Image Compression Benchmarks — Real Numbers

Compression advice is usually folklore: a blog post says WebP is “30% smaller”, someone else says 25%, and nobody says on which files, at which quality, or how it was measured. We got tired of arguing from vibes, so we ran the numbers ourselves and published them. This page is the data annex of this site: two local test campaigns — one from August 30, 2026 (27 real-world photographs) and one from August 31, 2026 (12 synthetic 1920×1080 frames) — all encoded on our own machine with open-source encoders.

The short version, for the impatient: at matched visual quality (~39.6 dB PSNR) AVIF q70 was -22.6% smaller than JPEG q80, and WebP q85 was -7.5% smaller. The JPEG sweet spot is real: q75–85 delivers most of the achievable quality for a fraction of q95’s weight. PNG photographs shrink through palettes and resizing, not through optimizer flags. And every byte of that size advantage has a CPU price: AVIF encoded the same batch 17× slower than JPEG.

JPEG vs WebP vs AVIF at matched quality

Comparing “quality 80” across formats is meaningless — q80 means something different inside every encoder. The fair comparison is at equal visual quality, so we matched by PSNR: JPEG q80 is the anchor at 39.6 dB median PSNR across the 27 photos; the nearest settings for the other encoders are WebP q85 (39.2 dB) and AVIF q70 (40.4 dB). At that footing, the median file sizes diverge sharply.

Deltas are computed on exact byte medians (JPEG 321,833 B · WebP 297,726 B · AVIF 248,979 B), before the KB rounding shown above.

AVIF describes the same pixels with roughly a fifth fewer bytes at this quality level — measured file-by-file on real photographs, not extrapolated from marketing. WebP’s smaller advantage at q85 is the cost of its simpler block model. If maximum compatibility matters, WebP still shrinks files decisively; if bytes matter most, AVIF is the measured winner.

PNG strategies — what actually shrinks a PNG

PNG is lossless, so an “optimization” can’t throw detail away — it can only change how pixels are represented: encode more efficiently, reduce the palette, or reduce the image itself. We saved the same 27 photographs as PNG seven ways to separate those effects. Note that the resized rows change the image dimensions, so they are a different trade-off rather than a free lunch.

The lesson is blunt. optimize=True alone shaved -0.2% — the encoder was already doing the smart thing by default. Reducing the palette to 256 colors cut -68.2%; halving the resolution cut -74.7%. If the image must stay a lossless PNG, palettes are the honest lever; if it doesn’t have to stay PNG at all, WebP lossless (-34.7% smaller here) or a lossy format beats everything.

The separate timing campaign tells the same story on 12 synthetic frames: raw 24-bit PNG median 2,433 KB, optimize=True 2,228 KB (-8%), palette-256 1,048 KB (-57%), palette-64 665 KB (-73%).

How free online tools compare — documented limits, not our measurements

Popular “free” compressors come with constraints. To be completely explicit about method: the table below is research, not benchmarking. We collected publicly advertised limits from each vendor’s own pages and third-party write-ups; we did not measure these services, the limits can change at any time without notice, and none of the measured numbers on this page depend on them. Treat this as a snapshot of documented behavior.

Our benchmark tooling works the way Squoosh does — fully local — but with batch capability, and that is also how our own tools are built: files never leave your device during processing, and we impose no session caps or per-file MB ceilings; your browser’s available memory is the only practical limit.