[Feature] Merge Faces from Metadata with detected faces

[hjed]

I have searched the existing feature requests to make sure this is not a duplicate request.

• Yes

The feature

In #4348 immich implemented reading face metadata from files.

These faces show up in the people view, etc but they are often duplicated with immich's own detected faces. When there is a significant overlap with an ML detected face immich should merge the detected faces with the one provided by metadata (and ideally their linked groups too).

My observation is that most detected face bounding boxes detected by digikam very closely match the size and shape of the bounding boxes generated by immich for the same faces so this would be very useful for portability.

Platform

• Server
• Web
• Mobile

[tye-mustafa]

my guess the tags come from xml are bad i tried with Lightroom and exported it to immich ended up same,

[hjed]

The ones from digikam work great except they don't merge with the autodetected ones which is really annoying because it means to use immich's better algorithm I'd have to go through a retag all faces (and having gone through this with several over programs and since immich doesn't support exporting faces its a big time sink that I'll have to do again next time I change programs w/ the associated data loss, etc with that)

[5TDg9]

Is this really needed? This is probably going to get a little bit messy... i would be statisfied, if the face tags generated by immich would be just somehow "exportable" in a format, that other software (Digikam, XnView, etc.) is able to read or extract.
Instead of merging two face tag datas - one from immich and one from another software (e. g. Digikam), they could surely co-exist both in the XML-File with an special description-tag like mentioned here
(how mylio is doing it). This would be lot easier to implememt than the merging.

[hjed]

It means that every face has to be retagged in immich at the moment to use the immich features and is duplicated in the UI. The data should be portable.

[rhatguy]

I think the way forward for something like this is a comparison from a % of size perspective. For example if Immich detects a region that overlaps say 90% (guessing at an appropriate % here) with a region from metadata, it should assume those two faces are the same. Using a % should help with different size faces rather than an absolute like number of overlapping pixels.

It would be super nice if Immich had a function somewhere that could report these overlaps and "correct" them if the user confirmed. For example, metadata said this face was here....Immich detected it here (overlapping by 90%). If user confirms, update the pictures metadata to match Immich's face region.

[dnnr]

For what it's worth, digikam internally appears to do a very similar thing. Sometimes there can be multiple face regions (detected but unrecognized) on the same face, so that they overlap by a lot. If the user assigns a specific name to one of them, the other one disappears. Maybe it's worth checking their implementation for how they decide this?

[Finke3]

Totally agree. If Immich detects an existing face from meta data which overlaps it should assume those two faces are the same!

[Finke3]

I'm facing the same. My photos are mounted via external library and parts already tagged with faces by Lightroom. Now every photo contains two face regocnition boxes and will shown twice by searching via mobile app.

[Banewraith]

Using V2.3.1
I would like this as well. I spent a lot of time getting the names for faces found in old family photos. These assets are spread across several Immich instances. I wrote some api calls to extract the faces regions and names from the original Immich instance and burned them into a copy of the asset using exiftool. I then drop these exported version into the external folder of my Immich V2.3.1 instance.

After time each asset will show two faces. One with the name from metadata and the other recently detected. The bounding boxes look to be identical. When I name the version without a name, it will integrate with the correct person. No duplicated faces or regions after that.

So I can just manually name all these new detected faces since I have the names from the metadata and they shows up in the asset info screen. Thanks for adding the face import option. Not sure when it was added. I've been on v1.135 for a while

I'm working to update my api calls for v2.3 but, it would be great if there was a automagical way of combining faces based on bounding box cords.

In V1.133.x
They were found with an api call to "/api/assets/$AssetUUID"
The resulting json has a people array and that has a faces array that contains the bounding box info.

This snip does not show an example of two faces with the same bounding box but it is a start.

  "people": [
    {
      "id": "abf1205a-16a6-4bdd-b779-d58d40bfec33",
      "name": "xxx, Lyddie",
      "birthDate": null,
      "thumbnailPath": "upload/thumbs/fc8718a2-d975-40a9-8ec9-357641b1f49d/ab/f1/abf1205a-16a6-4bdd-b779-d58d40bfec33.jpeg",
      "isHidden": false,
      "isFavorite": false,
      "updatedAt": "2025-06-13T15:03:49.320112-05:00",
      "faces": [
        {
          "id": "2c89ecdc-548f-4922-86d0-f298ef6e8576",
          "imageHeight": 1440,
          "imageWidth": 2120,
          "boundingBoxX1": 1143,
          "boundingBoxX2": 1434,
          "boundingBoxY1": 326,
          "boundingBoxY2": 588,
          "sourceType": "machine-learning"
        }
      ]
    },
    {
      "id": "4bc1b036-30db-41ef-be85-05e946e3f242",
      "name": "xxx, Sally",
      "birthDate": null,
      "thumbnailPath": "upload/thumbs/fc8718a2-d975-40a9-8ec9-357641b1f49d/4b/c1/4bc1b036-30db-41ef-be85-05e946e3f242.jpeg",
      "isHidden": false,
      "isFavorite": false,
      "updatedAt": "2025-11-06T17:55:05.727535-06:00",
      "faces": [
        {
          "id": "de2fc5bf-a653-4cfc-8649-1ecbef515ddc",
          "imageHeight": 1440,
          "imageWidth": 2120,
          "boundingBoxX1": 271,
          "boundingBoxX2": 585,
          "boundingBoxY1": 70,
          "boundingBoxY2": 492,
          "sourceType": "machine-learning"
        }
      ]
    }
  ],
  "unassignedFaces": [],