Meta Adds Facial Recognition Capability to Smart Glasses

Analysis of the Android build of Stella, Meta's companion app for its smart glasses, has revealed a full technical stack for on-device Facial Recognition.

The key technical characteristics are as follows.

• 3 ExecuTorch models: These include SCRFD for face detection, KPSAligner for alignment, and the SFace model, which generates 2048-dimensional biometric embeddings. These models are distributed in .pte format via Meta's asset delivery system.
• Vector database: The system leverages the sqlite-vec extension to build a cosine-similarity-based 2048-dimensional vector index, enabling a structure for local face search.
• Pipeline operation confirmed: Using test images, it was confirmed that the entire pipeline—from face detection to embedding generation, local index search, and sending a 'Person Recognized' notification—operates normally.

However, no direct evidence was found that this feature is enabled by default for general users, or that Meta is transmitting identity data via servers. In other words, all the components for facial recognition are assembled and operable within the device, but appear to be controlled by Meta's settings.