Quantization

Where it sits in the pipeline

A lossy encoder transforms blocks of pixels into frequency coefficients — how much coarse structure, how much fine detail. That transform is reversible and loses nothing. Quantization is the next step: each coefficient is divided by a value from a quantization table and rounded to an integer.

Rounding is where the loss happens, and it is the only lossy step in the process. Everything after it — entropy coding — is lossless again.

The quantization table

The table holds one divisor per coefficient position, and the values are not uniform. Low-frequency positions, which carry the broad shapes and colours, get small divisors and survive nearly intact. High-frequency positions, which carry fine texture, get large divisors and are frequently rounded all the way to zero.

That asymmetry is a model of human vision expressed as arithmetic. It is why a heavily compressed photograph still has recognisable shapes and colours while its fine texture has turned to mush.

The quality slider does not choose a table. It scales one: higher quality means smaller divisors, finer rounding, more coefficients surviving, and a larger file.

Why it explains the artefacts you see

Blocking — visible square edges in flat areas — happens when so many coefficients in a block round to zero that the block becomes a single flat value that does not match its neighbours.

Mosquito noise around text is the opposite failure: a hard edge needs many high-frequency coefficients to describe, coarse rounding produces the wrong ones, and the error appears as a shimmer along the edge.

Banding in a gradient comes from the same rounding applied to a signal that changes very gradually — the smooth ramp collapses into steps.

Knowing this tells you where to look when checking a compressed image: flat gradients, high-contrast edges, fine repeated texture, and deep shadows.