Your phone is the AI. Not someone else's server.

Most galleries that call themselves smart get their smarts the same way: your photos go to a server, a large model looks at them, and the results come back. Pixel Gallery takes the other road. Everything that makes the library searchable and organized is computed by models running on the phone itself, in a multi-step pipeline that works through your media in the background.

What the pipeline computes

Each photo and video passes through a series of independent steps. A step that cannot run on your device is skipped, never outsourced.

• Faces: detection with ML Kit, then face embeddings and clustering that turn thousands of faces into a handful of people. See People.
• Labels: ML Kit image labeling gives every photo a set of tags such as food, mountain or document, which power Discover and search.
• Text and codes: ML Kit text recognition in five scripts, plus barcode and QR detection. See Text & Codes.
• Captions and descriptions: Gemini Nano writes a short caption and a longer description on supported devices. See AI captions.
• Semantic embeddings: the optional SigLIP2 model turns images and video frames into vectors so you can search by meaning. See Semantic search.
• Subject segmentation: separates the subject from the background for background blur and cutouts in the editor.
• Colors: dominant and accent palettes for browsing by color.
• Quality score: sharpness, exposure, faces and composition combine into a score used by Stories and Cleanup.
• Provenance: Content Credentials and generator metadata mark AI-generated and AI-edited media. See AI content detection.
• Places and stacks: reverse geocoding of embedded GPS data into places, and clustering of bursts into photo stacks.

Why on-device is a real trade-off, not a slogan

On-device models are smaller than the ones in a data center. A phone cannot run a model with hundreds of billions of parameters, and Pixel Gallery does not pretend it can. What it can do is run purpose-built models for detection, labeling, recognition and embeddings, and a compact language model for captions, fast enough to index a whole library while you sleep.

The upside is total: no account, no upload, no monthly cost for servers, and search that works on a plane. The downside is honest: some queries a cloud model would understand, such as “photos of my grandfather”, are out of reach because no on-device model knows who your grandfather is. You name the person, and then the query works.

Where the results live

All derived data lands in a local database that is encrypted with SQLCipher. The key is generated on the phone and kept in the Android Keystore. Your original files are never modified. If you delete the app’s data or uninstall, the index is gone and your photos are untouched.

Managing the models

The AI Models screen lists the models that are optional, their size, and whether they are installed. The semantic search model is roughly 375 MB and can be restricted to Wi-Fi. Models that Google Play services provides on demand, such as the text recognizers, are handled by the system.

The Media Processing screen shows each step, how far it has progressed, and lets you reset a step if you want the app to recompute it. It also carries the plain statement that describes the whole design: your media never leaves your device.