Benchmarks

Benchmarks

vs precomp

precomp is the reference byte-exact PNG recompressor (also preflate-based). It treats a PNG as a deflate stream and re-compresses it with LZMA2; packPNG un-filters to pixels and uses an image codec. Both reproduce the original file byte-for-byte.

17 real-world PNGs (1.52 MB), 56-core Xeon E5-2690 v4. Every row byte-exact (17/17). Lower ratio = smaller = better.

backend	approach	ratio	encode	decode
TCIP (default)	preflate + WebP-lossless	45.7 %	2.38 s	0.72 s
TMCP (-preflate-max)	preflate + kanzi-TPAQX	47.6 %	11.8 s	11.5 s
TPCL (-tpcl)	preflate + MT-LZMA2	64.3 %	1.68 s	0.58 s
TVCP (-fast)	kanzi + zstd	85.6 %	0.39 s	0.04 s
precomp 0.4.8	preflate + LZMA2	75.3 %	2.21 s	0.98 s

• TCIP (the default) is 39 % smaller than precomp — modelling the image with WebP-lossless beats LZMA2-on-deflated-bytes.
• TPCL uses precomp's exact recipe (preflate + LZMA2) yet wins on all three axes (smaller and faster), thanks to a newer preflate + multi-threaded LZMA2.
• precomp's -intense / -brute don't change PNG results (~75 %); they only help raw/embedded zlib in other container types.

Speed methodology: packPNG times are -th0 (parallel across files); precomp is per-file (no batch mode). For a single file precomp's speed is competitive (it multi-threads LZMA2 internally); packPNG's edge is in batch. Ratio is mode-independent.

Robustness

• 162/162 PngSuite valid PNGs round-trip byte-exact.
• 14/14 intentionally-corrupt PngSuite files handled without crashes.
• Encode is deterministic (same input → identical SHA-256), incl. Linux ↔ Windows.