Compress JPG

JPEG does not store pixels. It splits the image into 8×8 blocks, converts each one into frequencies with a discrete cosine transform, and then divides those frequencies by a quantisation table before rounding. High-frequency detail — the fine texture the eye is least sensitive to — rounds to zero first and is gone permanently. The quality slider is really a scale factor on that table: lower quality means coarser division, more zeros, and a smaller file.

This is why compression results vary so much between images. A photograph of foliage or gravel is dense with high-frequency detail and resists compression; a studio portrait against a plain backdrop is mostly smooth gradients and collapses to a fraction of its original size at the same quality setting. Two 4 MB photos can land at 900 KB and 200 KB from identical settings, and neither result is a bug.

JPEG also throws away colour before it touches detail. Chroma subsampling stores brightness at full resolution but colour at half, in both directions, because human vision resolves luminance far more sharply than hue. That single step removes half the colour data before the quality setting is applied at all, and it is invisible on photographs — which is exactly why it is the default.

Quality mode is the right default when the file just needs to be smaller and no specific number has to be met. Quality 80 is the usual sweet spot: on most photographs the difference from the original is not visible at normal viewing distance, while the file is 60–80% smaller. Below about 60 the 8×8 block structure starts to show as visible squares in smooth areas like skies, and around edges as a faint halo known as ringing.

Target size mode exists for the cases where a number is imposed on you — a form that rejects anything over 200 KB, a marketplace listing cap, an email limit. Rather than making you guess, the tool encodes the image repeatedly and bisects the quality setting until it finds the highest quality that still fits under your ceiling. You get the best-looking image that satisfies the constraint, instead of an arbitrary guess that undershoots and looks worse than it needed to.

File size scales with pixel count, so halving both dimensions removes three quarters of the pixels — a far larger saving than any quality setting can achieve without visible damage. A 6000×4000 photo from a modern camera is 24 megapixels; displayed full-width on a 1920px monitor, more than 90% of those pixels are discarded by the browser at render time anyway.

The practical consequence is that if a photo will not reach your target size at acceptable quality, the answer is almost never a lower quality number. Cap the longest edge instead. Going from 6000px to 1920px typically does more than dropping quality from 80 to 40, and unlike aggressive quantisation it introduces no blocking or ringing artefacts — the image is simply smaller, and sharp at the size it is actually viewed.

JPEG compression is lossy and it is not idempotent. Every save re-runs the transform on data that has already been quantised, so artefacts compound. Compressing an already-compressed JPG a handful of times produces visible degradation even at high quality settings. Always compress from the most original file you have rather than from a copy that has already been through the process, and avoid using a compressed export as the source for the next edit.

JPEG is also the wrong container for some images entirely. It has no alpha channel, so transparency is flattened — usually to black or white. Its frequency-based model is built for photographic gradients and handles hard edges badly, which makes screenshots, text, line art, logos and diagrams look smeared and haloed even at quality 90. Those belong in PNG, where a screenshot is frequently both smaller and lossless.