What is photo culling? Meaning, ratios and how it's done

By Ayush Patnaik, founder of AccioPix · Published 4 September 2026 · Updated 20 September 2026

A photographer at a modern wedding captures several thousand frames. Only a few hundred ever reach the client.

Between the raw files on the memory card and the delivered gallery lies photo culling. It is the review where a photographer looks at every frame and throws out the failures: the missed focus, the blinks, the bad exposures, and the redundant copies in each burst. On a large shoot, this mechanical sorting takes more time than colour grading itself.

The filtering gate in raw capture

What is culling in photography, stripped down? It is the gate that picks which pictures make it to editing. Everything else gets a reject flag.

It helps to separate the four stages of the workflow:

• The shoot: Capturing bursts and single frames across the event.
• The cull: The survival filter. Nothing about the image is changed; you only decide which frames exist.
• The edit: Colour, tone, exposure, and retouching.
• The album selection: Picking the narrative arc from the delivered keepers.

Consider the arithmetic of a typical assignment. You return from a wedding with 3,200 raw files on two memory cards. A third of those are bursts: three, five, eight frames of the same second because the camera was firing continuously during a speech. Several hundred more are test shots, transit frames, or flash recycle failures. Culling is the work of turning 3,200 raw files into the 700 that you will open in Lightroom.

The arithmetic of the shoot

Every ratio you will read for this comes from a company that sells culling software, so treat them as claims rather than measurements. Narrative's culling FAQ says a typical full-day wedding generates 2,000 to 3,000 captured frames, that the average delivered gallery runs to around 800 images, and that the ratio from shot to delivered is "roughly 3:1 or 4:1 at minimum". AfterShoot's guide puts the capture side higher, at "4,000 to 8,000 RAW files after a 12-hour wedding shoot". Both pages were read on 20 September 2026. Nobody has surveyed this, so the honest summary is a range: a few thousand frames in, a few hundred out, and the delivered count set by your contract rather than by any rule.

Time is the second cost. Going through 3,000 frames by hand takes between two and four hours. For a photographer who shoots 40 events a year, culling is a tax. It means over 120 hours spent looking at near-identical images.

Counting the mechanical rejects

Ratios of 3:1 invite an assumption worth testing: that most of a cull is throwing out broken frames. So we measured one. The folder is a single wedding session, 245 frames shot raw plus JPEG on a 24 megapixel Nikon, numbered DSC_0001 to DSC_0250 with five frames missing, which makes it a whole card rather than a set somebody had already picked over.

Add up what a machine can throw out without an argument, the badly soft frames, the blown one and the spare frames inside bursts, and you get about 7 per cent of the shoot. The vendors put the cull at 70 to 75 per cent. The gap between those two numbers is the whole job: defects are a small minority of what gets cut, and the rest is a photographer deciding which of four technically fine frames of the same moment is the one worth keeping.

Three limits on that figure, and they all point the same way. It is one session of 245 frames, not a full wedding day. Blinks and expressions were not measured at all, because that needs a face landmark model, so the mechanical share is an undercount. And the sharpness thresholds are relative to this shoot's own median, because an absolute number for the variance of the Laplacian means nothing across different cameras and scenes. That measure comes from microscope autofocus research by Pech-Pacheco and colleagues in a 2000 ICPR paper, popularised for ordinary photographs by Adrian Rosebrock's blur detection tutorial.

Do not take ours on trust. Run it on your own shoot. The script reads a folder in place. It sends nothing anywhere and prints counts only. It needs Pillow, NumPy and SciPy:

python cull-lab.py "/path/to/shoot" --ext jpg

# sharpness   variance of the Laplacian on a 1024 px grayscale copy
# exposure    share of pixels clipped to pure black or pure white
# bursts      64-bit dHash plus EXIF capture time; frames within 2s
#             whose hashes differ by 5 bits or fewer are one burst

A warning from writing it. The first run showed no bursts at all. DateTimeOriginal sits in the Exif sub-directory, not at the top level. The reader quietly returned nothing. A measurement that fails silently looks just like a finding.

Seven technical failure modes

Most discarded frames fail on plain technical errors. The main categories are familiar:

• Missed focus: The camera missed the eye. The whole image is soft or the subject moved. Soft frames are cut unless the moment is unique.
• Blinks: In a group of ten people, someone is usually blinking. Bursts exist to find the single frame where everyone has their eyes open.
• Bad expressions: Mid-sentence mouths, chewing, or the awkward fraction of a second before a smile.
• Exposure errors: Flash recycle delays, harsh backlighting changes, or blown highlights on white clothes.
• Burst duplicates: Ten frames of the ring exchange. You keep the single best frame.
• Accidental frames: Lens cap checks, shots of the floor, and metering frames.
• Obstructions: A waiter walking across the frame, a stray arm, or bad framing cutoffs.

Culling is not about artistic taste. A boring photo of the hall survives the cull as long as it is sharp. It can always be dropped later, when making the album. The job of culling is modest: throw out the broken frames so that creative choices happen on a clean deck.

Decoupling rejection from selection: a two-pass protocol

Looking at every frame and deciding both whether to reject it and which variant is best is slow. It is better to split the work into two clean passes:

Pass 1: Throwing out obvious defects

Move through the shoot at high speed, flagging only clear technical failures: dark frames, obvious blur, and blinks. Do not compare frames, and avoid zooming in unless focus is genuinely in doubt. On 3,000 frames, this pass takes under 45 minutes and removes a third of the shoot.

Pass 2: Picking the winner of each burst

With outright failures gone, inspect the surviving bursts. Put the candidates side by side, zoom to the eyes, and mark the single winner. The remaining copies get rejected. This is where Lightroom Classic shortcuts and survey views save hours.

Pass 3 (Optional): Trimming to the count

If your contract promises 600 photos and you have 800 keepers, a final pass cuts the weaker moments. You already have one clean frame per moment, so this pass is about balance across the day.

The software landscape

Photo culling software comes in three kinds, and photographers usually end up with two of them:

1. Raw editors (Lightroom Classic, Capture One): The standard suites. They offer ratings, flags, and survey modes. The drawback is preview rendering: unless embedded previews are set on import, preview lag slows down the cull.

2. Fast preview viewers (Photo Mechanic, FastRawViewer): Built for sports and wire photographers. They read the JPEG embedded in the raw file, opening frames instantly without waiting for raw decoding.

3. AI culling software: Desktop apps (AfterShoot, Narrative Select, FilterPixel) that automate passes one and two. They score sharpness on faces, check for blinks, group bursts, and propose picks.

For a browser-based option, our Photo Culler runs the entire engine locally on your computer via WebAssembly and WebGPU. It checks edge sharpness on faces, detects closed eyes, groups bursts, and exports XMP sidecars that Lightroom reads on import. Related tools solve smaller tasks: the Photo Selector picks a clean set across the shoot, the Blur Culler drops soft frames, and the Closed-Eye Checker flags blinks in group portraits.

The limits of computer vision

The mechanics of ai photo culling are straightforward. Sharpness is edge contrast on facial landmarks; blinks are measured by eye aspect ratios; bursts are grouped by capture timestamps.

What software cannot see is human context. A slightly soft photo of a grandmother crying during a farewell matters far more to a family than a sharp photo of an empty mandap. Software can clear the technical failures. It cannot know which people matter. That final check is why the photographer is in the loop.

The delivery bottleneck

Culling fast only matters if the rest of the delivery is fast too. If a gallery is culled in two days but takes two months to reach guests, the client experience remains broken.

The delivery layer is where AccioPix operates. The photographer uploads the culled, finished gallery once. Guests submit a selfie via WhatsApp, and the platform delivers each individual their personalised collection within seconds. No applications to install, no passwords, and zero manual sorting for the host. More details on scaling studio operations are covered in our guide on how to grow your photography business.

Frequently asked questions

What does culling mean in photography?

Culling is the review pass right after a shoot where you throw out the frames you cannot deliver: soft focus, blinks, bad lighting, accidental shots, and duplicates. It happens before editing so you never waste time working on a frame that should have been cut.

How many photos should you cull from a wedding?

On a typical full-day wedding of 2,000 to 3,000 raw frames, a photographer usually delivers 400 to 800 photos. That means cutting roughly three out of every four frames. For big multi-day weddings with two or three cameras, total frames easily cross 5,000 and the share cut is even higher.

How long does photo culling take?

Culling a full wedding by hand takes two to four hours. AI culling tools do the first mechanical pass in a few minutes, leaving a 20-minute human check at the end.

Is culling the same as editing?

No. Culling decides which photos survive. Editing decides how those survivors look (exposure, colour, and skin cleanup). Culling comes first so you never spend editing time on a frame that gets rejected.

Can AI cull photos reliably?

AI is good at the mechanical checks: finding blur, closed eyes, exposure blowouts, and duplicate burst frames. But software cannot understand family ties or real emotion. A quick review by the photographer is still essential.

References

• Narrative, Photo culling: the complete FAQ for photographers. Read 20 September 2026.
• AfterShoot, Best AI culling software for wedding photos. Read 20 September 2026.
• Pech-Pacheco, J. L., Cristobal, G., Chamorro-Martinez, J. and Fernandez-Valdivia, J., 2000. Diatom autofocusing in brightfield microscopy: a comparative study. 15th International Conference on Pattern Recognition (ICPR), vol. 3, pp. 314-317. The origin of the variance-of-the-Laplacian sharpness measure used above.
• Rosebrock, A., 2015. Blur detection with OpenCV, PyImageSearch. Read 20 September 2026.
• The 245-frame measurement in "Counting the mechanical rejects" was made with the script above, run on one photographer's own wedding session. The photographs are private and are not published.