BUG: Face clusters show fewer photos than face search for the same person

[varsharaodevaraj]

Is there an existing issue for this?

• I have searched the existing issues

What happened?

What happened?

After AI tagging completes, Face Collections (People view) show very few photos for a person,
while face search returns many more photos for the same person.

Example from my local setup:

• AI Tagging → Face Collections → Person 5eb9 shows 3 photos
• Home → Face search using the same person returns 13 photos

This is confusing from a user perspective since both features appear to answer the same question:
"Show me all photos of this person."

---

How to reproduce

1. Add a folder with many photos of the same person
2. Enable AI tagging and wait for it to complete
3. Go to AI Tagging → Face Collections
4. Click a person cluster and note the photo count
5. Use Home search with a photo of the same person
6. Observe that search returns significantly more photos than the cluster view

---

Expected behaviour

Face Collections and face search should be consistent, or the UI should clearly explain
why results differ.

---

Actual behaviour

Face Collections show only a small subset of photos, while face search returns many more
for the same person.

---

Technical analysis (initial one)

From tracing the backend pipeline:

• Face clustering uses DBSCAN with stricter similarity thresholds and may exclude faces as noise

• Face search compares raw embeddings with a looser threshold and ignores clusters entirely

• Some faces remain unclustered but still appear in search results

• I agree to follow this project's Code of Conduct

[github-actions]

👋 Thanks for opening this issue! The maintainers will review it shortly. Till then, do not open any PR. This is a strict rule we like to follow around here :)

[varsharaodevaraj]

@rahulharpal1603 could u assign this issue to me

[Arunodoy18]

Hi @varsharaodevaraj,

I’d like to investigate this inconsistency.

From the description, it seems clustering (DBSCAN-based) and face search use different thresholds and possibly different embedding comparison logic.

My plan:

• Trace both clustering and search pipelines in backend
• Compare similarity thresholds used
• Identify whether discrepancy is due to DBSCAN noise filtering or threshold mismatch
• Propose either:
  • Aligning thresholds
  • Including unclustered faces in People view
  • Or clarifying behavior in UI

I’ll share findings before proposing a fix.

Let me know if I can proceed.