How to Get AI Face Recognition on Your Photos (Without Sending Them to Google)

You know the Google Photos trick where you tap a friend's face and instantly see every photo of them across ten years? That feature is the main reason people feel stuck on Google Photos. The good news: you can have the exact same thing running on a little box in your own home, with the AI looking at your photos locally and nothing ever leaving the house. The app is called Immich, it comes preinstalled on SparkBox, and this guide is about the one feature that makes people switch — grouping every photo by who's in it.

Tested on: UGREEN DXP4800 Plus running SparkBox v1.6.159 and Immich v2.7.5. The same steps work on any SparkBox install (NAS, mini-PC, or VPS).

1. What face recognition actually gives you

Once this is working, Immich gives you a People view. It's a row of faces — your kids, your partner, your parents, that friend who's in half your vacation photos. Tap any face and you get every single photo that person appears in, sorted by date, going back as far as your library goes.

It also quietly powers other things. Search for "Mum at the beach" and it can combine the person with the place. Make a shared album and Immich can suggest which photos to add based on who's in them. None of this needs you to tag anything by hand — the whole point is that the computer does the looking.

The part that matters for this guide: every bit of that runs on your own hardware. The AI model that recognises faces lives in a small piece of SparkBox called the machine-learning container. Your photos are never uploaded anywhere to make this work. That's the difference between this and Google Photos — same feature, but the face data stays in your house.

2. Why this normally only works on Google's servers

Face grouping feels like magic because for fifteen years it only existed inside giant cloud services. Google, Apple, and Amazon ran the AI on their machines, on your photos, and handed you back the tidy result. The catch was always the same: to get the feature, you had to give them the photos.

Doing it yourself used to mean installing AI tools by hand, wrangling Python, and downloading models from the internet — the kind of thing that makes most people close the laptop. Immich packages all of that into one app, and SparkBox preinstalls Immich along with its AI engine, the database it needs, and a cache, so the hard part is already done before you start. What's left is just flipping the right switches and being a little patient. That's the whole job.

If you haven't set up Immich at all yet, start with the full Immich setup guide first — get logged in and your photos importing — then come back here for the face part. And if the whole idea of running your own apps is new, the no-terminal beginner guide explains how SparkBox does the heavy lifting.

3. Make sure the AI brain is switched on

SparkBox ships Immich with the AI engine turned on by default, so for most people there's nothing to do here except check. But it's worth a thirty-second look, because if someone turned it off to save memory, no faces will ever appear and you'll wonder why.

Open Immich in your browser (it's the camera tile on your SparkBox dashboard, or go straight to http://YOUR-NAS-IP:2283). Log in as your admin account — that's the first account you registered when you set Immich up; there's no default login. Then go to:

Administration → Settings → Machine Learning

Make sure the top toggle is Enabled. While you're here you'll also see toggles for "Facial Recognition" and "Smart Search" — leave those on too. That's it. You don't need to touch any of the technical boxes below them; the defaults SparkBox ships are the right ones.

One thing worth knowing: the very first time the AI engine runs, it downloads its face-recognition model — about 500MB — from the internet and saves it permanently. If your NAS is fully internet-blocked (some privacy setups lock it right down), that download can fail silently and faces never appear. If that's you, allow the box online just long enough for the first run; once the model is saved, you can lock it back down and face recognition keeps working offline forever.

4. Let Immich find the faces

Here's the part that trips people up, so read this bit slowly: faces don't appear instantly. When you add photos, Immich runs a series of background jobs on each one, and finding faces is two of those jobs working in order:

1. Face Detection — the AI looks at each photo and works out "there's a face here, and here, and here." This is the slow one.
2. Facial Recognition — it then compares all the faces it found and groups the ones that look like the same person together.

You can watch both happening live. Go to Administration → Jobs and you'll see a list with a counter next to each one. Find Face Detection and Facial Recognition — when their counters are ticking down, the AI is working. When they hit zero, it's done, and your People view will be full.

How long does it take? On a low-power box like a NAS or a mini-PC, expect a few hours per ten thousand photos for the full run. A fresh library of a few hundred recent photos is done in minutes. A decade-long Google Photos export is an overnight job — start it before bed and the People view will be waiting in the morning. The box can be doing other things meanwhile; this just runs quietly in the background.

If you ever add a big batch of old photos and the faces don't show up, come back to this Jobs page and click Facial Recognition → Missing to nudge it to process anything it skipped.

5. Name your first ten people

This is the satisfying part. Once Facial Recognition has run, open the People view — it's in the main menu, sometimes under "Explore" depending on your version. You'll see a grid of faces, each one a cluster of photos Immich thinks is the same person. They're unnamed for now — just faces.

To name one:

1. Click a face that you recognise.
2. You'll see every photo Immich grouped under that person. Have a quick scroll to check they're all actually the same person (they usually are).
3. Click the name field at the top, type the person's name, and confirm.

That's it — that name now sticks to every one of those photos, and any new photos of that person get sorted in automatically from now on. Do this for your ten most common people first: family, close friends, anyone who shows up a lot. Those ten names cover most of your library and make search genuinely useful straight away.

A nice touch: once a person is named, you can set their best photo as the thumbnail, and you can mark people you care about as favourites so they sit at the front of the People view.

6. Merge, hide, and fix its mistakes

The AI is good, not perfect, and the most common quirk is that it splits one person into two or three clusters — usually because of big changes like glasses, a beard, a baby photo versus a grown-up photo, or just a bad angle. Fixing this takes seconds:

• Merge duplicates: name the first cluster (say "Sarah"). When you start naming a second cluster that's also Sarah, Immich notices the name already exists and offers to merge them. Say yes, and the two piles become one.
• Hide junk clusters: you'll get clusters of strangers in the background of photos, statues, posters, that sort of thing. You don't have to name those — just hide them so the People view stays tidy. There's a hide option on each person.
• Reassign a stray photo: if one photo got filed under the wrong person, open it, and there's an option on the face to detach it or move it to the right person.

You don't need to do all of this on day one. Name your main people, hide the obvious junk, and let the rest sort itself out as you browse. The library gets tidier every time you touch it.

7. The honest RAM and speed reality

This is the section the cloud services never have to write, so here's the straight version. Face recognition is the most demanding thing Immich does, and on small hardware you'll feel it.

• Memory matters most. The AI engine wants room to work. As a whole, face recognition is happiest on a box with 4GB of RAM or more — that's SparkBox's own recommendation for the machine-learning features. On a 2GB machine it'll still run, just slower, and you may want to do your big import overnight.
• It competes with everything else. While the AI is grinding through an old library, the rest of Immich (and your other apps) get less of the processor. That's why a fresh import can feel sluggish for a day and then settle down completely once the jobs finish.
• You can turn it off. If your box is tiny and you decide you only want the photo timeline and albums, not the AI, you can switch Machine Learning off in Settings and free up the memory. You keep timeline view, albums, search by date and place, and phone auto-backup — you just lose face grouping and the smart "find my dog" search. The main Immich guide walks through turning it off cleanly if you go that route.

The trade is simple and it's yours to make: the cloud gives you face recognition for "free" by reading all your photos; doing it at home costs you a bit of memory and some patience on the first run, and in exchange nobody else ever sees a single face. For a lot of people that's an easy call.

Next steps