On a Photograph Nobody Knows the Right Answer. On a Logo They Do

How it works

The file is enlarged four times by a superresolution model, then quantized to a small palette with dithering off, and saved. There is no diffusion model in the chain and no randomness anywhere in it, so the same input always produces the same bytes.

The quantization is the part that is specific to logos. General upscaling is tuned for photographs, where soft transitions are correct; on flat artwork those transitions are the defect, and snapping the result back to a few solid colors undoes them.

What the chain cannot do is anything about shape, which is the subject of the rest of this page.

Plausible is enough for a leaf and not for a circle

Superresolution works by having learned what detail usually looks like at a given scale, and adding detail of that kind. The question is whether anyone can check the answer.

On a photograph of foliage, no one can. There is a true arrangement of leaves that the low-resolution file did not record, and a plausible arrangement is indistinguishable from it to every viewer. The invention succeeds precisely because the ground truth is unavailable.

On a logo, the ground truth is available and often sitting in someone's inbox. The circle in the mark has a center and a radius. The rounded corner has a value in millimeters. The two strokes are the same weight because somebody set them to be. When the model produces a circle that is a fraction oval, the error is not invisible — it is measurable, and at large sizes it is visible to people who cannot say what is bothering them.

That is the whole argument of this page, and it is why the check it asks for is geometric: overlay a real circle, run a straight edge along the straight parts, compare stroke weight at the top and at the side.

Compression artifacts get promoted

The second failure has nothing to do with shape and is more common in practice, because of where small logos come from.

A logo pulled from a screenshot, a website, or a slide deck is usually a JPEG, and JPEG puts a faint halo of ringing around every hard edge. At original size, against a white background, that halo is nearly invisible.

Enlarge four times and the halo enlarges with everything else. Then quantize, and the halo — now several pixels wide — gets rounded to its own flat color, turning a soft artifact into a hard band that reads as a deliberate outline.

Which means the input format matters more than the upscaler does. A 240-pixel PNG enlarges better than a 400-pixel JPEG of the same mark, and it is worth going back to whoever sent the JPEG before running anything.

The fix that is not on this page

It is worth being blunt about the ordering, because people arrive here having skipped the two steps that actually work.

First, the vector file exists. Logos are drawn in vector tools; that is how they are made. Somebody has an EPS, an AI, a PDF, or an SVG — the designer, the agency, the person who commissioned it, or the brand guidelines PDF, which very often has the mark embedded as vector inside it. Ask.

Second, if it truly cannot be found, the answer is a redraw. A competent designer retraces the mark to correct geometry, and you end up with the file you should have had, which then works at every size forever.

This page is for the third case: you need a bigger raster today, for something that does not warrant either of the above. That is a real situation, and it is a narrower one than the traffic to pages like this suggests.

Publication gate

This page ships once the workflow has been run against a small flat PNG, a JPEG crop with visible ringing, a mark containing a true circle, and a mark with thin strokes, with circularity, straight-edge deviation, and stroke weight measured against the source geometry in every case.