ElasticFace
Elastic margin loss from CVPRW 2022: lets the decision margin breathe so identities separate more naturally.
About
ArcFace and CosFace fix one penalty margin for every identity, which is unrealistic when some identities have wild intra-class variation (age gaps, pose) and others don't. ElasticFace draws the margin from a Gaussian each training iteration, giving the decision boundary room to expand and retract as it learns.
On the authors' evaluation it beat both ArcFace and CosFace losses under the same geometry, advancing the state of the art on six to seven of nine mainstream benchmarks. The official PyTorch training code, evaluation scripts, and pretrained models are public. Like most academic releases it is weights-and-code rather than a product, but it slots straight into an InsightFace-style pipeline.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: LFW 99.53%, AgeDB-30 98.35% (best of nine tested on the age-gap set), CFP-FP 98.73%, and best or near-best across nine mainstream benchmarks (CVPRW 2022 paper). 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
Trust score — 7/10
Scored 7/10 (metrics, metrics-v1). Evidence: peer-reviewed CVPRW 2022 paper (arXiv 2109.09416) by named Fraunhofer/TU Darmstadt researchers; official training code, evaluation scripts, and pretrained models released. Strengths: rigorous nine-benchmark evaluation, reproducible code, elastic margin is a genuine loss-function contribution. Docked for: (1) no independent published benchmark; (2) modest community adoption and no commercial support. License: no explicit license stated on the release — treat as research use until confirmed.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · CVPRW 2022 (arXiv 2109.09416) by Fraunhofer IGD / TU Darmstadt researchers.
- · Elastic margin drawn from Gaussian; LFW 99.53%, AgeDB-30 98.35%, CFP-FP 98.73% provider-reported.
- · All numbers provider-reported; no independent benchmark found.