Blog

How Face Recognition Works Without a Server

Detect, embed, cluster. How Pixel Gallery groups faces locally, how you correct it, and exactly what is stored and where.

September 14, 2026 · Maximilian Keppeler

“Face recognition” is a loaded phrase. In a gallery it means something narrow: putting the photos of the same person together. Here is how Pixel Gallery does that without any face ever leaving the phone.

Step one: detect

ML Kit’s face detector scans each photo and returns bounding boxes and landmarks for every face. This is the same on-device library many camera apps use for autofocus. It does not know who anyone is; it knows where faces are.

Step two: embed

Each detected face is cropped, aligned and passed through a small on-device model that outputs a compact vector. Two crops of the same person produce vectors that are close together; different people produce vectors that are far apart. The model was trained once, elsewhere, on public data. It is never trained on your photos and it never updates itself.

Step three: cluster

Vectors that sit close together become one person. The clustering is conservative: when two groups might be the same person, the app proposes a merge instead of deciding. You accept or dismiss. When a group contains an odd face out, you select it and split it into a new person.

What you add

Names. Cover photos. Which people are pinned, which are hidden. Optionally, a link to a phone contact, which adds call and message shortcuts to the person’s page. All of it is local.

What is stored, and where

Face crops are not stored. The vectors, the cluster assignments and the names live in the metadata database, which is encrypted with SQLCipher using a key held in the Android Keystore. That database is excluded from Android backup. Nothing is compared with an external database, and the app has no account that could carry face data to a server.

What this is not

It is not identification. Nobody is looked up. If you never name a cluster, it remains “Unnamed” forever. If you hide a person, they vanish from the People screen without deleting a single photo.

Corrections most libraries need

The first week usually involves a few merges and one or two splits. The guide Name, merge and link people to contacts walks through them. The People page lists everything the screen can do.

Features mentioned

More from the blog