A folder with 40,000 photos and no tags is a folder you cannot really search. You half-remember a shot of a red door somewhere in Lisbon, but scrolling by filename or capture date will never surface it. An AI image tagger closes that gap: it looks at each picture, works out what is in it, and attaches words you can later search or filter by. Below, you will see what an AI image tagger does, where the different approaches run, and how Memora handles tagging entirely on your own Windows PC.
What an AI image tagger actually does
At its core, an image tagger reads the pixels of a photo and produces text that describes the content: a caption such as "a dog running on a beach," plus individual keywords like beach, dog, and ocean. Those words get stored next to the file so you can filter your library by them or find them in a search. Old-school tagging meant typing every keyword by hand, which almost nobody keeps up with across tens of thousands of images. Most AI taggers do the first pass for you, then let you correct or add your own tags on top.
Two things separate a useful tagger from a frustrating one. Accuracy is the obvious one, since a model that labels your sunset as "a plate of food" wastes more time than it saves. What matters just as much is where the work happens, because sending your personal library to a server to be analyzed is a very different privacy trade than doing it on your own machine.
Auto-tags versus the tags you add yourself
Many people blur these together, but they are different jobs. Auto-tagging is the model guessing keywords from the image. Manual tagging is you deciding that a set of photos belongs to "Hendricks wedding" or "needs reprint," something no model can infer from pixels alone. Good workflows use both: let the AI blanket your library with rough content keywords, then you layer your own meaningful tags on the photos that matter.
Memora keeps both in the same place. Its indexing step writes an automatic caption and a set of keywords for every photo, and any tags you type are stored right alongside them in the local catalog. When you later filter by a tag, it makes no difference whether a human or the model put it there.
Where the tagging runs: cloud versus on-device
Cloud taggers, and the big photo services built on them, upload your images to their servers, run the model there, and send the labels back. That approach is fast and needs no capable hardware, but every photo you tag has now left your computer. For a wedding photographer holding client galleries, or anyone with family photos they would rather not hand to a third party, that is the sticking point.
On-device tagging flips the arrangement: the AI runs on your own CPU or GPU, and nothing about the image is transmitted. Memora is built this way, so all of its tagging, captioning, and search run locally and your photos never leave your computer. One honest caveat is that local models download once the first time you use them (the captioning model is a sizeable one-time download from Hugging Face), and the app does contact shutterdev.com for licensing and update checks, sending only an anonymous hardware hash, your license ID, the app version, and a photo count. None of that includes your pictures.
How Memora tags your photos
Under the hood, Memora runs two local models while it indexes. A BLIP model writes a short natural-language caption for each image. A zero-shot CLIP model assigns keywords by scoring the photo against candidate labels. Both run on your CPU, which is why captioning takes longer than ordinary browsing, and both write their output into the local database rather than anywhere online. Because the captions are generated by a model, they can occasionally be generic or plain wrong, so it helps to treat them as a searchable first draft rather than gospel.
Tagging is only half the story. Memora pairs it with AI semantic search, which lets you find a photo by describing it even when no tag matches. A local CLIP model converts both your photos and your typed query into vectors and ranks results by similarity, blending visual match, caption, and keywords. Type "person holding a red umbrella" and you get results whether or not anyone ever tagged the word umbrella.
Tags or semantic search: which do you reach for?
Tags earn their keep when you want a stable, repeatable bucket: every photo marked "portfolio" should always appear under that filter. Semantic search shines when you are hunting for something you never anticipated tagging, like "foggy morning" or "latte art." Having both means you are not forced to predict, at import time, every way you might later want to find an image. Having both is what makes a large library genuinely navigable instead of merely stored.
What to look for in an AI image tagger
- Where it processes your photos. Local processing keeps private images off other people's servers; cloud processing does not.
- Whether you can edit the tags. Expect to fix and extend them, since auto-tags are only ever a starting point.
- Bulk tagging and filtering. Tagging one photo at a time defeats the purpose on a big library.
- Search that reaches past the tags. Content-based search covers everything the keyword list happened to miss.
- Format coverage. If you shoot RAW, the tagger needs to read CR2, CR3, NEF, ARW, and the rest, not just JPEG.
Memora reads all the common formats, from JPEG and PNG through HEIC and the main RAW types, so the tagger works on your real originals and not only on phone snapshots. Worth noting: its AI features apply to photos, so videos are managed and played but not auto-tagged.
Free versus Pro
Basic AI semantic text search ships in the free tier of Memora, so you can try finding photos by description before paying anything. Managing tags and filtering by tag sit in Pro, alongside face recognition, AI culling, and smart albums. Memora uses a one-time Pro license rather than a subscription, and current pricing lives on the Memora Pro page. It runs on Windows 10 and 11. For a wider look at how it all fits together, the photo organizing software overview walks through the full picture.
Frequently asked questions
How does automatic photo tagging work?
A vision model looks at each photo and predicts words that describe it, both a sentence-length caption and individual keywords. Memora does this during indexing with local BLIP and CLIP models, writing the results into your catalog so you can filter and search by them.
Are my photo tags and images private?
With a local tagger like Memora, yes. The models run on your own PC, and neither your photos nor the generated tags are uploaded. Outbound traffic is limited to licensing and update checks (an anonymous hardware hash, your license ID, the app version, and a photo count), which never include the images themselves.
Can I tag photos in bulk?
Automatic captions and keywords are generated for your whole library as it indexes, so the first tagging pass is already a bulk operation. Editing or adding your own tags across a selection is part of the Pro tag tools.
Does an AI image tagger work offline?
Once the models have downloaded the first time, Memora tags and searches with no internet connection at all. Offline capability is the real advantage of on-device AI: after setup, your library and the computer in front of you are all it needs.
Want to tag and search your own library privately? See how Memora's semantic search finds photos by description, or read more about the app on the Memora overview.