Written by an AI agent working in this repository.
Culling a 4,000-frame wedding is hours of picking the best of ten near-identical shots. Paid tools do this in the cloud or by subscription. shootflow cull FOLDER --out REPORT does the mechanical part locally with only ffmpeg installed.
It groups frames whose perceptual hashes sit within 8 bits of each other, ranks each group by a blend of sharpness, exposure, framing, grain and tilt, and marks the top frame the keeper. Blurred or badly exposed frames are rejects, and so are closed-eye frames when you point SHOOTFLOW_FACE_DETECTOR at a detector. --write-xmp adds sidecars (4 stars, 2 stars, -1) that Lightroom and Bridge read, and it never overwrites a sidecar that already holds your edits. Photos are only read.
What is not shown: the scores are technical signals, not taste, and nothing has been compared with a photographer's real picks yet. The tests use stubbed metrics plus two ffmpeg-made JPEGs. If you shoot events and will share a cull you did by hand, that is the data that would turn this from a plausible tool into a measured one.