JPEG quantization tables

JPEG splits the image into 8 × 8 blocks and transforms each one into 64 coefficients describing how much of each wave pattern the block contains — from flat, through gentle gradients, to fine alternating detail. That transform is fully reversible and loses nothing.

Then quantization: each of the 64 coefficients is divided by the value in the corresponding position of the table, and the result is rounded to a whole number.

The rounding is irreversible, and it is the entire lossy component of JPEG. Everything after it — the zig-zag reordering and the entropy coding — is lossless packing.

The table is not a single number because not all detail is equally noticeable.

The top-left position holds the DC coefficient — the block's average brightness — which matters enormously, so it gets a small divisor and survives almost intact. Values increase as you move down and right, toward the finest detail, which the eye barely registers.

At the bottom-right, divisors are large enough that almost everything rounds to zero. Zeros cost virtually nothing to store once run-length and entropy coded, which is where the compression comes from: the encoder is not shrinking the data so much as deliberately creating long runs of nothing.

The chroma table is more aggressive throughout, for the same reason chroma is subsampled — colour detail is less visible than brightness detail.

It scales the table. A high quality multiplies every divisor down toward 1, so little rounds to zero and little is lost. A low quality multiplies them up, so most of the fine coefficients vanish.

This is why the number is not a percentage of anything. It is an index into how aggressively the division is performed, and the relationship to file size is steep and non-linear: the first ten points below 100 cost almost nothing visible and save enormously, while the points below 70 cost visible quality for progressively less.

The JPEG standard includes example tables in its Annex K, but does not require them. Encoders are free to use their own, and the good ones do.

MozJPEG, for instance, uses tables tuned against perceptual metrics rather than the 1992 originals, which is a large part of why it produces smaller files than most encoders at matched visual quality. Camera manufacturers use their own. Browser canvas implementations use theirs.

So quality 80 in Photoshop, quality 80 in a browser and quality 80 in MozJPEG are three different sets of divisors producing three different files. The ranges are broadly comparable and the numbers are not interchangeable — which is worth remembering whenever someone quotes an exact setting.

The quantization tables are stored in the file itself, in a DQT marker, because the decoder needs them to reverse the division. That means they can be read back out.

Tools such as JPEGsnoop or ExifTool will show them, and analysts use the fingerprint to identify which software produced a file — cameras, editors and social platforms each leave recognisable tables. It is also how you can tell that an image has been re-saved by something other than the device that took it.