AI vs manual culling: the same shoot, both ways

Arguments about culling speed are usually arguments about two different shoots. The only honest version of this comparison is one shoot, one standard, walked through both routes with the clock running on each stage separately. Here is that walkthrough, so you can run it on a folder you have on a drive right now.

One shoot, two routes, one standard

This is a method, not a verdict. We have not run your shoot and we are not going to hand you a table of minutes we made up, because the number that matters is the one your own folder produces on your own machine.

Before you start, four ground rules. They are the difference between a comparison and a story you tell yourself.

• Same shoot, or as close as you can get. Comparing last week's wedding to last month's is comparing two weddings. If you cannot face doing one shoot twice, split it: run the ceremony by hand and the reception AI-first, then swap on the next job.
• Same delivery standard. Decide your target count before you start either route. Otherwise the fast route just delivers fewer photos and calls itself fast.
• Time each stage, not the total. The total is the least interesting number here. What you want to know is which stage gave the minutes back.
• Accept the memory problem. The second pass is always easier because you have seen the frames. Do the AI route second and it looks better than it is. Do it first and the hand cull looks better than it is. Note which order you used and hold the result loosely.

A typical wedding is 2,000 to 4,000 frames and 400 to 800 delivered, so that is the scale to picture throughout. If what you want is the argument about what each approach is fundamentally good at, that is a different piece and we wrote it in AI culling versus manual culling. This one is just the walkthrough.

Route one: the hand cull, stage by stage

Ingest and previews

Copy the card, then wait for previews to build. Nothing you do here is skilled work and none of it is optional. On a large shoot this stage alone can outlast your patience, which is why so many people start a cull, get bored, and come back to it two days later.

Pass one: the mechanical sweep

You go frame by frame looking for the disqualifiers: missed focus, motion blur, closed eyes, someone walking through the shot, the flash that did not fire. On most frames this decision takes a fraction of a second and is completely uninteresting. This is where the bulk of your clock goes, and it goes there not because the decisions are hard but because there are thousands of them.

Pass two: the burst groups

Now the slow part. Six frames of the same moment, all sharp, all fine, and you have to pick one. This is genuinely difficult work because absolute judgment is hard and comparison is easier, so you flick back and forth between two frames trying to see a difference that may not exist. Per frame, this is by far the most expensive stage of the whole job. It is also where your keyboard shortcuts earn their keep.

Pass three: the story pass

You stop looking at photos and start looking at the gallery. Is the dress covered? Did the groom's family get as much attention as the bride's? Is there a gap between the speeches and the first dance? This stage is fast, it is the most valuable thing you do all evening, and it is the first thing to be dropped when you are tired.

The trim to number

You are over your target, so you go back and cut. This is the stage everyone forgets to count, and on a big shoot it is not small.

Route two: the AI-first cull, stage by stage

Ingest and point

Same copy off the card, then you point the tool at the folder. If it runs on your own machine, that is the end of your involvement for a while. If it runs in the cloud, the shoot uploads first, and on a normal home connection a full wedding is not a five minute wait. Count that honestly, because it is part of the route.

The analysis run

The machine reads every frame. This is the stage that changes the shape of the whole evening, and not because it is fast in absolute terms. It is because it is not your time. You are not in the chair. The photos get looked at while you make dinner. Route one has no equivalent stage: there, every frame that gets looked at costs you attention.

Review the picks

You open the shortlist and go through it. Critically, this is not the same job as pass one above. You are not deciding on every frame from scratch, you are checking somebody else's homework, and agreeing is much faster than deciding. It is not free, though, and anyone who tells you the review is instant has not done it on a real wedding.

Audit what it did not pick

The stage that does not exist in route one, and the one people skip. You sample the frames it passed over, looking for the ones it got wrong. Do not skip it, especially on your first few shoots, because this is the only way you find out what your tool is blind to. What to actually do with that pile is a question in its own right, covered in what to do with the photos you do not pick.

The burst groups, pre-sorted

The six near-identical frames are now sitting next to each other with one suggested, rather than scattered across a timeline for you to find. The choosing is still yours and it is still the hard part. What you have been given back is the finding and the arranging.

The story pass

Identical to route one. Nothing about it has changed, and nothing about it should.

Where each route spends its minutes

Fill in your own times as you go. The shape of the table is the point: two stages move off your clock entirely, one shrinks, one is new, and one does not move at all.

Read the middle rows again. The saving does not come from the machine deciding faster than you. It comes from the largest block of work moving off your clock and onto a processor, and from two stages of hunting, finding the burst groups and finding the disqualifiers, stopping being your job.

What actually changed, and what did not

Three honest observations from the shape of that table.

Your first AI-first shoot will probably be slower. You will check everything twice, because you do not trust it yet, and you should not. That is not the tool failing, it is you buying information about it. Budget for it, and do not judge the route on shoot one. If you want to make that trust-building deliberate instead of anxious, run the scoring exercise in how to test AI culling accuracy on a gallery you already delivered.

The hard part is still the hard part. Choosing between six good frames of the same moment did not get automated. It got organized, which helps more than it sounds like it should, but you are still the one making the call. The skill involved is real and it is worth developing regardless of what software you own, which is the argument in choosing the best frame from a burst.

The story pass never moves. No model knows the groom's father was ill, or that the couple asked for more of the dog. That stage is the same length in both routes and it is the one worth protecting, because it is the difference between a set of good photos and a gallery.

Why the totals mislead you

Suppose both routes finish in a similar wall-clock time. Most photographers would call that a draw. It is not, for two reasons.

The first is attention. An hour of deciding on three thousand frames and an hour of reviewing a shortlist are not the same hour. Decision quality degrades as you make more decisions, and research on decision fatigue suggests that later choices in a long session are made worse than earlier ones, though parts of that literature are debated. You already know the practical version: the last two hundred frames of a midnight cull get a treatment the first two hundred did not.

The second is when the minutes happen. This is the part the totals hide completely. Culling is a real slice of the job, not a rounding error: a UK survey of more than 300 wedding photographers put it at 11 percent of working time, against 55 percent for editing and just 4 percent for shooting. But the pain of that 11 percent is not only its size. It is that it sits between the shoot and everything else, on a desk, at home, on a Sunday night, blocking the edit that blocks the delivery.

Which is the whole reason we are building the first pass to run on a phone. The analysis stage does not need you in the chair, so there is no good reason it has to happen at the desk at all. Approve picks on the drive home or on the sofa, then open the same shoot on the Mac when you need two frames side by side at full size, with everything it passed over still there. That trade-off between screens is worth thinking about on its own terms, and we went into it in culling on a phone versus a computer. The picking app is still being built. The booking page is live and free today.

How to run this without fooling yourself

Five things that turn this from an impression into evidence.

1. Use a timer, per stage. Not a guess afterwards. Guesses are generous to whichever route you already prefer.
2. Write the numbers down immediately. At the end of each stage, before you start the next one.
3. Count what you delivered from each route. If one route gave you 520 keepers and the other gave 610, you did not run the same test, and the faster one may simply have been stricter.
4. Compare the two galleries, not just the two clocks. Frames that appear in one and not the other are the most interesting output of this whole exercise. Look at every one of them.
5. Do it twice, in the other order. The memory effect is real and this is the only way to see past it.

What you end up with is not a verdict on AI culling in general. It is a stage-by-stage map of where your own evening goes, which is more useful, because it tells you whether your problem is analysis time, comparison time, or the fact that none of it can start until you get home. Those are three different problems and only one of them is fixed by faster software. If your suspicion is that the whole cull is simply taking longer than it should, the ranges to check yourself against are in how long culling should actually take.

FAQ