Upscale an Image 2x

How it works

Two steps, and the second one is the interesting one.

RealESRGAN reads your image and enlarges it four times in each direction. This is the model's native factor — the only one it does — and the stage where new detail is reconstructed, because the network has learned what photographic texture looks like at higher resolution.

That four times result is then scaled by half using lanczos resampling, landing on exactly double your original dimensions.

Why the detour is the feature

Downscaling a generated image is called supersampling, and it is the oldest quality trick in rendering.

When four pixels become one, their differences average out. Consistent detail — a real edge, a genuine texture — survives the average because all four pixels agree about it. Inconsistent detail does not. Ringing along a hard edge, the faint checcurboard some upscalers leave behind, invented pores on skin that never had them: these are exactly the sort of high-frequency disagreement that averaging destroys.

So the model gets to be confident and the resample gets to be skeptical. What reaches you is the part both agreed on.

This is why a 2x output can look cleaner than a 4x one examined at full size, even though it came from the same model pass. It is not doing less work. It is discarding the least reliable quarter of that work.

What it costs

The whole four times enlargement is computed before anything is reduced.

That means processing time and memory are identical to the 4x page. You are not buying a cheaper job by asking for a smaller output — you are spending the same compute and using it to buy cleanliness instead of size.

Whether that is a good trade depends entirely on what happens next. For a file going into a layout at a known size, it is. For a source so small that every generated pixel matters, it is not, and the 4x page is the better page.

When double is the right number

Doubling suits images that are already close.

A photo that is slightly under a print requirement, a screenshot that needs to survive a retina display, a product shot that is fine at its current size and needs headroom for a crop — these want a modest, clean enlargement rather than the largest one available.

The failure case is the opposite: a genuinely small source, a thumbnail, a tiny avatar, an image saved at 400 pixels years ago. Doubling that gives you 800 pixels, which is usually still not enough, and you will end up running the 4x page afterwards anyway.

Judging the result

Do one comparison and it will tell you everything.

Open your original, enlarge it to the same doubled size with ordinary bicubic interpolation in any editor, and put the two files side by side at one hundred per cent. On a clean source the gap will be small. On a compressed or soft source the model's reconstruction of edges and texture will be obvious.

If you cannot see a difference, use the resize. Nothing here is worth waiting for a GPU when a filter already answered the question.

Publication gate

This page ships once the workflow has been run against a compressed phone photograph, a clean high-quality source, and a portrait with visible skin texture, with each result compared against a plain bicubic doubling at full zoom.