What Is a Face Embedding?

Neural networks trained on millions of faces learn to extract embeddings that position faces in a high-dimensional space — similar faces cluster together; different faces are far apart. ScanErase extracts an embedding from your reference photo and compares it against 2.4 billion embeddings indexed from content across the web and 200+ platforms. When two embeddings are close enough in the vector space, they are flagged as a potential match for human confirmation. The embedding captures the mathematical essence of your facial geometry without retaining the original pixel data.

Key facts about this term

1. Embeddings are computed from geometric facial relationships An embedding captures distances and ratios between facial landmarks — eye spacing, nose bridge width, jaw angle — not specific pixels. This makes it robust to changes in lighting, makeup, age, and image quality.
2. Similarity between embeddings determines match probability Two embeddings within a threshold distance in the vector space are flagged as a potential match. The threshold is calibrated to balance detection sensitivity and false positive rate.
3. Your embedding is deleted after each scan session ScanErase does not retain the face embedding extracted from your reference photo between sessions. Each scan generates a new embedding from your current reference photo.