Upscaling a heavily compressed JPEG without cleaning it first, which sharpens the artefacts along with the subject.
AI and on-device processing
Super-resolution
Also called ai upscaling, sr or neural upscaling. Here is what it means, when it changes what you export, and what people get wrong.
AI and on-device processingRuns on your deviceUpdated 10 August 2026
The short answer
Super-resolution uses a trained model to produce a higher-resolution version of an image, adding detail that plain interpolation cannot. It does not recover the original detail, because that information was lost. It predicts what plausibly belonged there, which is convincing for texture and dangerous for evidence.
Updated
The short version
Super-resolution at a glance
| Quick fact | Detail |
|---|---|
| Input | One low-resolution image, or several frames |
| Common factors | 2x and 4x |
| What it adds | Plausible detail, not recovered detail |
| Best on | Natural texture at modest enlargement |
| Worst on | Text, faces used to identify, heavy artefacts |
| Honest alternative | Bicubic or Lanczos when fidelity matters most |
Slide the table sideways to see every column.
How is it different from ordinary upscaling?
Classical resampling, bicubic or Lanczos, works out each new pixel by weighted averaging of the pixels around it. It is fast, predictable and honest: it never introduces anything that was not implied by the neighbours. That is also why enlargements look soft, because averaging cannot create high-frequency detail.
A super-resolution model has been trained on millions of pairs of low and high resolution images, so it carries a learned prior about what a blurry patch of brick, skin or fabric normally looks like when sharp. Applied to your photo, it synthesises detail consistent with that prior. The result looks sharper because it genuinely contains more detail. That detail is invented.
Why is single-image super-resolution fundamentally ill-posed?
Run the maths backwards. Downscaling averages many pixels into one, throwing information away irreversibly. That means an enormous number of different high-resolution images reduce to exactly the same low-resolution one, and nothing in the small image can tell you which of them you started with.
So the model is not solving for the truth, it is choosing one plausible answer from a huge set of candidates, weighted by what it saw during training. It will make that choice confidently every time. This is also why two models, or the same model at different settings, give different plausible faces from the same blurry face, and none of them is a recovery.
Multi-frame super-resolution is a different case. When you have several slightly offset frames of the same scene, each captures a marginally different sample of it, and combining them adds real information rather than guessing at it.
What should you never do with an upscaled image?
Treat it as evidence. If a face, a number plate, a serial number or a line of document text was unreadable in the source, an upscale has not read it, it has written a legible-looking guess. The output can be sharp, plausible and completely wrong, and its confidence is indistinguishable from accuracy by eye.
The practical rules follow from that:
- Never identify a person from an upscaled face
- Never transcribe text or a number that was illegible before
- Never submit an upscaled image where the original is required, including identity and passport photographs
- Keep the original whenever the upscale might later need to be checked
Where does it work well in practice?
The useful cases are the ones where nobody is relying on the detail being true, only on it being believable. Modest enlargement, roughly two times, of natural texture is where it is strongest: fabric, wood grain, foliage, skin at normal viewing distance. Product photographs shot too small for a marketplace minimum, older images that need to fill a modern layout, and prints that need more pixels than the file has all qualify.
It is weakest on images that were already damaged. Heavy JPEG artefacts, banding and noise are structure as far as the model is concerned, so they get amplified and baked in rather than removed. Motion blur and out-of-focus regions are similar: there is nothing there to reconstruct, so the model invents plausible sharpness that does not match the real subject. Clean the source first, then upscale, and be more conservative with the factor than the settings tempt you to be.
What people get wrong about super-resolution
Each one is a real failure mode, not a style preference.
Reaching for 4x when 2x was enough, since error compounds with the factor and skin and fabric start looking plastic.
Reading a number plate or serial number off an upscale and treating the result as what the camera saw.
Upscaling a passport or identity photo to meet a pixel minimum, when the guidance expects an original of sufficient resolution.
Questions people ask
Super-resolution, answered
The follow-up questions people search for once they have the definition.
Can AI upscaling recover detail that was never captured?
No. It generates detail that is statistically consistent with what it learned, which looks like recovery but is prediction. If the information was lost at capture or in compression, no model can retrieve it. The output can still be genuinely useful, as long as you know what it is.
Why do faces sometimes look wrong after upscaling?
Because a face is the case where you have the strongest prior of your own. Small invented differences in eyes, teeth and skin texture that you would never notice on a brick wall read immediately as uncanny on a person. Lower factors and gentler models help.
Should I upscale before or after cropping?
Crop first, then upscale. The model works on what you give it, so upscaling the full frame and then cropping wastes effort on discarded pixels and can leave the kept region at a lower effective quality than if it had been the whole input.
Is 4x always better than 2x?
No. Larger factors mean the model invents proportionally more of the image, so texture drifts further from the original and errors become more visible. Choose the smallest factor that reaches the size you actually need, rather than the largest the tool offers.
Same cluster
More from the glossary
Neighbouring entries, so every term in the set is one hop from every other.
SVG
SVG (Scalable Vector Graphics) is an XML file that describes shapes, paths and text as mathematical instructions rather than pixels.
sRGB
sRGB is the default colour space of the web and of most consumer cameras, phones and monitors.
Thumbnail
A thumbnail is a small version of an image used in search results, grids and previews.
From the studio
Tools that do this
Each one runs on your own device, so the photo stays where it is.
Keep reading
Elsewhere on Photomop
Platform image rules
Image size, file type and background rules for 22 selling platforms, each one sourced.
Read itConvert image
JPG, PNG, WebP and AVIF, with an honest answer about HEIC and alpha.
Read itAll the tools
The whole toolbox in three families, with what each one needs before it can run.
Read itOn-device privacy
The pipeline in five steps, what leaves your device, and how to check it yourself.
Read it
Now do it to a photo
Photomop is a photo studio that runs on your own device. Resize, crop, compress, convert, batch edit, remove a background and change one all work in the browser tab you are reading this in, at full resolution and with no watermark. The on-device page shows you how to check that the photo stays put.