People and face recognition

How faces are grouped and named, what happens when PDR gets one wrong, and exactly where the face data lives.

Applies to: Photo Date Rescue for Windows, version 3.0.4. Last verified: 30 July 2026. Every statement on this page was checked against the shipping 3.0.4 build. If you are on an older or newer version, the behavior may differ — check What’s new for changes.

PDR can group the faces in your library, let you put names to them, and then find the rest of the photos that person appears in. This page describes exactly what that involves, what it costs you in time and disk space, and — because it is the question people ask first — where the face processing actually happens.

Face recognition is off when you install PDR. Nothing in your library is scanned for faces until you turn it on yourself.

At a glance

Capability	In 3.0.4
Runs face detection and face matching on your own computer, with no upload of photos or face data	Yes
Leaves face recognition switched off until you choose it	Yes
Groups similar faces and lets you name the group	Yes
Lets you verify, unname and reassign individual faces	Yes
Uses the faces you verified to find more photos of the same person	Yes
Renames people and merges duplicates	Yes
Removes a person and lets you get them back	Yes, with a time limit on one of the two undos
Imports face names from another program	No
Names a face PDR did not detect	No

Where the face processing happens

On your computer. The models that find faces in a photo and turn each face into a numeric fingerprint are installed as part of PDR. They are not fetched from a server, and they do not need one to run. When PDR looks for faces it reads the image file from your disk, runs the models on your processor in a background task, and writes the results into PDR’s own library database on the same machine.

No photo, no cropped face and no face fingerprint is sent anywhere. There is no account to sign in to for this feature, and the results are not synchronized between two computers — if you install PDR on a second PC, its People list starts empty.

Alongside face recognition, PDR’s optional AI can also tag what is in a photo. That part is separate, and it is the only part that involves a download: the first time you enable the AI features PDR fetches about 300 MB of model files, after which the tagging also runs entirely on your machine. Subject tagging chooses from a fixed list of 193 terms built into PDR — it does not generate free-form descriptions of your photos.

Limitations

Face data lives in the library database, not in your photos. Naming a person does not write that name into the image file. Your photos are not modified when you use People at all. Nothing you do here is visible to another photo program.
The face crops are cached on disk. To keep the People window responsive, PDR saves small cropped images of each detected face in its own data folder so it does not have to re-open the original every time. They stay there until you clear PDR’s data — see where PDR keeps its data.
It uses your processor, not a graphics card. Face work runs on the CPU. On a large library it takes a long time and competes with a Fix that is running at the same time.
Turning the AI features on turns on more than faces. The same master switch covers subject tagging. If you want faces and nothing else, turn the tagging off separately.

Turning face recognition on

PDR ships with the AI features disabled. Until you switch them on, no face is detected and the People list stays empty. There are two switches, and they are in different places.

To do it

Open Settings.
Go to After Fix and turn on Recognize people and content. This is the master switch for all of PDR’s AI, and the first time you use it PDR downloads the model files it needs for subject tagging.
Go to People and check that Face Detection is on. It is on by default, but it does nothing until the master switch above is on too.
Leave PDR running. Faces are found in the background, a photo at a time, and the People window fills up as it goes.

Two other switches sit nearby. Visual face suggestions offers visually similar people from your library when you are typing a name. In the AI section, Auto process decides whether newly added photos are analyzed automatically or only when you ask.

Limitations

Photos have to be in the library first. Face recognition works on photos PDR has indexed. If you have turned off “Make Fixed photos searchable”, newly fixed photos are not indexed and so are not scanned for faces.
There is no progress bar for the whole library. You get a running count of what has been processed, not an estimate of when it will finish.
It is slow, and honestly so. PDR’s own estimate is around a minute per hundred photos. A library of a hundred thousand photos is a multi-day job on a typical PC.
Detection quality is not something we promise. Faces in profile, small faces in group shots, heavy blur and very old scans are frequently missed. There is no way to correct a miss — see known gaps.

Grouping faces and naming a person

As faces are found, PDR groups the ones that look like the same person into a cluster. A cluster is a guess, not a person — it becomes a person when you give it a name. You do that in People Manager, which opens in its own window from the sidebar.

To do it

Open People Manager. The counts across the top tell you how many people are named and how many face groups are still unnamed, unsure or ignored.
Find an unnamed row you recognize and enter a name. You can also give a full name separately from the short name you will see on the tiles.
The row moves into the named section, and every face in it is attached to that person.
Use the Match slider at the top to change how tightly faces are grouped. Sliding towards Loose puts more auto-matched faces into each row; sliding towards Strict keeps only the most similar ones. Moving it regroups the unnamed faces.

Limitations

A cluster is not guaranteed to be one person. Two people who look alike can land in the same row, and one person can be split across several rows. That is why the naming step is yours and not PDR’s.
The Match slider only regroups faces that are not named yet. It does not revisit people you have already named.
Only faces PDR detected can be named. If PDR did not find a face in a photo, there is nothing to name.
The name is stored in PDR only. It is not written into the photo, and it does not appear in Windows, in File Explorer, or in any other photo application.

Verifying, unnaming and reassigning faces

Within a named person, individual faces are either verified by you or matched automatically. The distinction matters: only the faces you have verified are used as the reference when PDR goes looking for more photos of that person, so a wrong auto-match cannot spread.

To do it

Open a person’s row in People Manager and look through their faces. There is a filter for showing only the faces that are not verified yet.
Confirm the ones that are right. You can verify a whole person’s faces in one action rather than clicking through them.
If a face is not that person, either send it back to Unnamed or reassign it directly to the person it belongs to.
If you unname faces, a bar appears offering Undo. It is available for 20 seconds.

Limitations

The undo on unnaming faces expires. After 20 seconds the bar is gone and there is no second chance. Restoring the database from a backup is then the only route back — see backing up your library.
Verification is manual and there is a lot of it. Nothing verifies itself. On a large library this is the single most time-consuming thing in PDR.
Reassigning a face does not tell PDR why it was wrong. The correction is recorded, but the model is not retrained on your machine.
Photos are never touched. Unnaming or reassigning a face changes PDR’s records only. The image files on your drives are not modified or moved.

Finding more photos of someone you have named

Improve Facial Recognition takes the faces you have verified and searches the unnamed faces for more of the same person. You can run it for one person or for everyone you have named. It reports how many new faces it matched, and you can run it again whenever you have verified some more.

To do it

Verify at least one face for the person — the tool needs a confirmed example to work from.
Set the Match slider to how strict you want the matching to be.
Click Improve Facial Recognition for everyone, or use the per-person version on a single row.
Watch the progress and check the new matches it added.

PDR takes a snapshot of its library database before each run, so an Improve that goes badly can be rolled back from the Backup section of Settings.

Limitations

Faces it matched automatically are never used as examples. Only faces you verified count, deliberately, so a bad match cannot compound into worse ones.
It competes with a Fix for your processor. Running both at once makes both slower, and PDR warns you rather than blocking it.
New matches arrive unverified. They are suggestions. You still have to look through them.
A loose setting will produce wrong matches. There is no setting that is right for every library; a face-heavy family library and a library of crowd shots need different ones.
There is no undo button for a run. The recovery route is the database snapshot taken beforehand, not a one-click reversal.

Renaming people and merging duplicates

You can rename a person, set their full name, choose which face is used as their picture, and merge two people who turn out to be the same person.

To do it

Use the actions on a person to rename them or edit their full name.
To change the face shown for them, pick a different face from their photos and set it as the one to use.
To combine duplicates, merge one person into the other. Every face from the person you merged away is moved across.

Renaming a person to a name that already exists merges them. If you rename “Mum” to “Sarah” and there is already a Sarah, you end up with one Sarah holding both sets of faces. Capitalization is not treated as a difference.

Limitations

A merge is not reversible from the People window. Once two people are one, there is no un-merge; you would have to reassign the faces by hand or restore a database backup.
An accidental merge is easy to cause by renaming. Because a matching name merges automatically, a typo that happens to equal an existing name will combine two people without a separate confirmation of the merge.
Merging also updates anywhere else that person is referenced, including the family tree, so the change is not confined to the People window.

Removing a person and getting them back

There are two different ways to get rid of a person, and they recover differently. Neither of them deletes, moves or edits a single photo.

To do it

Discard the person. They leave the People list and go to a discarded list. Their faces stay attached to them, so restoring them later brings the person back intact. There is no time limit on this.
Send them back to Unnamed. The person record goes and their faces return to the unnamed pool, where you can name or reassign them again. An Undo appears for 20 seconds; taking it restores the person with every face and its previous verified status.
To restore a discarded person, open the discarded list and restore them.
To remove someone permanently, discard them first, then delete them from the discarded list.

Limitations

Permanent deletion cannot be undone. It is only offered for people you have already discarded, which is the confirmation step, but once done the only way back is a database backup.
The 20-second Undo is the whole window. If you close the message, switch away, or simply take too long, it is gone. PDR points you at the Backup section of Settings after that, which rolls the whole library back to an earlier point — not just that one person.
Restoring an unnamed person creates a new record. The person and their faces come back with their verified status, but this is a fresh entry rather than the original one, so anything that pointed at the old record elsewhere may not follow.
Discarding does not free up disk space. The detected faces and their cached crops remain in PDR’s data folder.

Known gaps in 3.0.4

The honest list, in one place:

You cannot name a face PDR did not find. There is no way to draw a box around a missed face and label it. If the detector skipped someone, that photo will not show up under their name.
There is no import of face names from other software. If you have already tagged people in another photo program, none of that work carries over — 3.0.4 has no control anywhere for reading in face names or face regions, and you start from nothing.
No un-merge. Merging two people is one-way from the interface.
Face results do not travel between computers. There is no sync. Two machines mean doing the naming twice, or restoring a library backup onto the second one.
The undo on unnaming lasts 20 seconds. Every other safety net for People is the library backup, not a per-action history.

How to check this page is still true

Every claim here describes behavior you can observe in the shipping build. If something on this page does not match what PDR does on your machine, that is a defect in one of the two, and we want to know: contact support. Tell us the anchor — for example /help/people#summary — and we will either correct the software or correct the page.