[Feature] Facial Recognition: "Re-cluster" job that preserves manually assigned names

[Jero1ste]

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

• Yes

The feature

Related discussions:

• [Feature] Facial Recognition Enhancements #2472 (closed) — several users requested similar functionality,
  including geoleven and Germs2004
• [Feature] Face Detection - unmerge / split faces and re-evaluate faces #3356 (closed) — unmerge / split faces and re-evaluate

---

The problem

When adjusting facial recognition settings (e.g. lowering the
maximum recognition distance to better separate similar-looking
people), there is currently no way to re-cluster only the
automatically assigned faces without losing all manually assigned
names.

The two current options are:

Job	Behavior
Missing	Only processes unassigned faces — does not re-cluster existing ones
Reset	Re-clusters everything — destroys all manually assigned names

There is no middle ground.

---

Real-world scenario

I am building a library of 200,000+ photos spanning 25 years.
The library includes two brothers with an age difference of
approximately 1.5 years. At young ages, their facial embeddings
are very similar, causing the model to frequently cluster them
as the same person.

After manually correcting hundreds of misassignments and lowering
the maximum recognition distance from 0.45 to 0.40, I would like
to re-run clustering on the automatically assigned faces only,
so the new setting takes effect — without losing the manual
corrections I have already made.

A full Reset would destroy weeks of manual work. The Missing job
does not help because the incorrectly clustered faces are already
assigned (just wrongly).

---

Proposed solution

Add a "Re-cluster automatically assigned faces" job (or a
"Refresh" button, similar to the one on Face Detection) that:

1. Re-runs the clustering algorithm using the current settings
2. Skips any face that has been manually assigned
   (i.e. where manually_assigned = true or equivalent flag in
   the database)
3. Only re-evaluates faces that were assigned automatically by
   the model

This would allow users to:

• Adjust recognition settings and apply them without data loss
• Correct systematic errors (e.g. similar-looking siblings)
  without a full reset
• Iteratively improve clustering quality over time

---

Technical note

Immich already stores manually assigned names separately in the
database. The distinction between manually and automatically
assigned faces therefore appears to already exist at the data
level, making this feature potentially feasible without major
architectural changes.

Platform

• Server
• Web
• Mobile

[bo0tzz]

#12931

[Jero1ste]

"Thanks for pointing to #12931 @bo0tzz! I see the strong overlap: both tackle the same core problem of separating AI guesses from human verification. Where #12931 focuses on the UI/UX confirmation workflow, my proposal addresses the specific backend job ('Re-cluster unconfirmed/auto-assigned only') needed to safely tune clustering thresholds on large libraries without data loss. I'll add my technical context and use case to #12931."