Image compressor benchmark: Smol against sips, Squoosh and Caesium

A compressor can always make a smaller file by keeping less of the picture, so file size alone says nothing. Each encoder was run at several settings on every image, each output was scored against its original, and we then worked out how many bytes each encoder needs to reach the same score. Totals are bytes for all 24 images; the percentage is the gap to Smol’s AVIF.

SSIMULACRA2 is a perceptual quality metric from Cloudinary that ships with the JPEG XL reference software; its authors describe a score of 70 as high quality and 90 as very high quality. “n/a (2 of 24)” means none of the settings we ran reached that score on 2 of the 24 images, so no fair total exists. Caesium’s WebP files are byte-identical to Smol’s at the same quality number (72 of 72 files). Its totals differ slightly only because we ran it at more settings, which gives the interpolation more points and a wider range.

The older SSIM metric tells a flatter story. By SSIM the two AVIF encoders are 2% apart, and Apple’s JPEG encoder beats MozJPEG:

Squoosh is the free reference most people compare against, and for two of three formats there is nothing to separate them. At the same quality number, Smol’s JPEG files were within 6 to 51 bytes of Squoosh’s MozJPEG on all 72 files compared, with identical SSIM, and its WebP files were byte-identical on 24 of 24. Smol is not a better JPEG or WebP encoder than Squoosh. It is the same output for a whole folder at once.

AVIF differs. We ran Squoosh’s AVIF at its default setting only, then found the bytes Smol’s AVIF needs to match each file’s score:

The two metrics disagree, and we report both. An image is “comparable” when Squoosh’s score falls inside the range of settings we ran for Smol.

Wall-clock time to convert all 24 PNGs to one format, on an M2 Pro.

Smol is the slowest of the three by a wide margin. MozJPEG is built to trade encoding time for smaller files, so some of the gap is the encoder; we did not measure how much is the encoder and how much is Smol itself. Squoosh is not in this table because it has no batch mode.

The 24 source PNGs are 15,394,305 bytes. Smol has no lossless mode, so it is absent here.

• Speed. A sips loop was about 13 times faster and caesiumclt about 37 times faster.
• JPEG at high quality. At a SSIMULACRA2 score of 85, Apple’s JPEG encoder needed 5.8% fewer bytes than Smol’s. By SSIM, Apple’s JPEG was smaller at both scores we matched.
• AVIF, by SSIM. Squoosh’s default AVIF was 4.2% smaller than Smol’s at the same SSIM, on every comparable image.
• Lossless. Smol always re-encodes. For pixel-identical output use ImageOptim, OxiPNG or Caesium’s lossless mode.
• JPEG and WebP against Squoosh. A tie, not a win: the files are the same.

• Images. The 24 images of the Kodak Lossless True Color Image Suite, 768 × 512 pixels each, 15,394,305 bytes of PNG in total. They are public, so anyone can repeat this.
• Machine. Apple M2 Pro, 16 GB, macOS 27.0.
• Tools. Smol 1.0.35 driven through its MCP server; sips-316 (built into macOS); caesiumclt 1.5.0; squoosh.app in headless Chrome 154; Zipic 1.10.2 on its free tier, which allowed three images and no AVIF.
• Settings. Each encoder at four to ten quality settings per format, except Squoosh’s AVIF (default only). 1,335 output files in all.
• Scoring. Every output against its original PNG with SSIMULACRA2 (libjxl 0.11.2) and SSIM (11-tap Gaussian window, sigma 1.5, mean of the red, green and blue channels).
• Matching. For each image and encoder, the bytes needed to hit a target score are interpolated in log-bytes between the two nearest settings measured, then summed over the 24 images. If any image cannot reach the score, no total is reported.
• Timing. Wall-clock seconds for the whole folder, two or three runs each, all listed above.

• Twenty-four small film scans are not your photo library. Phone photos, screenshots and product shots on white will give different numbers.
• A quality metric is a model of what people see, not people. SSIMULACRA2 and SSIM disagree in places, which is why both are shown.
• One machine, one version of each tool, one day. Smol’s JPEG output changed between versions 1.0.34 and 1.0.35, and other tools change too.
• Squoosh’s AVIF was measured at one setting and Zipic on three images, so those comparisons are narrower than the rest.
• We make one of the products tested. The data and scripts below exist so the result does not depend on trusting us.

• results.json and results.csv: one row per scored file (image, tool, format, setting, bytes, SSIM, SSIMULACRA2), 1,335 rows.
• ssim_lib.py, harvest2.py and match2.py: the scoring and matching code.
• smol-run.mjs and squoosh.mjs: how Smol and Squoosh were driven. README.txt explains the rest.

For what the numbers mean when you are choosing a tool, including four compressors we assessed without measuring, read the best image compressors for Mac. Found a mistake? Tell us and we will correct it.