Upscale an Image with AI

How it works

The upload is handed to Real-ESRGAN, a reconstruction model trained on pairs of degraded and clean images. It does not stretch pixels. It predicts what a larger version of this image would have looked like, one tile at a time, and assembles the tiles into the output.

The model has one fixed output scale: four times the input, in each direction. That is a property of the trained network, not a setting. So when you ask for a smaller factor, the graph still runs the full four times enlargement and then scales the result down to your chosen size with a Lanczos filter.

That detail is worth knowing because it explains the quality difference between the presets.

Why the smaller factor looks better

A four times upscale shows you the model's raw output. Every pixel in it was predicted, including the ones it got wrong, and reconstruction errors are visible at full size.

A two times upscale takes that same output and halves it. Averaging four predicted pixels into one is a very effective way of hiding a prediction that was slightly off — errors that disagree with each other cancel, and detail that the model was confident about survives. The result is not just smaller than the four times output, it is measurably cleaner.

So the useful rule is to pick the smallest factor that gives your layout the pixels it needs, rather than the largest one available. Going bigger does not get you more real detail; past a point it only shows you more of the guess.

Where the invention shows first

Look at these before you use the output anywhere it matters.

Text is the fastest tell. The model has seen a great deal of text and will happily produce letterforms that are crisp, confident, and not the characters that were in your source. Small type on packaging, a license plate, a street sign, a name badge — check every one of them against the original.

Faces at small sizes are the second. A face occupying forty pixels does not contain enough information to reconstruct, so what comes back is a plausible face rather than that person's face. It will look sharp, which is exactly the problem.

After those, look at repeating fine structure: fabric weave, brickwork, foliage, hair. The model tends to regularise these into a pattern that is tidier than reality, which reads as slightly synthetic once you notice it.

When the browser resizer is the right tool

If you need an image at specific dimensions and the source already has enough detail, use the ordinary resizer instead. It is instant, it runs on your own machine, and every pixel in the output is derived from a pixel you supplied.

Reconstruction is for the case where the source genuinely does not carry the detail the final use needs, and where you are willing to review what came back. A photo destined for print at A4 from a small web-sized original is a reasonable candidate. A screenshot you want at exact dimensions is not.

Checking the output

Compare at the same displayed size, not at the same zoom level. Enlarging a source and putting it next to its upscale at one hundred percent compares two different things and always flatters the upscale.

Then check the file. A four times enlargement of a two megapixel photo is a thirty-two megapixel image, which is slow to place in a design tool and may be rejected by an upload form with a dimension limit. Crop after upscaling rather than before, so the model works from every pixel the camera recorded.