How to Upscale a Photo Without It Looking Worse

Enlarging a photo used to mean choosing between soft and blocky. A trained model does better — it reconstructs detail at the larger size instead of stretching what is there. It also has real limits, and knowing them saves disappointment.

The short answer

1. Open AI Enhance in PhotoTray.
2. Drop in a photo.
3. Choose 2× or 4×.
4. Let it run. The first upscale downloads the model once; after that it works offline.

No account, no credits, no per-image charge, no watermark, and nothing uploaded.

Why a plain resize looks bad

Enlarging an image means inventing pixels that were never recorded. Traditional scaling methods invent them by averaging neighbours — which is mathematically reasonable and visually poor.

Averaging produces exactly two failure modes, and every enlarged photo you have ever been disappointed by is one of them:

• Soft, where the averaging smooths a hard edge into a gradient.
• Blocky, where it does not smooth enough and you see the staircase.

There is no setting that fixes this, because the information genuinely is not in the file.

What a model does differently

A super-resolution model has been trained on an enormous number of image pairs — a full-size photo and a shrunken version of it — until it has learned what kinds of detail typically get lost when an image is reduced.

Given a small image, it produces the larger one that would plausibly have shrunk into it. Edges stay sharp because it has learned that edges are usually sharp. Texture stays textured. The result looks like a photograph rather than like an enlargement.

Note the word plausibly. This is the important caveat and it comes back below.

What it cannot do

Three things worth being clear about, because upscaling is routinely oversold.

It does not un-blur. Upscaling addresses resolution. A photo that was out of focus, or smeared by camera shake, is still out of focus at 4× — just larger and more obviously so.

It invents, and the invention is not evidence. The detail it adds is plausible, not recovered. It should never be used anywhere the actual content matters — reading a number plate, identifying a face, examining a document. What comes back is the model's best guess at what might have been there.

It struggles with faces at high factors. Human faces are the thing people look at most closely, and a general-purpose upscaler renders skin and eye detail that reads as slightly artificial. 2× is usually safe; 4× on a face frequently is not.

2× or 4×

2× is the reliable choice. It roughly doubles the linear dimensions, looks natural on almost any source, and runs in a fraction of the time.

4× works well on a clean, sharp, low-noise source. On anything compressed, noisy or slightly soft, it amplifies the flaws along with everything else and starts adding texture that was not there.

The general rule: upscale as little as you actually need. If the print is going to be A4, work out the resolution you need for A4 rather than reaching for the biggest number.

Why local, and why slow

Every online upscaler meters by the image, watermarks the free tier, or both — and requires you to upload the photo.

Running the model on your own machine has no per-image cost because there is no server. It also means the photo does not leave your PC, which for family pictures and client work is not a small thing.

The trade is speed. PhotoTray runs the model on your CPU, deliberately: it means no GPU requirement, no driver problems, and no machines where the feature simply refuses to work. It also means a 4× upscale of a 12-megapixel photo takes minutes rather than seconds.

PhotoTray shows the estimate before it starts the queue rather than during it, so you can decide whether a folder of them is an overnight job you want to start.

Get the input right first

Upscaling amplifies whatever it is given, so it is worth spending a moment on the source:

• Start from the largest original you have, not from a copy that has already been resized or emailed.
• Avoid a heavily compressed JPEG if a better version exists — the model will faithfully enlarge the compression artefacts.
• Convert to PNG first if you are going to do further editing, so the working copy stops accumulating loss.

Then upscale once, at the smallest factor that gets you where you need to be.