Type what you remember. Find the photo.

Keyword search finds the word “beach” in a file name or a caption. Semantic search finds the beach in the picture. Pixel Gallery does the second kind on the phone, with no server involved.

How it works

The app uses SigLIP2, an open model that maps images and text into the same vector space. During indexing, every photo and a set of frames from every video are turned into a 768-dimensional vector. When you type a query, the text is turned into a vector too, and the app ranks your library by how close each item’s vector is to the query. Close means similar meaning.

Results are not purely semantic. The app fuses the vector ranking with its classic full-text search over captions, labels, recognized text and file names, so a query that contains a real word such as a place name still benefits from exact matches.

What you can type

• Scenes: “sunset over water”, “snowy street at night”
• Objects and activities: “birthday cake”, “someone playing guitar”
• Moods and styles: “moody black and white portrait”
• Combinations: “red car in front of a mountain”

Queries work in natural language. Short and concrete phrases work best.

What it costs

• Storage: about 375 MB for the model, downloaded once through Google Play.
• Time: roughly two seconds per photo on a recent flagship. Indexing runs in the background and prefers a charger.
• Memory: the feature is hidden on devices below a RAM threshold, because the model would not run reliably there.

Privacy

Nothing about the search leaves the device. The model, the vectors and your queries all stay on the phone. The vectors are stored in the encrypted metadata database alongside faces and locations.

Related

Semantic search sits on top of the regular Search & Discover index, and captions from AI captions make keyword search richer at the same time. Setup steps are in the guide Enable semantic search.