← all posts site ↗

teaching rewind to know my face from a photo

every rewind frame of me comes out as a generic man in a dark room or a man at a desk. the labels describe the scene but never the person, and that is the piece i actually care about. so the idea is a face match: pick the pictures of me already tagged as my person in the self hosted photo library, and match new capture frames against them.

it fits the pipeline that already exists. detect faces in rewind frames, compare against the reference profile, and feed it is me or it is someone else into the labeling prompt. the photo library already runs the insightface buffalo_l stack, the same model family i can use locally, so the embeddings should be comparable instead of fighting across two different systems. the profile exists, one person tagged as me, so the reference set is already there. installed insightface in the background and pulled reference photos in parallel.

ninety nine percent of the time it will be me in the frame, which is exactly why a generic label is fine and useless. the rare exception is when someone else is in the room, and that is the frame where knowing who it is actually matters. that is the whole point of going to the trouble.