ArcFace
The additive-angular-margin loss that became the default face-recognition embedding — LFW 99.83%.
About
ArcFace, published at CVPR 2019, introduced the additive angular margin loss: it pushes face embeddings apart on a hypersphere so that same-identity faces cluster tightly and different identities separate cleanly. The idea was simple enough to drop into any training pipeline, and it worked so well that ArcFace became the reference loss for nearly all face recognition research that followed — and the backbone of the open-source InsightFace toolkit.
The original paper reported 99.83% on LFW and 98.02% on YTF with a ResNet-100 trained on MS1MV2, essentially saturating those benchmarks. The honest caveat: this is the same technology that powers both phone unlock and mass surveillance. Face recognition deployments carry real privacy and consent obligations, and accuracy varies across demographics — anyone deploying it owes users a bias audit, not just a benchmark score.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 1801.07698): LFW 99.83%, YTF 98.02% (MS1MV2, ResNet-100); CALFW 95.45%, CPLFW 92.08%
- Third-party model card (py-feat): IJB-C 96.18% TAR @ FAR=1e-4 for the InsightFace release
- Widely reproduced; LFW is considered saturated by the field
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (TIER verified, metrics-v1). Evidence: peer-reviewed CVPR 2019 paper (arXiv 1801.07698) with LFW/YTF/CALFW/CPLFW evals; InsightFace toolkit with ~30k GitHub stars and released ONNX models; benchmarks reproduced across the field. Strengths: foundational, genuinely open, massively adopted. Docked for: (1) original training data (MS-Celeb-1M) was withdrawn over consent issues — later models use WebFace variants; (2) face recognition is dual-use with documented demographic bias and surveillance risks. License: MIT (training code per model card).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · CVPR 2019 paper arXiv 1801.07698; LFW 99.83%, YTF 98.02% (MS1MV2, ResNet-100); CALFW 95.45%, CPLFW 92.08%
- · InsightFace toolkit ~30k GitHub stars with released ONNX models; benchmarks reproduced across the field; third-party model card: IJB-C 96.18% TAR @ FAR=1e-4
- · Original MS-Celeb-1M training data withdrawn over consent issues; dual-use + demographic bias caveats