Skip to content
← Models
Open weightsMinchul Kim, Anil K. Jain, Xiaoming Liu (Michigan State University)/10 trustMetrics-checked

AdaFace

Quality-aware margin loss from CVPR 2022: face recognition that holds up on blurry, low-quality images where others fail.

Visit official page →

About

AdaFace starts from a simple observation: the same margin-based loss treats crisp studio portraits and blurry CCTV frames identically, and that hurts the hard cases. It approximates each sample's image quality from its feature norm and adapts the margin accordingly, emphasizing hard samples only when their quality justifies it.

The result is a face recognition model that shines exactly where older models degrade: low-resolution, motion-blurred, or heavily compressed faces. The official PyTorch code, training configs, and pretrained ResNet backbones are public, and the paper was presented as a CVPR 2022 oral. It is a research release with no commercial backing, so treat it as excellent open weights rather than a supported product.

Specs

Context window

Image input (aligned 112×112 face crops)

Pricing

Free — research code and pretrained weights

Free quota

N/A (self-hosted weights)

API access

No hosted API — download pretrained weights and self-host

Weights

Open weights

Released

Not disclosed

Benchmarks

Provider-reported: IJB-B TAR 94.42% at FAR 1e-4 (ResNet-50 trained on MS1MV2); reported error reductions of ~11% on IJB-B and ~9% on IJB-C vs the second-best baseline; strongest gains on low-quality sets (IJB-S, TinyFace). Numbers are provider-claimed; I found no independent published benchmark.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

face-recognitionlow-quality-face-recognitionface-verificationsurveillance-analytics

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics, metrics-v1). Evidence: peer-reviewed CVPR 2022 oral paper (arXiv 2204.00964) by named MSU researchers; official PyTorch code and pretrained weights released; demo shows clear low-quality advantage. Strengths: top-tier venue, reproducible public code, genuinely useful low-quality behavior. Docked for: (1) no independent published benchmark; (2) no explicit license stated on the release, and no commercial support or SLA. License: no license file visible on the official release — treat as research use until confirmed.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · CVPR 2022 oral (arXiv 2204.00964) by Minchul Kim, Anil K. Jain, Xiaoming Liu (MSU).
  • · Quality-adaptive margin; IJB-B TAR 94.42% at FAR 1e-4, strong on low-quality sets (IJB-S, TinyFace).
  • · All numbers provider-reported; no independent benchmark found.

Sources