-
Notifications
You must be signed in to change notification settings - Fork 14
Help People Faces
LrGeniusAI provides two face-related workflows:
- People — browse face clusters (persons), assign names, and jump to Lightroom collections per person. Its interface opens in your web browser.
- Find Similar Faces — select a photo containing a face and find other photos featuring the same person.
Face data is generated during Analyze & Index Photos — tick Enable face detection on the General tab. It needs the face models on disk; they arrive with the Download AI models button in the plugin settings, alongside the search and species ones. Without indexed face data neither workflow finds anything, and the People page stays empty until you have also pressed Cluster faces at least once.
Library → Plug-in Extras → People...
People opens as a page in your default web browser, served by the LrGeniusAI
backend at http://127.0.0.1:19819/v1/ui/people. Lightroom shows a progress bar
named People (open in your browser) while the page is live; it is what
carries your selections back into the catalog, so leave it running and cancel
it when you are done.
If the browser never comes up — a blocked pop-up, no default browser — the plugin tells you after about half a minute and gives you the address to open by hand.
A grid of persons from your indexed photos. Each card has:
- A thumbnail of the representative face.
- A name field, or Unassigned faces for the faces not yet grouped into a person.
- The number of photos this person appears in and how many faces are in the cluster.
Named persons come first, then unnamed ones; within each group the person with the most photos is first. Every face that belongs to no person — never clustered, or left over by the last run — shares a single Unassigned faces entry. The Filter by name box and Unnamed only checkbox narrow a long list; Refresh re-reads it from the backend.
Detection during indexing finds faces; grouping them into persons is a separate step. Press Cluster faces to (re-)group everything the backend has. The run reports how many persons and faces it produced, and the grid updates itself.
Anything you decided by hand — a face you moved, two people you merged — is kept. Those faces are pinned: clustering rebuilds the guesses around them instead of overwriting them, so re-clustering is never a step that loses work.
The threshold next to the button is the cosine distance clustering works at. Leave it at 0.5 unless the result is wrong in a specific way: lower it towards 0.45 when different people were merged into one cluster, raise it towards 0.55–0.65 when one person was split across several.
Tuning opens the quality gate underneath it. A blurred, tiny or half-covered face is the one that sits between two people and welds them together, so faces below the gate are not allowed to start a person — they are attached to a finished cluster afterwards, or left unassigned. Raise the minimum sharpness or detection score if strangers keep landing in the same cluster; lower them if too many faces end up unassigned.
After each run the page prints what actually happened: how many faces were pinned, gated out, attached afterwards, and — the two numbers worth reading — how tightly the clusters sit (spread) and how many pairs of separate people ended up closer together than your threshold. Many such pairs means one person is split across cards and the threshold should go up. It also says in plain words which way to move it.
Type into a card's name field and press Enter or click away — each name is saved on its own, and the card says Saved · confirmed when it lands. Names survive re-clustering and are used as context elsewhere.
As you type, existing names are offered below the field, best match first. It matches more than the beginning of a name: schm finds “Anna Schmidt”, and ms finds “Maria Schmidt” by its initials. Arrow keys move through the list, Enter takes the highlighted one, Escape closes it. Each suggestion shows how many photos that person has, which is what tells two similar names apart. Use it — a name typed slightly differently the second time makes a second person.
Naming a person confirms it. The name is your judgement that this cluster really is one person, so all of its faces are pinned exactly as if you had placed them by hand, and the card gets the confirmed badge. The next Cluster faces run then keeps them together instead of re-deciding them — otherwise a name could quietly end up on a different set of faces. Clearing the name later does not unpin them; take wrong faces off inside the person instead.
Entering a name another person already has is the other way to merge two cards. The page asks first, showing both people side by side, and offers three answers:
- Merge into one person — they become one, keeping that name.
- Keep them separate — both keep the name. This is the right answer when two different people really do share a name; the suggestion list then shows it once and says how many people have it.
- Cancel — nothing is saved and the field goes back to what it was.
Double-click a card to open that person and see every face in it, not just the one on the thumbnail. This is where you check whether the clustering was actually right.
Each face shows how much it looks like the rest of the group. A face further from the group than the clustering threshold is marked as an outlier, and a face that failed the quality gate says which measurement was bad. The sort control offers least like this person first, most like this person first and best quality first — the first of those is the fast way to spot a face that does not belong, so it is what the view opens on.
Click faces to select them (shift-click for a range), then:
- Move to… — pick another person in the dropdown next to it and the faces go there.
- Move to a new person — split the selected faces off into a person of their own, with a name if you give one.
- Not this person — detach them; they return to Unassigned faces.
Every one of these counts as a decision by you, so it survives the next Cluster faces run and the next re-index of those photos.
When the same person ended up on two cards, drag one card onto the other. A dialog asks which name the merged person should keep — either of the two, or one you type — and then all the faces move across. Typing an existing name into a card's name field does the same thing; see Merging by naming above.
A merge is permanent in the sense that matters: it is remembered as your decision, not a guess, so clustering at a different threshold will not pull the two apart again.
Click the thumbnails of the people you want; selected cards get a blue border. The bar at the bottom of the page then offers Show in Lightroom. With several people selected, the dropdown decides what you get:
- photos with all selected people — only photos where everyone appears together (the default)
- photos with any selected person — the union
Lightroom builds the collection and switches the Library to it. This is the one thing the page cannot do on its own, so it needs the People progress bar still running in Lightroom — if it is not, the page says so instead of leaving you waiting.
Library → Plug-in Extras → Find Similar Faces...
- Select a single photo in the Library grid that contains the face you want to search for.
- Open Find Similar Faces from the menu.
- The dialog lists every face detected in that photo, with a thumbnail and the person's name where one is known. Pick the one to search for.
- The plugin queries the backend for photos with matching face embeddings across all indexed photos.
- Results are placed into a new Lightroom collection, sorted by similarity.
There are no scope or result-limit options here — the search always runs over the whole indexed catalog.
- Better names = better workflow. Naming persons early makes it easy to find all photos of a specific subject across your entire catalog.
- Run indexing with face detection on all portrait-heavy shoots before using the People workflow.
- Clustering works by visual similarity, so identical twins or people who look very similar can land in the same cluster. Open the person (double-click) to see which faces are actually wrong, move them out, and merge cards that are the same person by dragging one onto the other. Fixing it by hand is worth more than re-running: your corrections are kept and make every later run start from firmer ground.
- The fastest way to judge a threshold is not to guess it but to look. Double-click your biggest person, sort by outliers first, and see whether the faces at the far end are still the same human. If they are, the threshold can go up; if strangers appear, it should come down.
- For culling portrait sessions, face data also feeds into the Cull Photos scoring (eye openness, blink detection, face sharpness). See Help: Cull Photos.
The face pipeline used to run InsightFace's buffalo_l models, which could not
be shipped with the plugin — you had to install them yourself. It now runs YuNet
and FaceNet, which arrive with the ordinary Download AI models button.
The two produce face embeddings that cannot be compared with each other, so faces detected by the old models are not silently reused:
- They are left unassigned by Cluster faces rather than folded into clusters, which would produce confident nonsense.
- Photos holding them are reported as needing processing again, so a normal Analyze & Index Photos run with Enable face detection on re-detects them.
- Nothing is deleted, and names you have given people are kept. As each photo is re-detected, its faces reclaim the person they were assigned to.
So: download the models, then re-index the photos with people in them. Until you do, the People page will show those faces as unassigned.