Understanding image compression
Smaller files, almost the same picture.
Why a JPG quality of 85 is the sweet spot — and why "lossless" and "smaller" aren't always the same.
What "quality" actually controls.
A JPG quality slider doesn't pick a percentage of pixels to keep — it picks how aggressively to round the discrete cosine transform coefficients in 8×8 blocks. Higher quality keeps finer frequency components; lower quality throws them away. The transform was chosen because the eye barely notices subtle high-frequency loss in photographs but sees flat-area banding immediately. Around quality 85 the file is several times smaller than the source and visually indistinguishable; below 60, blocky artefacts start to show.
quality 85 ≈ visually identical, ¼ the bytes
Lossless compression of PNG.
PNG already runs every image through a deflate compressor — the same algorithm gzip uses. A "PNG compressor" doesn't throw pixels away; it picks a smarter row-prediction filter, re-orders the palette, or runs a slower deflate variant (zopfli) that finds more compact bit patterns. On large flat- colour images you can save 20–30% with no loss. On photographs converted to PNG, savings are smaller because the data was incompressible to start with.
Quantising — fewer colours, less file.
Some PNG compressors (pngquant, the one behind TinyPNG) drop the image's 24-bit colour to an 8-bit palette of the 256 most representative colours. The output is technically lossy — pixels at the edge of two original colours blend toward one of the palette entries — but visually it's hard to spot, and the file gets 50–70% smaller. UI screenshots, icons, flat illustrations all do well; gradients and photographs show banding.
Re-saving compounds the loss.
Open a JPG, edit it, save it again — and the lossy round- trip happens twice. Block boundaries shift, colours drift, edges blur. The fix is to keep originals lossless (PNG, TIFF or the camera's RAW) and only encode to JPG for the final output. Tools that "re-save without re-encoding" exist for JPG (lossless rotate, lossless crop) and are worth using when you can.
What metadata you ship.
Photos from phones and cameras carry EXIF metadata: GPS coordinates, camera model, timestamp, sometimes a thumbnail embedded inside the file. That data adds bytes (sometimes hundreds of kilobytes for a thumbnail) and leaks information about you. Most serious image compressors strip non-essential metadata by default, keeping only the orientation tag and colour profile. The right setting for "share this on the web" is "strip everything but rotation and ICC profile".
Picking a target size.
For the web, a hero image should be under 200 kB; a thumbnail under 30 kB; a content image somewhere in between. For email attachments, under 5 MB total per message is the polite limit. Pick a target file size first, then pick the quality/format combination that hits it without visible damage. That's almost always WebP or AVIF for the web, JPG 85 for sharing, PNG for screenshots and icons.
Read next