Web & mobile apps

Is on-device AI more private than sending photos to the cloud?

The short answer

Usually, yes: if the model runs on the person's own device, the photo never leaves it, so there is no copy on a server to secure, retain or leak. The trade-off is that on-device models must be small enough for a browser or phone, which suits bounded tasks rather than everything.

Updated September 26, 2026

What on-device means

The model runs on the user's hardware — in the browser or app — rather than on a server anyone else operates. The input is analysed where it was captured.

Why it is more private

No upload means no stored copy, no transfer to secure and no retention policy to write for it. Atelier, our own room visualiser, was built this way on purpose: 100% of its segmentation runs in the browser and no image ever leaves the device. Saved projects stay in the browser's local storage.

What it costs

A model small enough for a browser does narrower work than a large server-side one. Deciding which pixels are floor, wall or cabinet is bounded work on one image, which is why it runs well on-device; open-ended reasoning over lots of data usually does not.

When it fits

Photos of homes, documents and people are the obvious candidates, because they are the inputs customers are least comfortable uploading.

Asked next

Does anything leave the device when Atelier analyses a room?

The room photo does not: segmentation, the perspective maths and the material rendering all run in the browser. What Atelier loads is the retailer's product catalogue, which is product data rather than anything about the customer.

Is on-device AI slower than cloud AI?

It depends on the device. Atelier uses WebGPU where available and WASM where not, so newer hardware is faster; in exchange there is no round trip to a server and no inference tier to scale behind it.