Meta’s companion app Stella ships a complete facial‑recognition stack on Android, including three neural models, a 2048‑dimensional embedding engine, and a cosine‑similarity vector index. The pipeline runs end‑to‑end on a test image, proving the machinery exists.

The models—SCRFD for detection, KPSAligner for alignment, and SFace for embedding—total roughly 100 MB. They match open‑source architectures but are scaled for Meta’s use, with larger embeddings and a 2048‑dimensional output.

A local SQLite database stores face embeddings and links them to person records. When a face matches, the app fires an Android notification titled “Person recognized.” Unmatched faces are written to a protected folder as JPEG‑embedding pairs, ready for future labeling.

Although the code is present, the UI and backend sync remain gated for non‑enrolled users. Meta keeps the recognition apparatus on the device, assembled but inactive for the average user.