Skip to content
github-actions[bot] edited this page Nov 21, 2026 · 2 revisions

clippa faces <stream> (#54) finds who is on screen: every keyframe (one every ~4 s on Kick) through two small face models OpenCV runs on the CPU, no GPU or Modal. An 8 h stream takes ~15 min on a 16-core laptop, most of it reading the video off the NAS share. It is a worker stage (faces in [worker] stages), run after the fast lane and before the slow cut, so the picks and Shorts reach review without waiting and the segment and compilation covers get it; run it by hand before clippa thumbs for anything else. Enquirer covers (prompt v5) then get each named person's best frames as candidates, labelled with who is in them, and use the found head box when the main face is one (face_found in thumbnails.json). The weekly takes each named person's best faces from every stream of the week ([weekly] face_frames, 3) ahead of the thumbnail stage's frames. A machine without the faces extra must take faces out of [worker] stages; clippa doctor fails until one or the other is done.

  1. uv sync --extra faces (add it to the other extras you use). The models (YuNet and SFace, from OpenCV's model zoo, MIT and Apache-2.0) download on the first scan into [paths] faces/models, checked against pinned hashes.
  2. clippa faces <stream> writes faces.json and faces/F01.jpg... one contact sheet per group of faces, biggest first. Groups that aren't people (game characters, puppets, a poster on the wall) are fine to leave unnamed.
  3. Name the regulars once from the sheets: clippa faces <stream> --name F01=Andy --name F03=Jer --enroll keeps them in [paths] faces/known.json, one file for every show. Later streams are named by face ([faces] match). Enroll from an ordinary stream, not a costume one; a second stream's --enroll adds another sample. clippa faces --enroll-photos adds the [thumbnail] photos portraits under their file names.
  4. Or do it in the war room (War room): name and enroll groups from their sheets, or draw a box round a face on a paused video ("Who's this?", B) to keep one that never made a group. clippa faces <stream> --at 1:23:45 --box x,y,w,h --person Name is the same from the command line; clippa faces --forget Name drops someone.
  5. clippa doctor --role gpu lists the known faces.
  6. For covers, keep a few photos of each regular in the cast library (clippa cast grab <stream>, or the war room's Cast window; see Cast library).

Clone this wiki locally