By Dik Raci · Creator of online tools · About the author

JPEG Quality 75 vs 85 vs 95: The 5x Size Question, Answered with Data

Somewhere along the way, quality 80 became a religion and quality 95 became a sin - or the reverse, depending on which forum you last read. Almost none of the people arguing have measured what the difference actually costs and buys. We did. As a second, independent replication of our earlier 27-photograph lab, we encoded 12 fresh test frames at every quality from 50 to 95 with progressive optimization enabled, and we kept three numbers for each setting: the exact file size, the PSNR fidelity against the original, and the encoding time. This article answers the only question that matters: when someone tells you to raise or lower JPEG quality, what exactly are they asking you to pay?

The size curve is not linear - it is an explosion

The median file across our 12 frames at 1920 pixels wide weighed 37 KB at quality 50 and 469 KB at quality 95. Between those extremes sits the finding most people have never internalized: the cost of quality accelerates. From 50 to 60 costs you 14 KB. From 85 to 95 costs you 299 KB - a 2.8x jump for the last ten points. The curve is exponential because JPEG's quantization tables compress coarse detail aggressively at low settings and then spend enormous amounts of bits preserving ever-finer high-frequency detail at high settings. Every step up the top of the ladder buys less perceived improvement than the last one did.

What PSNR says about the 75 / 85 / 95 question

Quality 75 measured 40.58 dB, quality 85 reached 41.41 dB, and quality 95 topped out at 43.50 dB on our three-frame fidelity subsample. Read those distances honestly: going from 75 to 85 - paying 75% more bytes - buys 0.8 dB. Going from 85 to 95 - paying 176% more bytes - buys 2.1 dB. Perceptual studies generally place the visibility threshold of common artifacts in the low-40s dB range for ordinary photographs, which means q75 already sits at the edge of what a paying viewer can detect on a photo, and q95 spends its enormous budget mostly on fidelity nobody perceives at normal viewing sizes. The exceptions to that statement are worth listing, because they are where q90-95 genuinely earns its cost.

When the top of the ladder is actually worth it

• Text and fine geometry: screenshots with small type develop visible ringing at q70-75 that only disappears in the 90s.
• Re-compression chains: files that will be cropped, resized or re-saved downstream lose quality at every step; starting at 90+ preserves headroom.
• Hard edges on flat backgrounds: product cutouts and logos on solid colors show mosquito noise early; high settings keep edges clean.
• Print or zoom workflows: any destination where a user can inspect pixels at 200% justifies the bytes.

The quality number is not portable - a warning that saves migrations

Our companion format benchmark measured the same frames with three encoders at identical numeric settings, and the results were wildly different: at quality 80 the JPEG weighed a median of 126 KB, the WebP 27 KB, and the AVIF 113 KB - three files with three different fidelity levels from one magic number. The same trap exists between JPEG encoders: libjpeg, mozjpeg and hardware encoders all interpret the scale slightly differently. The rule we now follow: quality settings are local dialects, not standards. When you migrate or automate, pick the output size or measured fidelity you want and tune the number to hit it - do not carry a number across tools and assume it means the same thing.

Speed: the quiet argument for staying moderate

We timed the batch encodes as well: all 12 frames compressed at quality 80 in 0.24 seconds total - about 20 milliseconds per image - which made JPEG roughly 9 times faster than WebP method 6 and 17 times faster than AVIF speed 6 on the same machine. For a nightly build this is trivia. For an image API compressing user uploads in real time, or a browser-based tool where the visitor's phone does the work - like the ones we run on this site - a codec that returns in 20 milliseconds keeps the experience instant. Moderate JPEG quality also produces smaller decode payloads, which speeds up rendering on the visitor's side too.

Our recommendation, in one paragraph

For web delivery of photographs, quality 75-85 remains the defensible range in 2026: our measured medians of 95-170 KB at 40-41 dB PSNR deliver images almost nobody can distinguish from originals at normal sizes, at a fifth to a third of the weight of quality 95. Reserve 90-95 for text-heavy screenshots, re-compression sources, and edge-critical graphics - the four cases above. And whichever number you pick, verify it on a sample of your real images rather than on folklore: run ten of your files through two adjacent settings, compare the byte medians and look at the crops at 100%. It takes five minutes and it ends the argument permanently.