LYNX
STANDBY

LYNX — DOCUMENTATION

OVERVIEW

LYNX detects faces in uploaded images, generates ArcFace biometric embeddings, clusters similar faces using HDBSCAN, and matches against a local reference database stored in your browser. No biometric data is stored server-side.

ANALYSIS PIPELINE

DETECTION

InsightFace RetinaFace detects faces in each image. Faces below a minimum area threshold are discarded as too small to generate reliable embeddings.

EMBEDDING

Each detected face is encoded as a 512-dimensional ArcFace embedding — a numerical representation of facial geometry that is comparable across images.

CLUSTERING

HDBSCAN clusters all embeddings by similarity. Faces within a cosine distance of ~0.35 are grouped into the same identity cluster (S001, S002, etc). Faces that don't cluster with anything become unknowns (UNK001, etc).

REFERENCE MATCHING

After clustering, faces are compared against your local reference database. If a face is within distance 0.35 of a saved reference, it is labeled with that person's name and shown with a green bounding box.

IDENTITY TYPES

REFERENCE MATCH (green box)

Face matched a person in your local reference database.

CLUSTER SUBJECT (colored box)

Face grouped with other faces across images but not yet identified. S001, S002 etc.

UNKNOWN (grey box)

Face detected but not similar enough to any other face to form a cluster.

KNOWN LIMITATIONS

CLUSTERING ERRORS

HDBSCAN may group two different people into the same cluster if their embeddings are similar (same age, ethnicity, photo conditions). It may also split one person across multiple clusters if their appearance varies significantly between images. Reference matching corrects both errors.

SMALL OR OBSCURED FACES

Faces that are too small, at extreme angles, or with eyes obscured (sunglasses) may not be detected or may generate unreliable embeddings.

EDITING IDENTITIES

To edit identity labels you must first enable edit mode by clicking EDIT TAGS in the bottom bar. Edit mode is indicated by a visible toggle state.

RENAMING A CLUSTER

In edit mode, click an identity card in the sidebar. A text input appears — type a name and click SAVE AS REFERENCE. This saves the face embedding to your local reference database under that name. The cluster label updates to the saved name.

MERGING TWO CLUSTERS INTO ONE PERSON

If the same person appears as S001 and S003 (split cluster), save any face from S001 as "Jane". Then click a face from S003, type "Jane" in the name field, and save. Both clusters will now match "Jane" on the next analysis.

ADDING A BOUNDING BOX TO AN EXISTING REFERENCE

In edit mode, click directly on a bounding box in the gallery image. A popup appears with the face crop. Type the exact name of an existing reference person and click SAVE AS REFERENCE. The embedding is added to that reference, improving future match accuracy.

COMPARING A FACE TO A REFERENCE

In edit mode, click a bounding box to open the popup. Use the COMPARE dropdown to select any saved reference person. LYNX computes the cosine distance between that face and the reference embeddings.

  • 0.00 — identical
  • < 0.25 — strong match (green)
  • 0.25–0.35 — likely match (yellow)
  • > 0.35 — unlikely match (red)
  • > 0.60 — different person

REFERENCE DATABASE

References are stored in your browser's localStorage under the key 'lynx_references'. They persist across sessions on the same device and browser but do not sync across devices.

To manage references, use the REFERENCE page to view all saved people, their face thumbnails, embedding count, and delete individual pictures.

PRIVACY

Images are sent to the LYNX backend for face detection and embedding generation. They are not stored. Embeddings and reference data remain in your browser only and are never transmitted to any server.