An upscaler does not recover detail; it invents plausible detail. Knowing which subjects survive that, and how to test it, is how you choose one.
The best AI image upscaler is the one that enlarges your kind of subject convincingly at the size you need, and the only reliable way to find it is to test on your own files. All upscalers invent plausible detail rather than recovering real detail, so they do well on skin, foliage and fabric and badly on text, logos, small faces and fine repeating patterns. Beyond that, choose between desktop software, a cloud service and a local open model on privacy, batch size and cost.
Every upscaler makes the same promise, and it helps to be precise about it. It does not recover detail that was never captured. It produces plausible detail consistent with what was captured.
Most of the time the distinction does not matter, because plausible is what you want. Occasionally it is the whole problem: plausible skin texture on a portrait is fine, and plausible lettering on a product label is a legal risk.
What an upscaler actually does
Traditional resampling methods such as bicubic and Lanczos interpolate: they calculate new pixels from their neighbours. Enlarge far enough and everything goes soft, because no new information enters the file.
A machine-learning upscaler is trained on pairs of images, each a high-resolution photograph and a reduced copy of it. It learns which kinds of detail usually correspond to which low-resolution patterns, and 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 sharp rather than merely large. It is also why it can be confidently wrong, and why the errors are hardest to spot exactly where accuracy matters most.
Subjects that upscale well
Clean, well-exposed, in-focus originals. The model has more signal and its guesses are better constrained.
Natural texture. Skin, foliage, fabric, stone, hair. If the reconstruction differs slightly from reality, nobody can tell, because nobody knows what that particular pore looked like.
Moderate factors. 2× is comfortable. 4× is often fine. At 8× and beyond, results start to look like a very sharp painting of your photograph rather than the photograph.
Subjects that upscale badly
Text and logos. The model produces letter-shaped marks that are not your letters. On packaging or signage that is an accuracy problem, not a quality one.
Small faces. A face 60 pixels wide does not contain enough information to identify a specific person. Upscaled, it becomes a plausible face, sometimes a noticeably different one. Fine for a crowd in the background; not for anyone the client can name.
Fine repeating patterns. Fabric weave, mesh, brickwork, grilles. Models tend to regularise them into a tidier pattern than the real one, which looks artificial once you notice it.
Damaged files. Compression artefacts, noise and motion blur are reconstructed as if they were detail. Upscaling a bad file produces a large bad file with more conviction.
Testing an upscaler against ground truth
The reliable test is simple, and hardly anyone runs it:
- Take a high-resolution photograph you already have, of the kind you actually shoot.
- Reduce it by the factor you care about, say 4×.
- Upscale the reduced copy back with the tool.
- Compare the result with the original at 100%.
Now you can see exactly what the tool invented and whether it invented it well. Run two or three tools on the same file and the differences that marketing pages describe in adjectives become differences you can point at.
Include difficult frames: one with text, one with fine fabric, one with a small face. That is where tools diverge.
Desktop application, cloud service or local model
The form an upscaler comes in matters as much as the model inside it.
Desktop applications run on your own machine, handle batches and keep files local. You pay a licence and need capable hardware: large files at high factors want a strong graphics card and plenty of memory, or a lot of patience.
Cloud services need no hardware and are quick to try. Every file is uploaded, output size and monthly volume are often limited by plan, and you should read what the service does with uploads before a client's photograph goes near it.
Open models run locally cost nothing to use and upload nothing, but take real setup and the same hardware as a desktop application.
Match the form to the work. Occasional web-sized enlargements suit a cloud service. Regular print jobs, files under a model release, or batches of dozens suit something that runs locally.
When you do not need an upscaler
Two situations catch people out.
The client wants a large print, but the file is already big enough. Prints are viewed from further away than screens, and the resolution a large print needs is often lower than expected. Make a small test print at final size before upscaling anything.
Or the frame is soft rather than small. Upscaling a soft image gives you a large soft image, which is a different problem with different fixes.
Upscaling comes after finishing
An upscaler should be one of the last steps, after retouching and grading, with no sharpening added on top; How to Upscale a Photo Without Making It Look Fake sets out the order.
Resolution is not a substitute for finishing either. An enlarged frame with uneven skin, a distracting background and a colour cast is a bigger unfinished photograph. The Editorial / Campaign package of High-End Photo Retouching, at $10.50 per image, is finished for web or print delivery, which makes it the version worth enlarging.
Building a shortlist in one afternoon
- Pick three candidates, ideally one of each form: desktop, cloud and local.
- Make a test folder of four high-resolution frames you own: a portrait, fine fabric, a frame with lettering, and a small face in the background.
- Run the ground-truth test on all four with each tool, at the factor you most often need.
- Check each output's pixel dimensions, file format and embedded colour profile, not only how it looks.
- Time a batch of twenty frames, including uploads and downloads for the cloud option.
- Read each tool's terms on uploads and file retention.
Judge the tools on your most difficult frame, not your easiest. That is the frame a client job will eventually hand you.
Common questions
How far can I enlarge before it looks wrong? 2× is comfortable on almost any decent file and 4× is usually fine; beyond 8× results tend to look painted. If you need more than 4×, enlarge in two passes and check at 100% in between.
Does an upscaler keep my colour profile and metadata? Not always. Check an output file's embedded profile and metadata before relying on a tool, because a silent conversion or a dropped profile changes colour, especially in print.
Why does upscaled text look wrong? The model draws marks consistent with letter shapes rather than your actual letters. On packaging, signage or a brand name, composite the original lettering back in at its own scale.
Do I need to upscale for a large print? Less often than you might think, because prints are viewed from further away than screens. Ask your print lab for the pixel dimensions it wants, and make a small test print at final size before enlarging anything.
Can I upscale a JPEG straight from a phone? You can, but compression artefacts and in-camera sharpening get rebuilt as if they were detail. Use the least-processed version of the file you have.
Related reading: Best Free Image Upscalers and RAW vs JPEG for Professional Retouching.
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.
