Reproduction kit — PNGToolbox and FitMyPic measurements ========================================================= Everything published in our format, JPEG-quality and resampling studies comes out of bench_kit.py. Plain Python with Pillow, NumPy and SciPy: no network access, no third-party images, nothing to install beyond those three. python3 bench_kit.py all # corpus + all three studies + charts Individual stages: python3 bench_kit.py corpus # writes corpus/*.png (~30 seconds) python3 bench_kit.py formats # study A: PNG / JPG / WebP output sizes python3 bench_kit.py quality # study B: JPEG quality sweep, subsampling python3 bench_kit.py resize # study C: nearest/bilinear/bicubic/lanczos python3 bench_kit.py charts # the SVG charts used in the articles Environment used for the published numbers ------------------------------------------ Pillow 11.0.0 (libwebp, libjpeg-turbo, zlib) NumPy 2.1.3 SciPy 1.13.1 Date 22 October 2026 corpus.zip contains the exact six test images the published numbers were measured on. Unzip it next to this file and the script will use those images instead of regenerating them, which is what makes a byte-for-byte reproduction possible. The generator is included as well (bench_kit.py corpus --force) for anyone who wants to rebuild the corpus from scratch: it is deterministic, and an independently generated corpus produces the same rankings with small per-file differences. Timing columns in study-c-resize.csv are the one exception to reproducibility: milliseconds depend on the machine and its current load, so treat them as indicative. Sizes, PSNR and SSIM reproduce exactly. Published result files ---------------------- study-a-format.csv one row per test image: PNG / JPG / WebP sizes study-b-jpeg-quality.csv quality 40-95: size, PSNR, SSIM, bits per pixel study-b2-subsampling.csv 4:4:4 vs 4:2:2 vs 4:2:0 at fixed quality study-c-resize.csv resampling methods vs a box-filter reference Corrections, or a reproduction that disagrees with ours, are welcome through the contact page.