against the original at 100%.
You now have ground truth. You can see precisely what the tool invented and whether it invented it well. Compare a few tools this way on the same file and the differences that marketing pages describe in adjectives become differences you can see.
Do this with the difficult cases too — a frame with text, a frame with fine fabric, a frame with a small face — because that is where they diverge.
Rules that keep you out of trouble
Upscale last. After retouching, after grading, before export. Upscaling first means every subsequent edit is applied to invented detail, and any sharpening you do is sharpening the invention.
Never upscale for print without a proof. A screen at 50% forgives everything. Paper does not.
Do not upscale anything with legal weight. Product labels, documents, licence plates, evidence. The output is a plausible reconstruction, and that is the wrong category of thing entirely.
Check the face. Every time, at 100%. It is the one region a viewer will study.
When you do not need an upscaler
Two situations catch people out.
The first: you are enlarging because the client asked for a big print, but the file is already sufficient. Print viewing distance is longer than screen viewing distance, and the resolution genuinely required for a large print is lower than most photographers assume. Make a small test print at final size before you upscale anything.
The second: the frame is soft rather than small. Upscaling a soft image gives you a large soft image. That is a different problem, and no upscaler fixes it.
Where this fits with retouching
Resolution is not a substitute for finishing. An upscaled frame with uneven skin, a distracting background and a colour cast is a bigger version of an unfinished photograph.
If a frame is important enough to enlarge, it is usually important enough to finish properly first — which is why our high-end tiers deliver at full resolution for web or print rather than optimising for a screen. Get the picture right, then decide how large it needs to be.
Building it into a workflow
Treat upscaling as an output step rather than a repair step. It belongs after retouching and grading, immediately before export, in the same place you would put output sharpening.
Keep the native-resolution master. The upscaled file is a deliverable for a specific size, and if the size changes you want to regenerate it rather than upscale the upscale.
Decide your own ceiling by testing rather than by rule. Downscale a high-resolution frame you already have, upscale it back, and compare — you will find the factor at which your files stop reconstructing and start illustrating, and that number is more useful than anything a vendor publishes.
And check whether you need it at all. A test print at final size answers that question in an afternoon and often removes the step entirely.
Common questions
How much can I enlarge before it looks wrong? 2× is comfortable on almost any decent file, 4× is usually fine, and past 8× the result reads as a very sharp painting of your photograph. If you need more than 4×, do it in two passes and look at 100% in between — you will often find the first pass was enough.
Should I upscale before or after retouching? After. Upscaling first means every subsequent edit is applied to invented detail, and any sharpening you do is sharpening the invention.
Why does upscaled text look wrong? Because the model produces letter-shaped marks consistent with the pattern, not your specific letters. On packaging, signage or a brand name that is an accuracy problem rather than a quality one. Composite the original text back in at its own scale.
Do I need to upscale for a large print? Less often than photographers assume. Print viewing distance is longer than screen viewing distance, so the resolution genuinely required is lower. Make a small test print at final size before you upscale anything.
Can an upscaler fix a soft photograph? No. Upscaling a soft image gives you a large soft image. Softness is a different problem with a different set of fixes.
Related reading: How to Upscale a Photo Without Making It Look Fake and Best Free Image Upscalers.
BeEdits Editorial Team
Post-production studio
A photography post-production company: high-end retouching, creative editing and scalable production support for photographers, studios and brands.
