How it works
Privacy
Everything runs on your Mac. No photo, face or log leaves it. The network is used only to download the face model once and to check for updates.
The scan
- References: every face in the reference photos is found and turned into an embedding, a list of numbers that describes a face. The person common to the photos is chosen, photos of anyone else are dropped, and the average of the rest is the reference face. See Reference photos.
- Detection: in each photo, or each sampled video frame, SCRFD finds every face and five landmarks (eyes, nose, mouth corners). When it finds none, it tries once more on the image with a border added, which catches faces that fill the frame.
- Recognition: each face is aligned on its landmarks and embedded with ArcFace.
- Scoring: a file's score is its best face's cosine similarity to the reference face. A video takes the best across its sampled frames.
- Copy or move: files at or above the threshold are copied or moved when you click Execute.
Performance
Both models run on the Neural Engine and GPU through Core ML. Files are read and decoded on four threads and the models run on two, ahead of the file being reported: reading, decoding and recognition overlap.
On an M1 Max, a scan of 12-megapixel JPEGs runs at about 20 photos a second. HEIC photos take longer to decode.
Folder listing reads each folder once, with no per-file lookups.
The face model
Vaquero uses buffalo_l, a face model pack from InsightFace, the open-source face analysis project.
| Pack | Files | Size | Recognizer |
|---|---|---|---|
buffalo_l |
det_10g.onnx, w600k_r50.onnx |
183 MB | ArcFace R50, WebFace600K |
On first launch, the app downloads buffalo_l.zip from InsightFace's own release, checks it against a pinned SHA-256, and unpacks the detector and recognizer into ~/Library/Application Support/Vaquero/models/buffalo_l/. Model and update downloads resume after a quit or a cancel, and need no GitHub account.
A pack needs a detector (det_* or scrfd*) and a recognizer (w600k* or glintr*); other files are ignored.
The InsightFace pretrained models are licensed for non-commercial research use only. They are not part of Vaquero and not covered by its license.
