Lossy compression

What is actually discarded

Lossy image codecs do not guess. They apply a model of human vision. The image is converted to a colour space that separates brightness from colour, the colour channels are stored at reduced resolution because the eye resolves colour detail poorly, blocks of pixels are transformed into frequency components, and the high-frequency components — fine texture, sharp transitions — are rounded most coarsely.

That final rounding step, quantization, is where the information goes. The quality slider scales how coarse it is. Everything else in the pipeline is reversible.

Why the curve has a knee

Because the discarded data is disproportionately the data you are least able to see, the size-versus-quality relationship is not linear. Dropping from quality 100 to 90 typically removes more than half the file with no visible difference at all. Dropping from 90 to 80 removes a third again, still invisibly on a photograph. Below about 70 the losses start to show — first on flat gradients, then on high-contrast edges, then everywhere.

For photographs, 75-85 is the range where files are far smaller and the result is visually identical at normal viewing size. That is the single most useful number in image work.

Generation loss

Each lossy save quantizes what the previous save already quantized, and the artefacts of the first round become detail the second round must encode and then damage further. Ten generations at quality 80 produces visible blocking in sky and shadow, and edge halos, from an image that was indistinguishable from the original after one.

Generations accumulate without anyone deciding to accumulate them: a camera JPEG, cropped and re-saved, sent through a messaging app that re-encodes, downloaded and colour-corrected, uploaded to a platform that re-encodes again. Nobody made a bad decision at any step.

The defence is a discipline, not a setting. Keep a lossless master, do every edit on it, and export a lossy file once, at the end.

When not to use it

Anything with text, hard edges or flat colour — the codec's assumptions are wrong in every particular, and the artefacts land exactly where they are most visible. Also masters, archives, and any medical, scientific or forensic image where a fabricated detail is a serious problem rather than a cosmetic one.