What an upscaler invents

Two different operations with one name

Interpolation — bicubic, Lanczos — estimates intermediate pixels from neighbours. It cannot add information. A blurred photograph enlarged this way is a larger blurred photograph. Every image editor has done this for thirty years.

Generative upscaling does something categorically different. A model trained on pairs of small and large images predicts what the large version would look like. Where the small image has a grey smudge, the model produces the texture that most often appears where such a smudge appeared in training: brick, or fabric, or an eye.

The output looks sharper because it is sharper. The detail is real detail of a plausible image. It is not detail of your image, because that information was never in the file.

Why this matters more than an aesthetic preference

Two consequences.

For creative work, it is straightforwardly good. If you generated the picture, invented detail is no less legitimate than the rest of it. Upscaling a generated image to print size is the normal final step and there is nothing to worry about.

For anything evidential, it is dangerous. A CCTV frame upscaled produces a face. The face is what a model thinks belongs in a smudge of that shape. It is not the person's face, and it can be a different person's face entirely. Published demonstrations have shown upscalers turning low-resolution images of well-known faces into confidently wrong different people, and the failure is not a bug — it is the model doing exactly what it was trained to do.

The rule worth carrying: an upscaler is a plausibility engine, not a recovery engine. Never present an upscaled image as evidence of what was in the original, and be sceptical when anyone does.

The families, and when to use each

Convolutional upscalers — the ESRGAN family and its many community variants, all free. Fast, deterministic, faithful to the input's structure. They add texture without changing content. Specialised variants exist for faces, anime, text and photographs, and choosing the right one matters more than choosing the newest.

Diffusion upscalers. A generation pass conditioned on the small image, usually at low denoise. Much stronger detail invention, much more drift. At denoise above about 0.4 they cheerfully add objects that were not there.

Tiled upscaling. Cut the image into overlapping tiles, upscale each, blend. This is how a 6-gigabyte-of-memory card produces a 6000-pixel image. The characteristic failure is tile seams and repeated content — a face appearing in each tile, because each tile is independently a plausible place for a face. Overlap and a low denoise value control this.

Face restoration — CodeFormer, GFPGAN and similar, all free. Purpose-built for faces and very effective. They have a documented tendency to regularise a face toward the average, so an unusual face comes back more conventionally attractive and slightly less itself. There is a fidelity slider; the default is usually too aggressive.

A workable ladder

For most work:

1. Generate at native resolution.
2. Masked detail pass on faces and any small critical element.
3. Convolutional upscale to twice the size.
4. Optional low-denoise diffusion pass at 0.2 to 0.3 to unify texture.
5. Sharpen and grade in a free editor, last.

Doing the detail pass before the upscale is the ordering people get wrong. Upscaling a wrong face gives you a large wrong face, and the upscaler has no idea it was wrong.

The limitation nobody removes

Every upscaler has a training distribution, and it invents what that distribution contained. An upscaler trained on photographs will render text as text-like texture. One trained on anime will smooth a photograph's grain into flat colour. A face upscaler applied to a painting will make it look like a photograph of a face.

So the practical discipline is to match the upscaler to the material, and to look at the result at 100% rather than trusting that bigger is better. A great deal of the "AI slop" look in circulation is not the generator at all. It is an aggressive upscaler applied by default, adding invented micro-detail everywhere, and turning it off is often the single biggest improvement available.

One habit worth borrowing from print. Work out the size you actually need before upscaling anything. A 300 dots-per-inch A4 page is 2480 by 3508 pixels; a full-page magazine image is not much more; most screen use never exceeds 2000 pixels on the long edge. People routinely upscale to 8000 pixels for a piece that will be seen at 1200, which quadruples the invented detail, the file size and the processing time for no visible gain. Decide the output dimension first and stop there.