Upscaling: Enlarging a Photograph Without Turning It to Mush

Upscaling means enlarging an image while building in detail. Classic resizing merely stretches pixels and gives mush. AI upscaling constructs what is not in the file, drawing on what it learned in training.

Where the detail comes from

The model has seen millions of pairs of small and large images and learned what an enlarged version usually looks like. Given a small frame, it predicts a plausible version. The operative word is plausible, not authentic.

What that means in practice

Upscaling invents. Skin texture, eyelashes, fabric — all of it is drawn in from statistics. For a portfolio or a print this is usually acceptable: the result looks convincing. For anything where documentary accuracy matters it is not: an enlarged number plate or text on a document will be fiction.

The sensible limit

Doubling almost always works well. Quadrupling works if the source is sharp. Eight times is the territory of invention: the model is no longer restoring but composing. No poster will come out of a 300-pixel photograph, whatever an interface promises.

The order of operations

• Remove noise and compression artefacts first, then enlarge — otherwise they get enlarged too
• Upscale before colour grading: grading adapts to the final level of detail
• Retouch after enlarging, not before: upscaling will amplify some defects
• Final compression last, once

Where upscaling is indispensable

Archives. Photographs from ten or fifteen years ago were taken on cameras whose resolution looks modest today. Doubling brings them back into circulation — and that is perhaps the most honest use of the technology.

This material is intended for adults over 18 and is informational in nature.