Smol image compression benchmark, 5 October 2026
https://smolmac.com/benchmark

results.json / results.csv
  One row per compressed file that was scored: image, tool, format, setting, bytes, ssim, ssimulacra2.
  1,335 rows. "image" is the file name in the Kodak Lossless True Color Image Suite
  (https://r0k.us/graphics/kodak/), which is not redistributed here.

ssim_lib.py     the SSIM implementation used (11-tap Gaussian, sigma 1.5, mean of R, G and B)
harvest2.py     scores a folder of outputs against the originals (SSIM and SSIMULACRA2 from libjxl 0.11.2)
match2.py       turns the rows into "bytes needed to reach the same score" (interpolated in log-bytes)
smol-run.mjs    drives Smol through its MCP server
squoosh.mjs     drives squoosh.app in headless Chrome

Machine: Apple M2 Pro, 16 GB, macOS 27.0. Smol 1.0.35 (standard compression), sips-316,
caesiumclt 1.5.0, Zipic 1.10.2 (free tier, three images), squoosh.app in Chrome 154.
The scripts assume a working folder of /tmp/smol-bench.