librefacerec-l

512-d face-embedding model for LibreYOLO's facial-recognition (embed) task: aligned 112x112 RGB face crop in, L2-normalized 512-d identity embedding out (iResNet100, ArcFace input convention). Verification and identification are cosine similarity on these embeddings.

Usage

from libreyolo import LibreYOLO, FaceGallery

model = LibreYOLO("librefacerec-l")          # auto-downloads this file

# 1:1 verification
model.verify("a.jpg", "b.jpg", threshold=0.4)

# 1:N identification
gallery = FaceGallery(embedder=model)
gallery.enroll("alice", ["alice1.jpg", "alice2.jpg"])
results = model("group.jpg", gallery=gallery)
print(results.identities.name)

libreyolo compare model=facerec-l source=a.jpg source2=b.jpg
libreyolo enroll  model=facerec-l source=people/ gallery=team.gallery.npz
libreyolo predict model=facerec-l source=group.jpg gallery=team.gallery.npz

The default face detector (librefacerec-det) is auto-downloaded separately.

Provenance and license

• Mirrored unmodified from fal/AuraFace-v1 glintr100.onnx (Apache-2.0), renamed to librefacerec-l.onnx.
• SHA-256: a7933ea5330113b01c9b60351d8f4c33003f145d8470ac5f0e52ee2effe25c60 (verified identical to upstream at mirror time, 2026-07-28).
• Training data: not disclosed by the upstream authors (described only as a commercially licensed face dataset). Evaluate fitness for your use case accordingly.
• Only this single file is mirrored; other files in the upstream repository are third-party artifacts under different terms and are NOT mirrored here.

Accuracy

Measured 2026-07-28 through the full LibreYOLO pipeline (detect, 5-point align, embed) on the standard LFW dev pairs, using the funneled originals. The decision threshold was selected on pairsDevTrain (1100 pairs) and accuracy is reported on the held-out pairsDevTest (1000 pairs, 500 same and 500 different), so the headline number is not tuned on itself.

Metric	Value
ROC-AUC	0.980
Accuracy @ threshold 0.227 (selected on train)	96.9%
Accuracy @ threshold 0.400 (library default)	95.6%
Mean cosine, same person	0.572 (std 0.159)
Mean cosine, different people	0.063 (std 0.063)

Operating points on the held-out split:

Threshold	False accept	False reject	Accuracy
0.30	0.0%	5.4%	97.3%
0.40	0.0%	8.8%	95.6%
0.50	0.0%	25.0%	87.5%

The distribution is heavily separated in the false-accept direction: no different-person pair in this split scored above 0.30. The library's 0.4 default is therefore conservative, trading recall for near-zero false accepts, which suits access-control style uses; lower it toward 0.3 when missed matches cost more than false ones.

For calibration, note that LFW is a saturated benchmark where leading ArcFace-family models report roughly 99.5% and above. This pipeline's 96.9% is usable but not state of the art, and it is a single benchmark on a frontal, celebrity-photo distribution. Evaluate on data representative of your deployment.

Responsible use

This model produces biometric identifiers. It is intended for consent-based applications (device unlock, photo-library organization, opt-in event galleries). Remote biometric identification of individuals in public spaces is prohibited or heavily restricted in several jurisdictions (EU AI Act, BIPA and similar laws), and compliance is the deployer's responsibility. Accuracy varies across demographics; calibrate thresholds for your population and application.