[Feature] New face-recognition backend & more models

[barti04]

I have searched the existing feature requests, both open and closed, to make sure this is not a duplicate request.

• Yes

The feature

Face recognition is one of Immich’s “wow” features; when it’s solid, it turns a big photo dump into something you can actually browse.

The problem I keep running into (and I’ve seen others run into too) is that the current setup feels “locked in”:

• If face grouping is slightly off for your library (split people too much, merge similar faces, struggle with older photos), there’s not much you can do besides manual cleanup.
• Model support is basically tied to InsightFace’s ecosystem and whatever Immich packages.

DeepFace is attractive here because it’s essentially a single interface to multiple established face models. All the InsightFace-family models Immich users already know (e.g. Buffalo_L) are available via DeepFace, but DeepFace also makes it easier to support alternatives without Immich having to directly integrate each one.

DeepFace also appears to be more actively maintained, with recent releases and model benchmarks (Python versions, NumPy changes, etc.). It even seems to be moving toward a usage model that matches Immich’s needs quite well.

Platform

• Server
• Web
• Mobile

[bo0tzz]

Are there specific models you'd like to see, or is this just a general suggestion for us to switch to another framework?
I wasn't directly involved in the design & implementation of the current system, but I'm pretty certain deepface will have been considered and we'll have gone with the current approach for good reasons.

[barti04]

Thanks for the swift reply!

There definitely were good reasons for using InsightFace at the time. Immich’s facial recognition relies on vector embeddings and clustering, and DeepFace historically did not provide an out-of-the-box workflow for vector search and embedding persistence. With the recent DeepFace additions, some of this functionality is now available directly, which may help abstract parts of the complexity currently handled in Immich (e.g. embedding storage and similarity search):

https://sefiks.com/2026/01/01/introducing-brand-new-face-recognition-in-deepface/

Models Immich supports right now

Name	Detection model	Recognition model	Model size
antelopev2	SCRFD-10GF	ResNet100@Glint360K	407MB
buffalo_l	SCRFD-10GF	ResNet50@WebFace600K	326MB
buffalo_m	SCRFD-2.5GF	ResNet50@WebFace600K	313MB
buffalo_s	SCRFD-500MF	MBF@WebFace600K	159MB

In Immich it is currently not possible to choose the detection and recognition models separately; only bundled model packs can be selected. The models are a good baseline for accuracy and performance, but far from ideal. An example:

Detection model benchmarks

Model	Easy	Medium	Hard
SCRFD_500M	90.57	88.12	68.51
SCRFD_2.5G	93.78	92.16	77.87
SCRFD_10G	95.16	93.87	83.05
SCRFD_34G	96.06	94.92	85.29

For comparison, RetinaFace with R50 reports:

Model	Easy	Medium	Hard
RetinaFace-R50	96.5	95.6	90.4

RetinaFace is a DeepInsight model and can therefore be used within the InsightFace ecosystem. A better facial recognition model with similar accuracy benefits, such as Facenet512, is, however, not supported by InsightFace.

If moving to DeepFace is too complex, I would be happy to open a separate feature request focused on updated or more flexible model options within the current InsightFace-based implementation.

cheers ✌️

[bo0tzz]

which may help abstract parts of the complexity currently handled in Immich (e.g. embedding storage and similarity search):

These things are not handled anywhere near where deepface would go.

If there are specific models you'd like to see in Immich then feel free to open a new FR for those, but I don't see any benefits in moving to deepface.