Local-first AI photo management for Mac and Windows: diagnosis, cleanup, people recognition and metadata
LumiLab is built for photography enthusiasts and people with years of family photos. It indexes your photos where they already live (it doesn't move, copy or upload the files), then runs a diagnosis of the whole library: how many duplicates, similar shots, bursts and screenshots there are, and how much space you could free. From there you can clean up by quality and similarity, group photos by person with face recognition, and batch-edit titles, tags, copyright and more. All AI inference runs on your machine; metadata is written to XMP sidecar files next to the photos, the originals are never touched, and every change can be undone.
Add one or more photo directories (they can span several drives). After the background scan you get a diagnosis: photo and video counts and space used, date range, format distribution, camera and lens statistics, duplicate groups and extra photos, similar groups, screenshots, burst groups, reclaimable space and named people. Every item jumps straight to the matching cleanup or browse page.
Scores each photo on overall sharpness, exposure, face sharpness, eye sharpness, closed eyes and aesthetics, and groups by similarity. Set how many photos or what percentage each group keeps, define blur and exposure reject rules, use built-in presets like Balanced, Sharpness-first and Portrait-first, or import and export your own scoring rules.
Once you confirm the cleanup, rejected photos can go to a trash folder inside the library, to the system Trash / Recycle Bin, or be permanently deleted. Permanent deletion must be chosen by hand and comes with a clear warning.
Detects and recognizes faces on your machine and clusters them automatically into a people wall. You can name and rename people and merge them; merges can be undone for a short while. "Merge suggestions" asks one by one "are these the same person?", and you confirm, reject or postpone. False detections like animals and statues can be marked "not a face".
Face detection threshold, minimum face size and clustering merge threshold take effect immediately; no need to re-analyze the whole library. New photos only get incremental computation.
Multi-select and batch-edit rating, flag, color label, title, caption, creator, copyright, city, country, keywords, capture date and orientation. Each field can be kept, cleared or replaced. Changes are staged as a draft and only saved after you confirm.
Saved metadata is written to XMP sidecar files next to the photos; the original files are never modified. On rescan the information in the XMP takes precedence.
Every writeback is recorded. Review added, removed and modified content field by field, filter by writeback, rollback or failure, and roll back single records or in batch.
Grid, list, monthly timeline, people and loupe views; folder tree and virtual collections (deleting a collection does nothing to the originals). Filter by media type, rating, flag, color label, camera, keywords, person, date, folder and collection; search by filename, title, caption and camera.
In the loupe view, P / X marks pick or reject, 1–5 sets the star rating, ←→ switches photos. Common actions all have shortcuts, and you can open the original in the system app, copy the path or reveal it in Finder / Explorer.
Scanning, cleanup and people analysis all report progress in the task center and can be cancelled or retried.
The cleanup analysis computes several metrics per photo, then combines them into one score using the weights you set:
| Metric | What it catches |
|---|---|
| Overall sharpness | Camera shake and missed focus |
| Exposure | Over- and underexposure |
| Face sharpness, eye sharpness | Whether faces and eyes are sharp in portraits |
| Closed eyes | Skip the frame where someone blinked in group shots and portraits |
| Aesthetics score | Pick the better composition among near-identical photos |
On top of that you can set hard reject conditions (for example, reject outright below a blur threshold) and how many photos each similar group keeps. Rules can be saved as presets and exported as files to share.
LumiLab never rewrites your photo files. When you save edits to rating, keywords, title, copyright and the like, they are written to same-named XMP sidecar files next to the photos. The benefits:
A rollback is itself recorded, but rollback records cannot be rolled back.
Basic is free and gives you the full browsing and AI analysis experience. Applying results to your photos requires Pro, which comes as a monthly subscription, a yearly subscription or a lifetime purchase; prices are shown in the in-app store.
| Capability | Basic | Pro |
|---|---|---|
| Browsing, timeline, filter and search | ✓ | ✓ |
| Library diagnosis | ✓ | ✓ |
| People recognition and cleanup analysis | ✓ | ✓ |
| Edit metadata drafts (stored locally) | ✓ | ✓ |
| Confirm cleanup, remove photos | ✓ | |
| Write metadata to files | ✓ | |
| Save people names, merge people | ✓ |
License info is cached on your machine, so your entitlement keeps working when the network is unreliable.
Photos, thumbnails, face features and the index all live in the app’s data directory on your machine. Face analysis only runs when you start people analysis; the data is never uploaded or given to third parties. LumiLab has no account system. The network is only used for:
No. AI inference such as face recognition, similarity grouping and aesthetic scoring all runs on your machine; photos and feature data never leave your device. The network is only used for the first download of AI models and store purchase validation.
Not when you add a directory; it only builds a local index. Edited metadata is written to XMP sidecar files next to the photos, so originals are never rewritten. Only when you confirm a cleanup are rejected photos moved or deleted the way you chose.
Yes. By default rejected photos move to a trash folder inside the library or to the system Trash, and the scanner automatically skips those locations. Permanent deletion has to be selected manually in settings, and it is not the default.
Basic lets you browse photos, read the diagnosis, run people recognition and cleanup analysis, and edit metadata drafts. Confirming cleanups, writing metadata to files and saving people names and merges require Pro. Pro comes monthly, yearly or as a lifetime purchase.
It indexes and browses common formats like MP4, MOV and MKV. In the current version video metadata is read-only and cannot be edited.
Only when you run people analysis, and only on your machine. LumiLab stores face locations, the numeric features used to compare similarity, group identifiers and the names you type, all in the local app data directory. Face data is never uploaded or shared with third parties. It is deleted when you uninstall the app or remove the photos; the originals are never modified. See section 4 "Face Data" of the privacy policy for the full description.
Find the record in the change history and roll it back, for single records or in batch, with every field's before and after visible.
LumiLab is designed for offline local processing and does not depend on cloud transcoding. Read ourprivacy policyfor full data-handling details.