Upscale an Image 8x

How it works

Two stages, and they are not equivalent.

First, RealESRGAN enlarges the image four times in each direction. This is the model's native factor and the stage where new detail appears — the network has learned what photographic texture looks like at higher resolution and fills in accordingly.

Second, that result is doubled again with lanczos resampling. Lanczos is a mathematical filter. It computes each new pixel by weighting the pixels around it, which produces a smooth, clean enlargement and creates nothing that was not already implied by the four times output.

Eight times overall. One model pass, one resample.

What the second stage is worth

It is worth being plain, because "8x" reads as twice as good as "4x" and it is not.

Interpolation cannot add information. Given a 4x image, lanczos produces the most reasonable 8x version of that same image — smoother than nearest-neighbor, less soft than bilinear — but every detail in it was decided by the model pass before it. Zoom to one hundred per cent on an 8x output and you are looking at 4x detail rendered across four times the area.

That is not a criticism of the approach. It is how essentially every high-factor upscale works, including the ones that do not tell you. Models have fixed native factors; anything beyond them is either repeated passes, which compound the model's errors, or resampling, which does not.

When dimensions are the actual requirement

The useful framing is to separate two different needs that both sound like "make it bigger".

The first is wanting a sharper image. For that, 4x is the answer and 8x adds nothing but file size and processing time.

The second is wanting an image that satisfies a specification. Large-format print services publish minimum pixel dimensions for each physical size, and their upload checks enforce them. A three meter banner may require a file no 4x output of a small source can reach. Here the requirement is genuinely dimensional — the print will be viewed from several meters away, where 4x-level detail is far more than the eye resolves, and the pixel count is what the machine needs.

Flat artwork is the other good case. Logos, diagrams, and vector-style graphics have hard edges and little texture, so there is not much for the model to invent and interpolation handles the enlargement cleanly.

Judging the result

Do not evaluate this at one hundred per cent. Almost nothing is ever seen that way at these dimensions, and inspecting a sixty-four times enlargement pixel by pixel will always disappoint.

Judge it at the size it will be used. Print a section at the final scale, or view it at the display size it is destined for. The question is whether it holds up there, not whether it survives a microscope.

Publication gate

This page ships once the workflow has been run against flat vector-style artwork, a photographic source, and a source large enough to test the memory ceiling, with each output judged at its intended output size rather than at full zoom.