Compress Image Files Without Wrecking Quality

Lossless compression is a repacking exercise. PNG, the most common lossless image format, runs each row of pixels through a prediction filter — each pixel is stored as its difference from a neighbor rather than as an absolute value — and then compresses the result with DEFLATE, the same algorithm behind ZIP files. Decompression reverses the process bit-for-bit, so the image that comes out is identical to the one that went in. This works brilliantly on screenshots, diagrams, and flat-color graphics, where long runs of identical pixels predict perfectly, and poorly on photographs, where sensor noise makes every pixel slightly unpredictable.

Lossy compression, as implemented in JPEG, works on completely different principles. The encoder first converts the image from RGB into a brightness channel and two color channels, then usually stores the color channels at half resolution — a trick called chroma subsampling that exploits the fact that human vision resolves brightness far more finely than color. Each channel is then split into 8 by 8 pixel blocks, and each block is transformed into frequency components: broad gradients become low frequencies, fine texture becomes high frequencies. Quantization then rounds the high-frequency components toward zero, aggressively at low quality settings and gently at high ones. Those rounded-away values are the loss, and they are unrecoverable.

The practical consequence: the quality slider on a JPEG or lossy WebP export is directly controlling how coarsely fine detail gets rounded away. WebP applies the same broad idea with better prediction between blocks, which is why it usually lands smaller than JPEG at matching visual quality.

Quality settings above 90 spend a lot of bytes on differences nobody can see; the size grows steeply while the visible image barely changes. The 75 to 85 range is the working sweet spot for photographs — detail loss exists but hides in texture. Below about 60, two artifact types start surfacing. Blocking shows up as visible 8 by 8 tiles in smooth areas like skies and skin, where the eye is exquisitely sensitive to banding. Ringing shows up as ghostly halos around hard edges — text over a photo is the classic victim, because sharp letterforms are made almost entirely of the high frequencies that quantization attacks first.

So judge compressed output where it fails first, not where it survives longest. Zoom to 100 percent and inspect smooth gradients, skin, and any text or line work. If those hold up, the busy regions are fine too. Keep the original open in another window and compare the two directly — memory of what the original looked like is unreliable after even a few seconds, so a real side-by-side beats squinting at the compressed copy alone.

One caution: judge at 100 percent zoom, not at 300. Every lossy image looks bad magnified three times; the question is whether it looks bad at the size people will actually see it.

The most common oversized image on the web is not an under-compressed JPEG — it is a photograph stored as a PNG. Screenshot tools and design app exports default to PNG, people photograph things and paste them into that pipeline, and the result is a photo carried by a format that must store its sensor noise exactly. Re-encoding a photographic PNG as a quality-80 JPEG routinely shrinks it several times over before you touch any slider, with no visible change.

The reverse mistake also costs you: a flat-color logo or a screenshot full of text saved as JPEG picks up ringing around every edge and often ends up larger than the clean PNG would have been. The decision rule is short enough to memorize.

Compression without a target is guesswork, and most of the time the target is imposed on you. These are the limits that generate most compression work in practice.

When a form imposes a hard cap well below what quality reduction alone can reach, the answer is dimensions, not more compression. A 6000 by 4000 pixel photo cannot gracefully become 100 KB, but resized to 1200 pixels wide first, it compresses there easily and looks better than the brutalized full-size version would.

Lossy compression is cumulative. Every re-encode quantizes the image again, and artifacts that were invisible after one save become blotchy after four — the same generation loss that made repeatedly photocopied pages unreadable. The single most important habit is therefore boring: never overwrite the original. Compress a copy, and when you need a different size later, go back to the master rather than compressing the compressed version.

For batches, pick your settings on the worst-case image — the one with smooth gradients or fine text — and apply them to the set, rather than tuning on an easy image and hoping. And when a result surprises you by being large, check the format and the pixel dimensions before blaming the quality setting; in practice one of those two is the culprit far more often than the slider.