Help: Cull Photos (beta)

Beta feature. Culling results are useful as a starting point but should be treated as suggestions, not final decisions. Always review Picks and Reject Candidates before deleting anything. No photos are ever deleted automatically.

What the culling workflow does

The Cull Photos workflow groups similar photos (bursts and near-duplicates), ranks them using technical and face-aware metrics, and creates Lightroom collections so you can quickly review:

Picks – the best usable photo per group (a group where nothing is usable, such as a single blurred frame, has no pick)
Alternates – reasonable alternatives you might still keep
Reject Candidates – clearly weaker shots
optional Duplicates / Near Duplicates
Brackets / Stacks / Panoramas (keep all) – see below

No photos are deleted automatically. All results are non-destructive and shown via collections.

Sets that must not be culled

An exposure bracket, a focus stack and a panorama all look like a burst to a grouper: same framing, seconds apart, near-identical. But there is no keeper to pick — every frame is part of one picture, and rejecting four of five destroys the shot. The backend detects these sets and switches ranking off for them: no winner, no reject candidates, and the whole group goes to the Brackets / Stacks / Panoramas (keep all) collection instead.

Detection is deliberately cautious and needs corroborating evidence from more than one signal, so it can miss a set rather than misread an ordinary burst. Exposure-bracket detection reads the exposure compensation value from your photos, so it works best on frames shot with AEB.

Prerequisites

The backend server is running and reachable from Lightroom.
The photos have been processed with Analyze & Index Photos so the backend has embeddings and culling metrics for them.

How to run culling

In Lightroom Classic, select the photos you want to cull or switch to a filtered view (for example a single shoot or folder).
Open the menu:
Library → Plug-in Extras → Cull Similar Photos....
Choose:
- Apply to – Selected photos only or Current view.
- Burst time window (seconds) – how far apart two frames may be and still count as the same burst. Auto (from preset) uses the preset’s own window: 1 s for Default, Portrait and Street, 2 s for Event, 3 s for Sports.
- Culling preset – tunes thresholds and weights: Default (sharpness, exposure, faces), Portrait (faces and expression), Street (candid moments), Event (people, emotion, moments), Sports (action and emotion). Every preset except Default also judges the moment; see below.
- Cull strength – how many near-misses become reject candidates. Gentle keeps more alternates, Strict aims for one keeper per burst, Balanced is the preset’s own margin. It only moves the score margin: a blurred frame, a blink or a blocked face is rejected at the same threshold whatever you choose.
- Create ‘Duplicates / Near Duplicates’ collection – on by default.
- Lightroom flags and ratings – see the next section.
Start the task and wait until the progress dialog completes.

Flags, stars and labels in Lightroom

Collections are good for review; flags are what you cull with. The Lightroom flags and ratings box writes the result the way you would by hand, so the Library filter bar, the flag shortcuts and Photo → Delete Rejected Photos all work on it directly:

Flag picks (P) – the best usable frame of every group gets the Pick flag. On by default.
Flag reject candidates as rejected (X) – reject candidates get the Reject flag. Off by default, because that flag is one menu command away from deleting files; turn it on once you trust the preset on your work. Nothing is ever deleted by the plugin.
Star ratings – picks 3 stars, alternates 2 stars. Reject candidates get no stars; the flag is their mark. Off by default.
Purple label on “look before deleting” rejects – a reject candidate that had its group’s strongest moment but a technical flaw is labelled purple, so it stands out in the grid before you delete rejects. On by default.
Also overwrite flags, stars and labels I set myself – off by default. The plugin remembers what each cull wrote on a photo (the Culling wrote field in the Metadata panel), so it can tell its own marks from yours. Marks an earlier cull wrote are always brought up to date with the new run — an earlier pick that is now an alternate loses its flag, an earlier “look before deleting” label goes — while a flag, rating or label you set yourself is left alone and counted in the completion dialog. Turn this on only to re-cull from scratch over your own marks.

Everything the run writes is one history step, so Cmd/Ctrl+Z takes it all back. Frames of an exposure bracket, focus stack or panorama are never flagged or rated.

The result collections are filled in review order: Picks and Alternates best first, Reject Candidates worst first. Set the collection’s sort to Custom Order to see it.

The plugin calls the backend culling endpoint, which:

groups photos into similarity clusters (single, burst, near_duplicate, and the keep-all types bracket, focus_stack, panorama)
scores each image per group
selects winners, alternates, and reject candidates — except in keep-all groups

Emotion and the moment

Sharpness alone picks the wrong frame from a burst surprisingly often: the sharpest frame is the one before the goal, the kiss with the eyes shut looks like a blink, the laugh is a little soft. So every preset except Default also asks what the photo shows:

Sports asks about action (is the ball at the foot, is someone in the air) and about emotion (celebrating, protesting, despair).
Event asks about candid moments and about emotion (laughing, crying with joy, hugging, guests reacting).
Portrait asks about expression, Street about candid, unposed moments.

The stronger of the two answers counts, not their average: the goal celebration has no ball in it, and the shot on goal has nobody cheering yet.

The decision then works the way a photographer culls: first usable, then the moment. Among the frames that are sharp enough, have the eyes open and the face clear, the strongest moment wins, even over a slightly sharper frame where nothing happens. A frame with a flaw no edit can repair (blur, unintended closed eyes, a blocked face) never beats a usable one, however strong its moment. Exposure is not on that list, because Lightroom lifts a stop without a trace.

Emotion beats the blink. At Event and Portrait, closed eyes in a kiss or a burst of laughter are recognised and not treated as a blink.

Singles and whole bursts can be rejected. A single photo used to be a pick whatever its quality. Now a clearly blurred single is a reject candidate, and so is every frame of a burst that is blurred throughout. Exposure alone never rejects a single, since a dark concert photo is often meant to be dark.

A blurred strong moment is never deleted unseen. When the frame with the strongest moment of its group is a reject candidate for a technical flaw, and its moment is clearly stronger than the pick’s, its explanation says so (“strongest moment in the group, but it has a technical flaw — look before deleting”), and the completion message counts these photos.

The moment is read from the image analysis (the SigLIP2 embedding). When you cull with a moment preset, the preparation step therefore computes it for any photo that does not have it yet. That takes longer than the culling signals alone, but no AI descriptions are generated. If the moment could not be judged, for some photos or for all of them, the plugin tells you why and what to do.

The emotion questions and weights are a first version and are still being measured against hand-culled shoots. If a preset keeps picking the wrong frame, export a culling fixture (Library → Plug-in Extras → Export Culling Fixture...) from a shoot you culled by hand and share it.

Result collections in Lightroom

For each culling run, the plugin creates a new collection set:

Name: Culling Results @ <timestamp>
Contents:
- Picks
- Alternates
- Reject Candidates
- optional Duplicates / Near Duplicates
- Brackets / Stacks / Panoramas (keep all) – only when the run found such a set

After creation, Lightroom automatically switches to the Picks collection inside that set so you can start your review immediately.

You can safely rename or move these collections later; they are standard Lightroom collections.

When the run finishes

The completion dialog counts the groups, picks, alternates, reject candidates and near-duplicate groups, and names the preset used. When flags, stars or labels were written it says how many of each, and how many were left alone because you had set them yourself. Two more lines appear only when they apply:

how many groups were recognised as brackets, stacks or panoramas, and that every frame in them was kept
Not culled: how many of the selected photos are missing from the collections, and why — the plugin could not compute a photo ID for them, or the backend grouped photos that are no longer in this catalog

If the backend reports a problem with the run itself, a Culling warning dialog shows it before the collections are created.

Understanding scores and explanations

For each photo, the backend stores culling-related fields such as:

group information: cull_group_id, cull_group_type, cull_group_rank, cull_group_winner
technical scores: cull_sharpness, cull_sharpness_peak, cull_exposure, cull_highlight_clip, cull_shadow_clip, cull_noise, cull_noise_flat (noise measured only where the frame is smooth, so fine detail no longer reads as grain; ranking prefers it to cull_noise, and photos indexed before it existed rank on cull_noise until re-indexed), cull_motion_anisotropy (camera shake vs. subject motion), cull_sharp_region_x / cull_sharp_region_y (where in the frame the sharpness actually sits, so a sharp background does not win over a sharp subject)
face scores: cull_face_score, cull_face_count, cull_face_sharpness, cull_face_prominence, cull_face_visibility, cull_eye_openness, cull_blink_penalty, cull_occlusion
aesthetic scores: cull_aesthetic (a contrast-and-colour heuristic) and the photo’s Aesthetic score from the AestheticSigLIP model, stored at indexing time (aesthetic_score) and read by culling as cull_aesthetic_model
moment scores, also computed at cull time: cull_semantic_iqa (action, expression or candid moment, depending on the preset), cull_emotion_iqa (Sports and Event) and cull_eyes_closed_intent_iqa (kiss or laugh, Event and Portrait)
explanations: cull_reason_codes, cull_explanation

The plugin writes only the result of culling into the photo’s metadata fields: Culling decision, Culling score, Culling rank, Culling group ID and Culling explanation. The sub-scores above (sharpness, exposure, noise, the face scores, the reason codes and the group type and winner) stay in the backend: they were never read back in Lightroom, and writing twenty fields per photo made every cull run touch every photo’s last-edit time. The explanation carries the reason codes in words. Photos culled before this change keep the values their old fields held, but those fields are no longer shown or updated.

Typical reason codes include:

sharpest_in_group
blurred
underexposed / overexposed
best_face_quality / weak_face_quality
eyes_open_best
possible_blink
possible_occlusion
no_face_detected_in_group
near_duplicate_weaker
peak_action / best_expression / strongest_moment / strongest_emotion – the moment decided the pick
eyes_closed_intentional – eyes closed in a kiss or a laugh, not counted as a blink
strong_moment_but_unusable – the group’s strongest moment, but a technical flaw put it among the reject candidates; look before deleting
bracket_frame_kept / focus_stack_frame_kept / panorama_frame_kept

These help explain why a specific frame was chosen as a pick or flagged as a reject candidate.

“Some culling signals are missing”

When the selected photos have no culling data yet, the plugin offers to prepare them — a fast pass that computes only what culling reads. It never runs a language model, so when the backend is on this machine it hands over a group of photos at a time and they are read and measured across several CPU cores rather than one after another. If a signal could not be computed, the plugin says so before the culling run starts instead of quietly grading photos without it. The usual cause is that the on-device models are not downloaded yet: open File → Plug-in Manager → LrGeniusAI and press Download AI models, then run culling again. Culling still works in the meantime, but face-aware ranking (eyes open, sharpness, occlusion) is inactive, and with a moment preset the moment is not judged either.

The same applies after the run: when the moment could not be judged for some photos because they have no image analysis yet, the culling warning names how many and what to do. Photos without it are ranked on sharpness, exposure and faces only, and so is every photo in the same group, so that the frames of one burst are always compared on the same terms.

The moment and emotion questions and the kiss-or-blink check need the AI search model once, to prepare a set of reference phrases; those are then kept in LrGeniusAI’s database, so later culling runs never load the model for them. If the model was not there for that first step, one culling warning names every score that could not be computed, says that ranking ran on the technical measurements alone, and how to fix it.

The aesthetic score is separate: every photo gets one when it is indexed (or prepared for culling) by the Photo quality scoring model, a second download in the same Download AI models button, and culling reads it from there. It is one input among several (sharpness, exposure and faces weigh more) and it only orders frames within a group, so a photo that scores low is not rejected for that alone. Photos indexed before it existed, or scored by an older version of the model, have none; the culling warning then says how many and that they were ranked on a simpler contrast-and-colour estimate instead, which can compare unevenly with scored frames in the same group. Run Analyze & Index Photos on them to fill it in.

Capture time and grouping

Burst grouping runs on each photo’s capture time. The fast preparation pass sends it from the catalog, and the backend reads it from the file’s EXIF when the catalog has none. Photos prepared by an older version without it group badly (whole sequences of similar frames collapse into one group with one pick), so the plugin offers to prepare them again once; a photo that has no capture time anywhere is remembered as such and not asked about again.

Rejects for a blink, a blocked face or a weak face are judged against the other frames of the group as well as against fixed thresholds: a frame is only a blink when its eyes read clearly less open than the best frame of its burst. On small faces the eye estimate reads “closed” on every frame, so without this every frame but the pick was a reject. The same treatment for blur exists in the backend but is switched off until it has been checked against hand-culled shoots of several genres.

Tips for best results

Run culling after you have narrowed down an initial selection for a shoot (for example by folder or date range).
Use Analyze & Index Photos with face detection enabled if you want face-aware ranking (eyes open, sharpness, occlusion).
Start with conservative presets (default) and treat the result as a review aid, not an automatic delete list.