What Is Photo Culling and How AI Does It

What culling is, and why wedding galleries make it hard

Culling means reducing a shoot to a delivery set by rejecting failures and choosing the best of each moment. A common difficulty with weddings is repetition: bursts of the same kiss, a group shot repeated until every face is open-eyed, and a second shooter's overlapping coverage. Frame counts differ by photographer and are not something this site can state. How to do the work by hand is on how to cull wedding photos; doing it inside Lightroom is on how to cull in Lightroom.

Culling normally has two jobs that need different judgement:

Technical rejection

missed focus, motion blur, blinks, misfires, severe exposure problems, test shots.

Selection

picking the strongest frame in a sequence, and checking that coverage is complete (family formals, key moments, details).

The vendors' pages describe mostly technical signals (sharpness, eyes open, exposure, duplicates), while which frame is the best moment depends on the couple, the brief and your style.

Culling versus editing, and two ways to cull

Culling is selection; editing is colour, tone and retouching, and it comes after. The words overlap in some usage: Narrative's guide notes that photojournalism sometimes calls selection "editing", which is why a client or editor may say "edit" when they mean "cull".

Within culling there are two directions. "Cull out" rejects failures and keeps the rest as maybes; "cull in" assumes nothing qualifies and marks only what you would deliver. Narrative says most experienced photographers recommend cull in; that is the vendor's statement. The practical steps are on how to cull wedding photos.

How many frames survive

Vendors publish keeper-rate and volume ranges, none with a method. Narrative's guide gives unsourced shot-to-delivered ranges by shoot type, and FilterPixel's guide gives its own unsourced estimates plus a self-reported single-shoot example. Treat them as prompts to measure your own galleries: delivered frames divided by shot frames, per shoot type, over your last several jobs. "Keeper rate" also lacks one definition; it can mean frames you deliver, frames you keep for edit, or frames a tool selects.

How AI culling works

AI culling software analyses each image and combines several signals into a keep, reject or rank decision that you can override. The vendors' own pages describe the signals; some describe working from previews (FilterPixel and Imagen describe uploading previews, per their pages).

Sharpness and focus

Adobe's Assisted Culling page lists subject sharpness and eye sharpness, each with a 0 to 100 slider. Imagen lists a Sharpness Check that analyses focus accuracy to find the sharpest photos in each set.

Eyes open

Adobe lists an Eyes open check, with an option to reject images where the eyes cannot be determined; Imagen lists blink detection that flags closed eyes; Narrative says its assessment shows whether subjects are in focus and whether eyes are open or closed.

Exposure and misfires

Adobe lists rejects for exposure issues, misfires and documents or receipts.

Grouping

Narrative groups similar moments into Scenes and ranks within each; Imagen says its AI groups similar shots and keeps the best, and can remove duplicates; FilterPixel describes groups of similar images and a genre-aware DeepCull mode that gives a score and a reason per photo; Excire Foto lists duplicate search and best-shot detection by sharpness, expression and aesthetics.

Learning your preferences

Aftershoot's homepage says Select picks up your preferences, patterns and creative instincts as you cull, and FilterPixel says custom tagging learns from your stars, labels and flags. How well any of that works is not shown by a published method.

How results come back differs: Adobe lets you apply a flag, star rating or colour label, and Aftershoot's homepage FAQ says RAW ratings are typically written to XMP sidecar files (it also lists one-click export to editing apps). Narrative says it never deletes a photo and FilterPixel says its Review Mode never removes an image; for other tools, whether anything is deleted: not found in sources reviewed.

Assisted, automated and the "first pass"

"Assisted" means the software proposes and you confirm; "automated" means it applies its decisions across the whole shoot and you review afterward. The terms are vendor language, not a standard.

• Adobe's Assisted Culling sorts photos into Selects and Rejects for you to review, lets you override rejects, and lets you organise results with a flag, star rating or colour label. Adobe's page says it is available in both Lightroom and Lightroom Classic and gives no release date or plan requirement; check Adobe's help documentation for the current menu path in your version.
• Aftershoot describes culling in automated and assisted modes, per its pricing guide.
• Narrative calls its ranking "AI First Pass".

The labels are not used consistently. Narrative's guide sorts tools into manual, AI-assisted (the AI analyses and you select) and fully automated (the AI selects without your input), and places Aftershoot, FilterPixel and Imagen in the last group, while Aftershoot's own pricing guide describes both automated and assisted modes. That is a rival's classification, so check each tool's actual review step instead of trusting the label.

Whatever the label, the useful question is who reviews the rejects and how long that review takes.

Local, cloud and hybrid

Where the analysis runs is a separate choice from how it works, and it matters for client files. Vendors describe three patterns: local (Excire says analysis runs on your computer; Narrative says it culls on your computer), cloud (Imagen's editing uploads smart previews), and hybrid (FilterPixel says originals stay on your computer while a small embedded preview goes to the cloud for DeepCull). Vendor statements, not verified here. The tool-by-tool picture, including conflicts, is in offline vs cloud culling.

Glossary: terms and why vendor accuracy figures are not comparable

Vendor accuracy figures describe different things measured different ways, so one cannot be set against another. The terms below are the ones needed to read them.

Several published figures show the problem. FilterPixel's comparison page (a competitor claim about Narrative) reports 94.7% keeper accuracy and 96.1% reject accuracy from "controlled testing with 2,740 wedding images" against Narrative, with no tester or method named; its page about Imagen gives Imagen 63.4%, undisclosed method, and it is a competitor claim. Imagen advertises culling time cut "by up to 75%", also without a method. FilterPixel's own guide says modern AI culling tools reach 90 to 95 percent agreement with professional selections on technical quality, a general statement with no dataset shown, and Narrative's guide repeats FilterPixel's 94.7 percent figure, attributing it to "independent testing" in early 2026 without naming a tester. Sources: FilterPixel on Narrative and on Imagen. A figure can only be compared if gallery, human reference, definition and target are shared, and none of these is. Weigh these as marketing until a method exists; this site's own method is on the methodology page and the scoring sheet is the culling test scorecard.

What AI culling may misjudge

The vendors' pages describe signals such as sharpness, eyes open, exposure and grouping. They do not claim to know which frame matters to a family, and that is the human part. When you trial a tool, include the cases where signals and intent diverge:

• Deliberate motion blur, silhouettes, backlight and flare, which can read as technical faults.
• Low-light reception frames with noise, where "sharp" depends on the scorer's tolerance.
• Glasses, profiles and veils, which can make an eyes-open check ambiguous (Adobe's page includes an option for images where eyes cannot be determined).
• Group frames with one person turned away, where a per-face score may differ from your judgement.
• A burst whose winner differs by a micro-expression.

These are not documented failures of any named tool; they are cases to put in your own test set.

Does AI culling replace you?

The documents describe AI culling as a first pass with human review. Adobe's page has you review rejects and override them, and Narrative says it never deletes photos. What fraction of your time it saves on your gallery is not documented independently; vendor time claims are unmeasured. The workflow question of where to hand off is covered on how to cull wedding photos.

Choosing where to go next

• Reading about specific tools: Aftershoot, Narrative Select, Imagen, Excire Foto.
• Staying in Lightroom: Lightroom vs paid cullers.
• Not wanting to pay yet: free and trial options.
• Judging accuracy claims or scoring a tool yourself: how to measure AI culling accuracy.

Check this in your trial

1. Pick 100 frames you have already culled. Run the tool and count the keepers it rejected before you look at anything else.
2. Ask what it does with a burst: does it keep one, rank all, or group them?
3. Confirm what the result writes back to Lightroom (flags, stars, labels) and where.
4. Check whether the reason for a decision is shown, so you can judge it quickly.
5. Watch network activity during a pass if file locality matters to you.

Original test pending: on a licensed wedding gallery per our method, we would capture screenshots of each tool's first pass and record keepers found, false rejects and review time against a human cull.

·