CMNT: A 3-arY Combinatorial tree image codec outperforms PNG on continuous art

CMNT is a lossless image codec and Python toolset that aims to reduce file sizes by combining reversible color decorrelation, adaptive spatial prediction, and a ternary prefix tree. It transforms RGB into a luminance-like channel and two color-difference channels, predicts pixels from nearby values, then encodes the prediction residuals as ternary symbols. Groups of five trits are packed into one byte, and the result is stored in a native .cmnt file. The project says decoding reproduces every pixel exactly, with round-trip verification using matching SHA-256 hashes. It also offers a compact near-lossless mode using quantization, plus command-line conversion and a Windows desktop viewer.

Benchmarks cover five examples: 3D art, digital art, a phone screenshot, CGI, and a photograph. With the optional MANIAC context-modeling stage, CMNT files are 6.9%-29.7% smaller than standard PNG files across those examples. The base CMNT results are mixed: they beat PNG on two samples but produce larger files on three. Google WebP Lossless is smaller than even the MANIAC-enhanced CMNT files in every listed comparison, so the demonstrated advantage is specifically over PNG, not the strongest tested alternative. The repository also claims the near-lossless mode can reach JPEG-like sizes without 8×8 blocking artifacts, and includes installation instructions, automated tests, and an MIT license.