How much do free PDF compression methods shrink everyday PDFs? A measured comparison on 16 public documents

I compressed 16 public PDFs (3 fillable forms, 2 research papers, 2 government reports, 2 photo documents and 7 library scans) with five free methods: lossless repacking with qpdf, bpdf's two levels (Recommended and Strong), and Ghostscript's /ebook and /screen presets. For each output I recorded the size, the time, whether it still opens, a visual difference score (PSNR of the first three pages, drawn by PDFium) and how many fillable form fields survived. No method was best on every kind of file. On photo documents bpdf Recommended saved 80 to 82% at PSNR 38 to 39 dB, against 63 to 66% at 30 to 31 dB for Ghostscript /ebook. On six of the seven forms, papers and reports Ghostscript saved 50 to 92%, where bpdf saved 1 to 25%, but on the forms Ghostscript removed every fillable field. On library scans already stored as JPEG 2000 all methods saved little. License: the paper and data are CC BY 4.0; study.mjs and fields.mjs are MIT (see their headers). Competing interests: the author builds bpdf.cc, one of the tools tested.