On-Device vs Cloud AI for Photos: What Actually Changes

“AI-powered” appears on every photo app’s listing. The word hides the only question that matters for privacy: where does the model run? This post explains the difference without jargon, and ends with a way to check any app yourself.

Where the bytes go

Cloud AI: the app uploads your photo, or a downscaled copy, to a server. A model there produces labels, faces, captions or search vectors. The results come back. Your photo now exists in at least two places, and the provider’s terms decide what happens to the copy.

On-device AI: the model is part of the app, or provided by the operating system. The photo is read from storage, processed in memory, and the results are written to a local database. Nothing leaves the phone. Airplane mode changes nothing.

Why cloud models are bigger

A server can hold a model with hundreds of billions of parameters and run it on hardware that draws more power than your house. A phone has a few gigabytes of memory and a battery. So on-device models are specialized: one for faces, one for text, one for labels, a compact language model for captions, a quantized vision-language model for semantic search. Each is very good at its narrow job and useless outside it.

The practical consequence: on-device apps deliver the same categories of features, sometimes with slightly lower accuracy, and cannot deliver the open-ended, conversational features that need a frontier model.

What on-device cannot do

• Understand who people are without you naming them.
• Answer questions that need world knowledge beyond what a photo shows.
• Run the largest generative edits.

If an app offers those and claims to be fully on-device, one of the two claims is wrong.

Latency, battery, offline

On-device processing has to fit into idle time. Well-built apps index while charging, in independent steps, and show progress. The reward is that everything works offline, forever, without a subscription that keeps the servers on.

How to check an app yourself

1. Read the permissions. Does it need location? A gallery does not; the GPS data is inside the photo. Does it need the advertising ID?
2. Turn on airplane mode. Does search still work? Do new photos still get labeled after a while?
3. Look for an account requirement. Cloud processing needs an identity to attach your library to. No account is a strong signal.
4. Read the privacy page. Does it list every network connection, or does it speak in generalities?

Pixel Gallery’s answers are on the Privacy first page, and the on-device AI page lists every step the pipeline runs. For a comparison with a cloud service, see Pixel Gallery vs Google Photos.