Glossary · Technology
Face embedding
A face embedding is a fixed-length list of numbers, typically 128 to 512 values, that a neural network produces from an aligned face image. Faces of the same person produce embeddings that sit close together in that numeric space; different people sit far apart. Every modern face search and verification system compares embeddings, not pictures.
Updated November 6, 2026 · 1 min read
How is an embedding made?
After face detection and alignment, the cropped face is fed to a network trained on millions of labelled faces. Training pushes embeddings of the same identity together and different identities apart; FaceNet (2015) made this "triplet loss" approach standard, and later losses such as ArcFace sharpened it. The output is a vector that can be stored, indexed and compared cheaply.
How are two embeddings compared?
By a distance or similarity measure, most often cosine similarity, which gives a similarity score. A threshold turns the score into a decision; where the threshold sits trades false matches against missed matches. In a search engine the comparison is one probe against millions of stored vectors, which is one-to-many matching.
| Model | Vector length | Year |
|---|---|---|
| FaceNet | 128 | 2015 |
| dlib ResNet | 128 | 2017 |
| ArcFace (InsightFace) | 512 | 2019 |
| Commercial engines | Undisclosed, typically 256 to 512 | Ongoing |
Can a face be rebuilt from an embedding?
Partly. Research has shown that an approximate face can be reconstructed from some embeddings, which is one reason regulators treat stored embeddings as personal data and, where they identify a person, as biometric data. Embeddings are not anonymous just because they are numbers.
Where does it come up in law?
The GDPR defines biometric data by the processing that allows unique identification, which an embedding does. Illinois BIPA covers "scan of face geometry", which courts have read to include embeddings produced from photographs. Whether an engine stores embeddings of people who never consented is the central question in the cases against face search engines.
Questions
Is an embedding the same as a photo?
No. It is a derived numeric description. You cannot view it, but it can be compared, and in some systems approximately inverted.
Do two photos of me give the same embedding?
Similar, not identical. Pose, lighting, age and resolution move the vector; the comparison threshold absorbs small moves.
Sources
- 1.Schroff, Kalenichenko, Philbin, FaceNet: A Unified Embedding for Face Recognition and Clustering (checked March 12, 2015)
- 2.Deng et al., ArcFace: Additive Angular Margin Loss for Deep Face Recognition (checked January 23, 2018)
Entry reviewed November 6, 2026. Found an error? Tell us.
Related entries
- Term Similarity score Similarity score defined: the number from comparing two face embeddings, what it does and does not mean, thresholds, and why a high score is a lead, not proof.
- Term One-to-many matching (1:N) One-to-many (1:N) face matching defined: a probe face against a gallery of many, how it differs from 1:1 verification, and why errors grow with gallery size.
- Term Biometric data Biometric data defined: identifying data from a face, fingerprint, iris or voice, when a photo becomes biometric data, and the GDPR, BIPA and DPDP rules.
- Concept Embeddings and similarity What a face embedding is, how networks place the same face close together, how cosine similarity becomes a score, and why a high score is not proof of identity.
Cite this entry
FaceSearch Wiki editors. "Face embedding". FaceSearch Wiki, updated November 6, 2026. https://facesearch.wiki/glossary/face-embedding/