JPG quality vs file size: what each step from 95 to 30 really does, measured
7 minutes, updated 2026-11-19
A JPG quality slider runs from 0 to 100, which invites the guess that every step buys the same saving. Measured, the steps are uneven. Two of them do most of the work, and the top of the scale can make a file bigger than the one you started with. This guide shows where those steps fall, so you can pick a setting with the numbers in front of you.
How we measured: on 19 November 2026 we saved three public-domain NASA photographs at qualities 95, 90, 85, 80, 70, 60, 50, 40 and 30 with Chromium's built-in JPG encoder, the same one every tool on this site uses in your browser. The photos are 1920 pixels wide: a crew group portrait (1920 × 1536, lots of fine detail), an astronaut with two dogs (1920 × 1536, soft painted backdrop) and a tall single portrait (1920 × 2400). A synthetic 1440 × 900 screenshot, lines of dark text on white, went through the same steps. Sizes are in KB of 1024 bytes, the way the tools here count them.
What PSNR means
Next to each size is a PSNR figure (peak signal-to-noise ratio). It compares the saved file with the picture it was made from, pixel by pixel, and expresses the average difference in decibels (dB). Higher means closer. The scale is logarithmic: every 6 dB lower roughly doubles the typical pixel error, so a drop of 5 dB is a large change and a drop of half a decibel is noise.
Our reference is each photo as it arrived, decoded. The figures therefore measure how much a re-save damages the picture you already have, which is the question you face when a file is too big. As a rough guide:
| PSNR (dB) | What you see |
|---|---|
| Above 40 | Hard to tell apart from the source |
| 35 to 40 | Good; differences show only side by side |
| 30 to 35 | Visible on close inspection: softer texture, faint blocks |
| Below 30 | Obvious blocks and smeared detail |
PSNR weighs every pixel equally, while eyes forgive noise in hair or foliage and notice it on a smooth cheek or a clear sky. Use it to compare settings on the same picture, not one picture against another.
The measurements
The three source files were already JPGs. The patterns below held for all three photos; where they differ, the range is given.
| Quality | Crew (KB) | Crew (dB) | Dogs (KB) | Dogs (dB) | Tall portrait (KB) | Tall portrait (dB) |
|---|---|---|---|---|---|---|
| Source file | 732 | – | 561 | – | 1,012 | – |
| 95 | 929 | 43.7 | 719 | 47.1 | 1,307 | 47.9 |
| 90 | 780 | 42.7 | 601 | 44.9 | 1,088 | 45.3 |
| 85 | 707 | 41.5 | 543 | 44.6 | 986 | 45.8 |
| 80 | 638 | 36.4 | 490 | 38.8 | 869 | 39.2 |
| 70 | 394 | 33.1 | 302 | 34.9 | 483 | 34.7 |
| 60 | 344 | 32.4 | 260 | 34.2 | 420 | 34.1 |
| 50 | 306 | 31.4 | 233 | 33.1 | 367 | 33.4 |
| 40 | 246 | 30.8 | 181 | 32.5 | 258 | 32.8 |
| 30 | 203 | 29.8 | 146 | 31.6 | 199 | 32.0 |
Quality 90 and 95 made the files bigger
Re-saving at 95 produced files 27 to 29% larger than the sources, and 90 still came out 7 to 8% larger. Quality 85 got them only about 3% under. The extra bytes record the source's existing compression artefacts faithfully; no detail comes back.
If a JPG is already the size you need, leave it alone. The compress tool checks this for you: when the result in the same format would be larger than the file you gave it, it hands back the original and says so in the row.
From 85 to 80: a small saving and the biggest loss
Moving from 85 to 80 saved only 10 to 12%, yet PSNR fell by 5.1 to 6.6 dB, from 41.5 to 45.8 dB down to 36.4 to 39.2 dB. No other single step on the scale cost as much. The result at 80 is still in the good band, but about half of all the fidelity lost between 95 and 70 goes at this one step.
From 80 to 70: the biggest saving
The next step is the mirror image. Going from 80 to 70 cut the files by 38 to 44%, the largest saving anywhere on the scale, for a further 3.3 to 4.5 dB. At 70 the three photos measured 33.1 to 34.9 dB: softening you can find if you zoom into texture such as fabric, hair or the painted backdrop, and little you would notice at normal viewing size.
Below 70, each step buys less
From 70 down to 30 the files shrank by another 48 to 59% in total, but it took four steps, each saving 10 to 30% for 0.6 to 1.1 dB. At 30 the crew photo, the busiest of the three, fell below 30 dB (29.8), where blocks around edges become obvious. The two single-subject portraits ended at 31.6 and 32.0 dB.
One oddity: the tall portrait measured slightly closer to its source at 85 (45.8 dB) than at 90 (45.3 dB). Differences under 1 dB between neighbouring settings are within the noise of this method, so do not read meaning into them.
We did not trace why Chromium's encoder jumps at 80 and 70; the pattern simply repeated on all three photos. Other programs map the quality number differently, so 80 in a desktop editor or a phone's camera app will not give the same file as 80 here. The figures on this page describe what a browser produces, which includes every tool on this site.
The screenshot behaves differently
Text on a flat background lost fidelity smoothly, 1.4 to 2.8 dB per step, with no cliff and no single big saving. From 95 to 30 the screenshot only fell to a third of its size, while the photos fell to 15 to 22% of theirs. Sharp letter edges are what JPG handles worst, so the screenshot dropped below 35 dB at 70 and below 30 dB at 40.
For text, stay at 85 or above, or use another format. WebP vs JPG, measured compares the formats on this same screenshot, and making screenshots smaller covers cropping and palette PNG.
| Quality | Screenshot (KB) | Screenshot (dB) |
|---|---|---|
| 95 | 304 | 44.5 |
| 90 | 250 | 41.7 |
| 85 | 219 | 39.4 |
| 80 | 198 | 37.5 |
| 70 | 170 | 34.7 |
| 60 | 149 | 32.6 |
| 50 | 135 | 31.2 |
| 40 | 120 | 29.8 |
| 30 | 103 | 28.2 |
Which setting to use
For forms, a quality number is the wrong thing to choose by hand. The same setting, 70, gave 394 KB on the crew photo and 302 KB on the dog portrait, both 1920 × 1536. Enter the limit in the fit-to-size tool instead. It encodes, measures and adjusts, and when a quality around 70 is not enough it reduces the pixel size before pushing quality lower. Why that order wins is measured in resize or lower the quality. Enter slightly less than the stated limit; what KB means on a form explains why.
For email and messaging, 80 is a safe default. It sits just before the large saving at 70, so if a batch is still too big, 70 is the one step worth trying before you reduce the pixel size. For a website, 70 to 80 covers most photos, and WebP gives the same fidelity in fewer bytes where it is accepted. The wider picture of how compression works is in the image compression guide.
| Use | Setting | What we measured |
|---|---|---|
| Archive, print or further editing | Keep the original | Re-saving at 90 to 95 made files 7 to 29% bigger |
| Screen viewing, sharing, email | 80 | 13 to 14% under the source, 36 to 39 dB |
| Website photo on a size budget | 70 | 46 to 52% under the source, 33 to 35 dB |
| Form with a KB limit | Fit-to-size tool | Searches the setting for you |
| Screenshot or scan with text | 85 to 90, or PNG or WebP | 39 to 42 dB on our screenshot |
Two habits that cost quality for nothing
- Raising the quality of a file that was saved low. A JPG saved at 60 and re-saved at 95 becomes a larger file of the same damaged picture, as the 95 row above shows on files that were never damaged to begin with.
- Saving the same JPG again and again while you edit. Each save adds its own loss on top of the last. Work from the original and export to JPG once, at the end.
Check one photo before a batch
Your photos will not match ours exactly; at the same setting and the same pixel size, the busy crew photo came out about 30% larger than the dog portrait. Before compressing a folder, drop one typical photo into the compress tool and try 80, then 70. The row shows the new size and the saving as a percentage, and Compare puts the result beside the original with a divider you can drag across a face or a patch of fine texture. If you cannot see the difference at the size the photo will be viewed, take the smaller file.