What on-device photo analysis actually means
More and more photo apps advertise themselves as “private” or “on-device,” but the term gets used in very different ways. Here’s the real difference, explained without marketing.
Two ways to analyze a photo library
With a server: the app uploads your photos (or a compressed version) to its own servers, where image recognition runs, faces get grouped, and categories get classified. It then returns the result to you. At some point, your photos existed outside your device.
On-device: recognition runs directly on your iPhone’s chip, using the frameworks Apple already builds into the operating system. No photo ever leaves the device, not even temporarily or in compressed form.
What frameworks make this possible
On iOS, a handful of system frameworks make serious analysis possible without anything leaving the device:
- Vision — image recognition: scene classification, face detection, text recognition (OCR), animal detection. It runs locally on the chip’s Neural Engine.
- Foundation Models — the Apple Intelligence language model that ships with the system. It can generate natural-language text from statistics the other frameworks already computed, on the device, with no network call. This is the piece that surprises people: a language model that isn’t an API.
- Natural Language — sentiment and entity analysis over text, including whatever OCR pulled out of your screenshots.
- Core Location — turning the GPS coordinates already embedded in your photos into place names.
- AVFoundation — composing video from the clips inside your Live Photos, locally, rather than rendering it on a server.
None of them require an internet connection. All of them keep working in airplane mode.
How to tell if an app actually does this
There’s no way to verify it from the outside with absolute certainty, but there are signals:
- It works offline. If photo analysis requires internet, something is probably being uploaded to a server.
- The privacy policy is specific. It should name the frameworks it uses (Vision, Core ML, Apple Intelligence) instead of generic phrases like “secure cloud processing.”
- There’s no “upload” delay. On-device analysis takes time because of local compute, not bandwidth — you shouldn’t see an “uploading your photos” progress bar.
How WrappedMe does it
All of WrappedMe’s analysis — categories, faces to group people, text recognized in screenshots, place names from the GPS your photos already carry — runs with Vision and Apple Intelligence directly on your iPhone. There’s no backend of our own, no photo upload, and the result (aggregated statistics, never the images themselves) is stored only on your device. You can read the full detail in our privacy policy.