How JPEG Compression Works and How Much Quality to Use

JPEG doesn't compress an image the way a ZIP file compresses text. A ZIP file has to be perfectly reversible — every byte has to come back exactly as it went in. JPEG doesn't play by that rule. It's a *lossy* format, which means it deliberately throws away information the human eye is bad at noticing, and keeps the information your eye is good at noticing.

Two things get discarded first: **Color detail** and **Fine texture in busy areas**. Human vision is far more sensitive to changes in brightness than changes in color. JPEG encoders exploit this by storing color information at lower resolution than brightness information — a technique called chroma subsampling.

Every JPG compressor gives you some version of a quality slider or percentage. Here's what it's actually doing: it controls how aggressively the encoder simplifies each block before writing it to the file. Higher number, less simplification, bigger file. Lower number, more simplification, smaller file — and past a certain point, visible artifacts.

Quality numbers are encoder-specific and the same setting can behave differently across images. Export a candidate, then inspect faces, fine texture, gradients and high-contrast edges at the intended viewing size. Keep the least-compressed result that meets the destination's size requirement rather than relying on a universal percentage.

Rather than assert what compression achieves, here is the complete measurement set behind FiloTool's three compression levels. Four fixtures were encoded and measured in bundled Playwright Chromium on 2026-08-13, using the same Canvas MIME types and preset quality values the production tool uses, including the production behaviour of keeping the original file when a candidate is not smaller.

The two lossy formats respond to the levels roughly as you would expect:

Both PNG fixtures came back byte-for-byte identical at every level, and that is the system working correctly rather than failing. Browser Canvas PNG encoding exposes no lossy quality control, so all three levels perform the same lossless PNG re-encode. When that re-encoded candidate is not smaller than the file you started with—which is common for an already-efficient PNG—the tool discards its own output and hands back your original.

The practical consequence is worth internalising: for a screenshot or a flat graphic already stored as PNG, a compression slider has nothing to pull on. If such a file genuinely needs to be smaller, the lever is changing format or reducing dimensions, not turning compression up. Converting a screenshot to JPG will make it smaller and will also put ringing around every letter, which is usually the wrong trade—compare formats on a real file before committing.

Read the JPEG row again as a progression rather than three separate options. Moving from Quality to Smart saved a further 4,279 bytes; moving from Smart to Extreme saved 3,726 more. The absolute savings shrink while the visual cost keeps climbing, because the encoder is now discarding information you were more likely to notice.

That shape is the argument for choosing the least aggressive level that meets your actual constraint. If a 12 MB upload limit is the constraint and Quality gets you under it, Extreme buys you nothing you needed and costs detail you cannot get back.

If your PNG screenshot will not shrink, the instinct is to convert it to JPG. A separate FiloTool benchmark, encoded with sharp 0.34.5 on 2026-08-02, measured exactly that case and the result argues against it. Starting from a 59,145-byte interface screenshot:

The same pattern held for a flat illustration, where JPG saved 13.3% and was the only output the reviewer marked as showing visible artifacts—ringing and colour bleed along hard diagonal edges—while WebP saved 61.4%. So the honest sequence for a stubborn PNG is: try a different format before you try harder compression, and reach for Image Converter rather than a quality slider.

Every JPEG save is a fresh lossy encode of whatever was decoded, not an edit of the previous file. Compress an image, open the result, compress again, and the second pass discards information based on artifacts the first pass introduced. This is generation loss and it accumulates, which is why a photograph that has been through several messaging apps looks soft in a way no single save explains.

Keep the highest-quality original you have and re-derive from it each time you need a different size or setting. Never treat a compressed export as your working copy.