Meta's smart-glasses app ships a complete face-recognition stack, dormant by default

A teardown of version 273.0.0.21 of Stella, the Android companion app for Meta’s smart glasses, found a fully assembled on-device face-recognition pipeline sitting unused on stock accounts. The package includes three ExecuTorch models (SCRFD for detection, KPSAligner for alignment, and an oversized SFace variant producing 2048-dimension embeddings), a SQLite database under Meta’s RLDrive sync framework with person and face tables, a cosine-similarity vector index built on sqlite-vec dimensioned exactly to the embedder’s output, and a hardcoded Android notification channel called nametags_recognition that fires a ‘Person recognized’ heads-up alert on a match.

Invoking the handler directly with a test photo runs the pipeline end to end: an empty index writes the cropped face and its L2-normalized fingerprint as a (uuid.jpg, uuid.emb) pair into a mode-0700 NameTagsPending directory that survives reboots, while a preloaded index triggers a production notification deep-linking to a fb-viewapp://name_tags screen. That destination screen is missing from the navigation graph in this build, the user-facing ‘Connections’ widget never surfaces on an unenrolled account, and the researcher did not observe Meta server-pushing identity records into person_profiles.

The author is careful about the gap between capability and activation: this is not evidence that Meta is identifying people through the glasses today. It is evidence that the entire apparatus to do so — models, schema, vector search, write path, notification channel — is shipped, wired, and gated only by Meta’s choice not to populate the index or expose the UI. The staging directory’s structure, a face crop next to its biometric embedding awaiting a label, matches exactly what retroactive identification would require.