How Image Compression Quality Works

The slider controls quantization, not 'quality'

Inside a JPEG encoder, the image is split into small blocks and converted into frequency data. The quality setting scales the quantization tables — the divisors that decide which frequency detail survives. Low quality uses aggressive divisors, zeroing more detail and shrinking files; high quality preserves more coefficients at the cost of bytes. Everything downstream — blocking, ringing, banding — is a consequence of what those divisors removed.

Understanding this explains the non-linear payoff curve. The first reductions from quality 100 cost very little visible detail because the highest frequencies are nearly imperceptible anyway. Past a tool-dependent midpoint, each further step removes progressively more visible information. The curve also differs by content: a gravel texture tolerates quality 60 cheerfully while a gradient sky falls apart at 75. The slider is one control; the image content decides what it means.

Why 'quality 80' differs between tools

Quality numbers are not an industry unit. One encoder maps the 0 to 100 dial directly onto quantization tables; another maps it onto a target perceptual metric; a third clamps chroma channels differently. Setting 80 in a photo editor, a command-line encoder, and a web tool can produce meaningfully different files — sometimes several points of visible difference at identical numbers.

The practical consequence: never carry quality numbers between tools as if they were measurements. Calibrate per tool by compressing one representative image at several settings and comparing outputs at display size. Within a single tool the dial is consistent enough to standardize on; across tools it is folklore. This is also why 'use quality X' advice online is only loosely transferable — the honest recommendation is always 'lower until artifacts appear at display size, then stop'.

The economics of the top of the range

Between quality 90 and 100, file size can double or triple while visible improvement becomes undetectable outside controlled comparisons. Those top settings exist for archival masters and intermediate editing steps, not for delivery. The bytes they buy encode detail that the human visual system discards anyway at normal viewing distance — chroma nuances in already-noisy regions, frequency components below perception thresholds.

Delivery workflows benefit from inverting the instinct: start aggressive and relax only where artifacts appear. Compress at a middle setting, run the artifact checklist on the difficult regions, and raise quality only for images that fail. This per-image negotiation routinely lands collections 30 to 50 percent smaller than blanket high-quality settings with no visible regression — the same total quality, distributed where it matters instead of spread evenly where it is wasted.

Content decides sensitivity: smooth, busy, and synthetic

Three content classes respond to compression differently. Smooth areas — skies, studio backgrounds, soft skin — expose artifacts first because there is no texture to hide quantization noise; they demand higher quality or dithering. Busy areas — foliage, gravel, crowds — mask artifacts almost completely and tolerate aggressive settings. Synthetic content — text, charts, UI — sits in its own category: any lossy artifact near sharp edges reads as a defect, which is why such content belongs in lossless formats entirely.

Mixed images are the negotiation cases: a portrait with a smooth studio background wants higher quality than the face alone would need, because the background betrays the encoding. When an image mixes classes, tune for the most sensitive region. Cropping is a legitimate escape — removing the problematic smooth margin often lets the rest compress far more aggressively without any visible cost.

How to judge quality honestly (most people do it wrong)

The standard mistakes: judging at 100 percent zoom, where every image looks damaged and every difference looks dramatic; judging on a calibrated studio monitor rather than the phone the audience uses; and comparing the compressed image against the original side by side instead of judging it alone. A/B at pixel zoom is a test of the encoder, not of the image's fitness for purpose.

The honest protocol: view at actual display size, at normal viewing distance, in the context it will appear — inside the page layout, the email, the slide. Ask 'does anything draw the eye as a defect?' rather than 'can I find a difference from the original?' Most images that fail the side-by-side test pass the real-world test completely. Reserve pixel-zoom inspection for the artifact checklist on specific suspicious regions, not as the primary judgment. Fitness for display, not fidelity to source, is the standard that matches what compression is actually for.

Quality settings across JPG, WebP, and AVIF

Modern formats move the whole quality curve downward: at equal visible quality, WebP typically needs 25 to 35 percent fewer bytes than JPG, and AVIF can need 50 percent fewer still, because their transforms and prediction are more efficient. The slider behaves the same way conceptually — it still governs how much detail survives — but the numbers are again tool-specific and not comparable across formats.

Two wrinkles matter in practice. AVIF encodes slowly, so aggressive quality presets that finish fast are often worth the slightly larger file for bulk jobs. And all lossy formats share the smooth-region weakness: banding in gradients appears at higher quality numbers in WebP and AVIF than blocking does in JPG, so gradient-heavy images still deserve a conservative setting regardless of format. The decision order stays constant: resize first, pick format by content, then tune quality by artifact inspection at display size.

What acceptable quality actually means

Acceptable quality is a perception threshold, not a technical one, and understanding that reframes the whole task. Viewers judge images at display size, at a glance, and in context — a photo inside an article competes with the reading experience, not with a calibrated monitor at hundred-percent zoom. The artifacts that ruin perceived quality are specific and few: blocking in smooth skies, ringing around sharp edges, color banding in gradients, and mushy text. Everything else — subtle texture attenuation, minor chroma smoothing — passes below notice for nearly every real viewing situation.

This explains the asymmetry that makes compression worthwhile: large portions of a file's bits serve detail the audience never perceives, while a small set of visible artifacts carries almost all the perceived quality. The practical consequence is that quality settings can move far below their nominal values before visible damage appears, provided the image content cooperates. Busy photographs hide artifacts in their own texture; clean graphics expose them instantly. Reading an image's content type before choosing settings predicts where the threshold lies far better than any benchmark number.

The evaluation routine that professionals converge on is simple. Compress at the intended setting, view at intended display size, and hunt specifically for the four artifact families — blocking, ringing, banding, mushy text. If none appears, the setting is validated for that content class; record it and reuse it. If one appears, move the slider until it disappears and note where. Over a few batches this produces a personal calibration table — photographs at this setting, screenshots at that one — and quality judgment stops being anxiety and becomes a lookup. The goal was never maximum fidelity; it was fidelity adequate to the viewing contract, reliably achieved.

Frequently asked questions