Why does Lightroom Classic face recognition stop working — and what can you do about it?

By David · April 19, 2026 · 8 min read

You probably noticed it the same way I did: Lightroom Classic face recognition works great on the first few hundred photos, then starts misfiring. It suggests the wrong name with alarming confidence. The "unnamed people" pile grows faster than you can clear it. At some point you just stop using People View entirely.

It's not you. It's not your library organization. The underlying engine has a fundamental architectural problem — and understanding it is the first step to actually fixing your workflow.

How does Lightroom Classic face recognition actually work?

When you open People View, Lightroom runs a face detection pass across your entire library in the background. It finds faces, crops them, and groups visually similar crops into clusters. You then name those clusters, and Lightroom uses that as a signal to surface new suggestions.

The key word is clusters. Lightroom doesn't build a per-person face model the way a modern neural network would. It groups faces that look similar to each other and asks you to confirm the grouping. The "suggestions" you see when tagging are based on which existing named cluster a new face most resembles — not a robust identity score.

This distinction matters more than it sounds. Cluster-based matching works fine on a small, consistent dataset. Add a few thousand photos — different lighting, angles, ages, varying image quality — and the cluster boundaries start overlapping. Lightroom can't reliably distinguish between siblings, family resemblance, or just two people with similar haircuts under studio light.

Why does accuracy get worse the more you use it?

Here's the counterintuitive part: the more you use People View, the worse it gets. Every false positive you confirm teaches the cluster that those two faces belong together. Over time, a person's cluster accumulates noise from incorrectly tagged lookalikes, and future suggestions inherit that contamination.

There's also a scale problem. Lightroom's background face detection runs as a low-priority task, and on libraries over 20,000 photos it can thrash — repeatedly re-scanning folders it's already processed, surfacing photos you tagged years ago as "unnamed," and generally behaving like it lost track of what it's already done.

And then there's the neglect factor. Adobe shipped People View in Lightroom Classic 6 (2015). The face detection model hasn't had a meaningful public update since. The rest of the AI industry moved on to deep metric learning; Lightroom Classic didn't follow.

What do modern AI face-recognition tools do differently?

The difference comes down to how a person's identity is represented in memory.

Lightroom's cluster approach is essentially: "this face looks like the other faces in bucket #7, so it's probably person X." Modern tools like InsightFace (ArcFace model) instead compute a 512-dimensional embedding vector for every face — a precise numerical fingerprint that captures identity independent of lighting, angle, and expression. Two photos of the same person in dramatically different conditions will produce vectors that are very close in L2 distance. Two different people, even siblings, will be measurably farther apart.

When I built the Face Tagger plugin, I used a specific technique called centroid matching: rather than storing every individual face encoding for a person, the plugin computes one mean (L2-normalized) embedding per person across all their training photos. Recognition then means computing the distance from a new face to each person's centroid — one comparison per known person, fast and stable even at scale.

The other signal Lightroom ignores entirely: temporal context. In my testing, photos taken within 5 minutes of a confident face match are substantially more likely to feature the same people. The Face Tagger plugin applies a small distance boost — up to 0.06 — for faces that appear in temporal proximity to confirmed matches. It's a small signal, but it measurably reduces false negatives in sequence shots and event photography.

Quick fixes to try before replacing the feature

If you're not ready to switch tools, a few practices will wring more accuracy out of People View:

• Don't let background detection run unattended. Turn off automatic face detection in Catalog Settings and run it manually on specific folders when you're ready to tag.
• Reject bad suggestions aggressively. Every time you confirm a wrong face, you pollute the cluster. It's better to leave a face unnamed than to confirm it incorrectly.
• Use "Find Faces Again" on individual photos (right-click a photo in People View) to force a re-detection pass on images where the initial scan missed faces.
• Tag in batches by shoot. Lightroom's suggestions improve when you work within a single event's photos — the visual consistency helps the cluster matching.
• Keep your training set clean. Remove confirmed faces that are blurry, small, or partially obscured. One bad training example can skew an entire cluster.

These are workarounds, not solutions. If your library is over 10,000 photos or you're tagging more than 10–15 people, you'll hit the ceiling quickly.

Native People View vs. AI plugins — honest comparison

The honest version: if your library is under 5,000 photos and you're tagging 3–4 people, People View is probably fine. Past that threshold, the degradation compounds faster than the convenience is worth.

Building a face-tagging workflow that actually scales

After a lot of trial and error building the Face Tagger plugin, the workflow that holds up at scale looks like this:

1. Train first, scan later. Select 10–20 clear, varied photos of each person and run Train People. The plugin builds per-person centroids from those examples before touching the rest of your library.
2. Scan in batches. Run Scan Selected Photos on a shoot or a date range. The plugin pre-filters with YOLO person detection — skipping photos with no person visible entirely — so the ArcFace scan only runs where it's needed.
3. Review unrecognized faces after each scan. The plugin automatically clusters unknown faces and presents them for naming. You're not wading through an infinite unnamed pile — you're reviewing tight visual clusters of 3–15 similar faces at a time.
4. Sync corrections to auto-tune. After you manually fix any tagging mistakes in Lightroom, run Sync Corrections. The plugin detects confirmations and rejections, then adjusts per-person recognition tolerance automatically — stricter for people who get over-matched, looser for people who get missed.

This is the approach that scales to 50,000 photos without falling apart. The correction loop is the part most face-tagging tools skip entirely — and it's also the part that compounds the most value over time.