Your phone probably holds thousands of photos — birthdays, holidays, screenshots, receipts, that sunset you meant to print. Finding one of them means scrolling, and scrolling, and scrolling. Cloud photo apps solved this years ago by letting you search by description: type "beach sunset" and the photo appears. The catch is that the search only works because your entire photo library was uploaded to someone else's servers.
On-device AI offers the same trick without the upload. A small model on your phone reads your photos locally, builds a searchable index on the device, and lets you find pictures by describing what you remember. Nothing leaves the phone.
How searching by description actually works
Behind the search box is a technique called image embedding. It sounds technical, but the idea is simple:
- The model looks at each photo. A small vision model — the kind that runs on a phone — analyses the image and produces a compact list of numbers that captures what the photo shows: the objects, the scene, the colours, the activity.
- Your search text gets the same treatment. When you type "red bicycle in front of a cafe", the model turns those words into the same kind of numerical description.
- The app compares the two. Photos whose descriptions are closest to your words rise to the top of the results.
This is why you can find a photo you never tagged. Nobody labelled that picture "dog on a beach" — the model worked out what it shows by looking at it. It is also why searches can be surprisingly flexible: "birthday cake with candles", "my kid on a swing", "receipt from the hardware store" all work because the model understands meaning, not just keywords.
The models that do this on phones are compact versions of the same idea behind large cloud vision models. They are smaller and faster, tuned to run on a phone's processor, and they trade a little accuracy for the ability to run entirely offline.
What you can actually search for
Once your gallery is indexed, the searches that work well are the everyday ones:
- People and pets. "Photos of the dog", "Mum at the beach", "the kids playing football". Face grouping on-device can cluster the same person across photos, though this works best with a modest library and clear faces.
- Places and scenes. "Mountain hike", "city at night", "the kitchen renovation", "airport". The model recognises environments, not just objects.
- Objects and food. "Red car", "birthday cake", "the blue dress", "coffee". Useful when you remember the thing but not the date.
- Documents and screenshots. "Passport photo", "boarding pass", "the wifi password screenshot", "receipt". Text in images is often readable by the model too, which makes this one of the most practical uses.
- Activities and moods. "Swimming", "camping", "christmas morning". These are looser queries, and results vary, but they usually surface the right album.
Searches that are still hard for small on-device models: very specific details ("the photo where Sarah is holding the blue mug, not the green one"), counting ("photos with exactly three people"), and fine distinctions between similar-looking objects. If a search misses, try a simpler description — one or two key nouns usually beat a long sentence.
Setting it up: index once, search forever
The workflow is the same across on-device photo search apps:
- Download the search model once. This is the only step that needs internet. The model is small — typically a few hundred megabytes — and it downloads over Wi-Fi.
- Let the app index your gallery. The app reads your photos and builds the local index. On a modern phone, a few thousand photos take minutes, not hours, and the app usually does it in the background while you do something else. Indexing uses battery, so it is best done while charging.
- Search by typing what you remember. From then on, results appear in a second or two, with no connection needed.
New photos get added to the index automatically when the app runs, so the search stays current without any maintenance. The index lives on your phone, alongside your photos — if you delete the app, the index goes with it.
The privacy difference, plainly stated
With cloud photo search, the deal is clear: your photos are uploaded, stored on a company's servers, and analysed there. That is how the search works, and it is also how the company can offer it for free — your photos become part of their system, subject to their policies, their breaches, and their business model.
With on-device search, there is no upload at all. The photos never leave the phone, the index never leaves the phone, and there is no account that ties your face, your family, and your travel history together in a database. Your gallery is, for the first time, searchable without being shared.
This matters most for the photos you would hesitate to upload: pictures of your children, medical images, scans of documents, photos of your home. With cloud search you have to decide photo by photo what is safe to upload. With on-device search, the question does not arise.
There are honest limits. On-device search cannot back up your photos — you still need a backup strategy, and that backup will live somewhere, whether it is a hard drive or an encrypted cloud service. And on-device search cannot find a photo on a phone you do not have with you. What it does is remove one specific exposure: the search itself no longer requires handing your library to a third party.
Photo organization beyond search
Search is the headline feature, but the same on-device understanding powers a few other useful tricks:
- Automatic albums. The app can group photos by trip, event, or person — "the beach holiday", "last Christmas" — without you creating a single folder.
- Duplicates and near-duplicates. Burst shots and accidental re-takes are easy to find and clean up when the model can tell two photos are nearly identical.
- Finding the best shot. Some apps can rank similar photos by sharpness and framing, so you can keep the best one from a burst and delete the rest.
- Memories, locally. A "this day last year" or "best of the month" view, generated on the device from your own photos, with no server involved.
None of these need the cloud. They are all just the same local understanding of your photos, applied in different ways.
Where on-device photo search falls short
It is worth being honest about the gaps, because they affect whether this is right for you:
- Accuracy is close, not equal, to cloud search. Cloud services run much larger models on server hardware. For most everyday queries you will not notice the difference, but for very specific or unusual queries, the cloud model is more likely to get it right.
- Face recognition across huge libraries is weaker. Cloud apps have spent years tuning face grouping on millions of photos. On-device grouping works well for a family-sized library, but it can struggle with very large collections or photos taken years apart.
- Indexing takes time and battery. The first index of a large gallery is a real job for a phone. Do it on Wi-Fi, plugged in, and let it finish before judging the results.
- It cannot search photos that are not on the phone. If your photos live only in a cloud backup, on-device search cannot see them. The search covers what is stored locally.
For most people, the trade is a good one: slightly less accuracy on edge cases, in exchange for a gallery that is never uploaded anywhere.
How this works in LLM Hub
LLM Hub includes on-device media search built on a small embedding model: it indexes the photos and videos in your gallery locally, and you search them by describing what you remember — no uploads, no accounts, no tracking. Because it runs fully on-device, it works in airplane mode and in dead zones, and your photos stay on your phone throughout.
To use it, download the media search model once over Wi-Fi, let it index your gallery while the phone charges, and then search by description whenever you need a photo. It sits alongside the rest of LLM Hub's on-device tools — chat, image generation, translation and transcription — so the same phone that searches your photos can also draft a caption for one, translate a sign in it, or turn a voice note about it into text.
A few tips for better results
- Describe the most distinctive thing. "Red bicycle" beats "bicycle on a street in summer". One or two strong nouns usually win.
- Try a synonym if the first search misses. "Beach" and "seaside" may return different photos; the model understands both, but the ranking shifts.
- Search for the scene, not the date. You do not need to remember when a photo was taken — that is the whole point. "Birthday cake" will find it faster than scrolling to March.
- Keep the index healthy. If you move photos between folders or apps, let the search app re-index. A stale index is the most common reason a photo seems to vanish.
- Use it for documents too. Screenshots of tickets, passwords, receipts and addresses are some of the most-searched images on any phone, and on-device search finds them without uploading your paperwork.
The bottom line
Photo search by description used to mean uploading your life to a cloud service. On-device AI removes that condition: the same search, running on your own phone, with your photos never leaving it.
If your gallery is the part of your phone you would least want uploaded — and for most people it is — that is the feature that matters most.

