Can you actually upscale a photo? What AI does and does not recover

Everyone has seen the scene: a detective points at a security still, says "enhance", and a reflection in someone's eye resolves into a licence plate. That is not possible and never will be. But "AI upscaling is fake" is the wrong conclusion to draw from it, because modern upscalers do something real and useful. The point is being precise about what.

The old way: interpolation

Doubling an image's dimensions means inventing three new pixels for every one you have. Traditional resizing does it arithmetically. Nearest-neighbour copies the closest existing pixel, which gives you visible blocks. Bilinear and bicubic average the neighbours, which gives you a smooth blur.

The mathematics guarantees the outcome: no new information is created. A bicubic upscale is the same picture, larger and softer. Fine detail does not sharpen, it spreads. This is why enlarging a small image has always looked bad — not a flaw in the algorithm, a property of averaging.

The new way: learned reconstruction

A super-resolution model is trained differently. Take millions of high-resolution photographs, shrink each one, and train a network to reproduce the original from the shrunken copy. Repeat until the network is very good at it.

What that network learns is not "how to enlarge". It learns what real detail looks like in the world — how a strand of hair behaves, how brickwork repeats, how an eyelash meets an eyelid, what a letterform does at a sharp corner. Then, given a small image, it does not average anything. It generates plausible high-resolution detail consistent with what it sees and with everything it has learned about how photographs are built.

Which means the honest description of the output is: the model is not recovering your detail, it is inventing statistically likely detail. That sounds damning. In practice, for the overwhelming majority of photographs, it is exactly what you want — because the invented detail is drawn from the same real world the photo came from, and the result genuinely looks like a sharper photo rather than a blurry one.

Where the distinction stops being academic

It matters enormously the moment the detail is information rather than texture.

Upscale a blurred number plate and you will get a crisp, confident, entirely fabricated set of characters. Upscale a face too small to identify and you get a crisp, confident face that is not that person's — this has been demonstrated repeatedly and embarrassingly. The model is doing exactly its job: producing the most plausible sharp thing consistent with the blur. Plausible is not the same as true.

So: use upscaling for photographs, never for evidence. If the pixels were not captured, the answer is not in there, and a model's confidence is not a substitute for having taken a better photo.

What it is genuinely good at

• Old scans and family photographs. A 900-pixel scan from 2004 becomes something you can print. This is the single best use of the technology.
• Small product images. The only copy of a product shot is 400px wide and the new site wants 1200. Upscaling handles it far better than bicubic ever did.
• Screenshots and low-res web images that need to appear at a larger size.
• Undoing compression damage. Many upscalers also clean up JPG blocking and ringing, because their training data taught them what those artefacts look like and what should be underneath.

And what it is not good at: anything where accuracy of small detail is the point. Documents where the text is already illegible. Faces at a scale where the features are a handful of pixels. Medical, forensic or measurement imagery, where an invented detail is worse than no detail at all.

Enhancement is a different job

Worth separating, because tools blur the two. Upscaling increases pixel dimensions. Enhancement keeps the dimensions and improves what is there — contrast, colour, noise, sharpness. Restoration repairs specific damage: scratches, dust, fading, tears.

They stack, and the order matters. Restore damage first — you do not want a scratch upscaled into a beautifully rendered high-resolution scratch. Then upscale. Then adjust colour and contrast last, on the final pixels.

We have separate tools for each: the image upscaler, the image enhancer and photo restoration. All three run the model in your browser, so the photo — often a family photo — is never uploaded.

Practical notes

2× is where the quality is. 4× and 8× are available and increasingly speculative; by 8× the great majority of what you are looking at was invented. If you need 4×, running 2× twice sometimes looks better than one 4× pass, because the second pass gets more to work with.

Feed it the cleanest source. An upscaler amplifies whatever is in the input, artefacts included. If you have the original rather than the version that has been through WhatsApp twice, use the original.

Check faces and text at 100%. These are where fabrication shows. Skin and foliage will look fine; a face will occasionally come back subtly wrong in a way that is obvious once you look at it directly.

It is not a substitute for reshooting. If you can take the photo again at a decent resolution, take the photo again. Upscaling is for the pictures that cannot be retaken — which, conveniently, is exactly the category most worth having.

The summary

"Enhance" as television imagines it is impossible: you cannot read information that was never recorded. But a model that has seen enough real photographs can produce a convincing, useful, printable enlargement of one that has gone soft — and for old scans and small originals, that is a genuine capability, not a trick. Just never treat the result as a record of what was there.