AI culling for portrait photographers: what to expect

Almost everything written about AI photo selection is written for wedding volume, where the promise is that three thousand frames become a shortlist. A portrait session is a few hundred frames of one person in light you set. Half of that promise does not apply to you, and the other half matters more than it does at a wedding. Here is which half is which.

A portrait session is not a small wedding

The sales pitch for automatic selection is a volume pitch. A wedding is 2,000 to 4,000 RAW frames across a long day, shot in rooms you did not light, with people walking through your frame and a ceremony you cannot ask to happen again. Software that takes that pile and hands back a shortlist is doing something obviously valuable, and the hours it saves are easy to count.

Now open a portrait folder. One person, or one family. Light you set and did not move. A backdrop or a spot you chose because it worked. Two or three hundred frames, maybe fewer. Every one of them is in focus, because you were standing three feet away checking. Nothing is going to be thrown out for technical reasons except a handful of frames where somebody blinked or turned away mid sentence.

So the headline benefit mostly evaporates. There is no pile to filter. What is left is a much narrower and much harder problem: you have forty frames of the same woman in the same chair with the same light, and they differ only in what her face was doing. That is the entire job, and it is the part the marketing rarely talks about. It is the same shape of problem that makes a newborn session so slow to pick from, arriving in a different costume.

What the software is actually looking at

Under the hood, current selection tools run every frame through a handful of checks. Sharpness and focus, whether it landed on the eye. Eye state, open or closed. Exposure. Face detection, so it knows who is in the frame and can group by person. And near duplicate grouping, so twelve frames of one pose arrive as one stack instead of twelve separate decisions. Nothing more mysterious than that, and the plumbing is worth understanding before you judge the output.

The interesting thing is how differently those checks pay off depending on what you shot.

Read that last row again, because it changes how you should shop. If you are a portrait photographer comparing tools on how aggressively they filter, you are comparing them on the feature you need least. Compare them on how well they group near identical frames and how fast the interface lets you flick between two of them at full size. Everything else is noise at your volume.

Expression is the job, and it is what software scores worst

A model can tell you that the eyes are open, that the mouth is turned up, that teeth are showing and the face is sharp. Those are measurable. What it cannot do is tell you that the smile arrived a beat too late and looks polite rather than pleased, or that this particular client dislikes how she looks laughing with her head back, or that in the frame where the dad is looking slightly off camera the whole picture suddenly reads as a family instead of four people arranged on a log.

Two failure cases show up over and over in portrait work.

The technically perfect dead frame. Sharp, well exposed, eyes wide open, expression empty. It scores well on every criterion the software has and it is not a photograph anybody wants. It will be near the top of your shortlist.

The half blink that reads as tired. Not a closed eye, so the blink check passes it. The lids are just low enough that the person looks sleepy, and nobody can say why they dislike the picture. This is one of the known hard edges of that check, covered properly in our piece on blur and blink detection.

Both are examples of the same thing: the margins that decide a portrait are smaller than the margins a scoring model was built to see. Be skeptical of vendor accuracy claims here too, particularly the numbers one vendor publishes about a rival, which are marketing rather than testing. The only accuracy figure worth anything is the one you measure yourself, and you can run that test in half an hour on a gallery you already delivered.

A portrait workflow that puts the software where it belongs

Use it as a sorting machine, not a judge. The order matters more than the tool.

1. Let it group first, and do not look at the scores yet. You want the folder broken into stacks of near identical frames, one stack per pose or per look. If the tool ranks within each stack, fine, treat the ranking as a suggestion.
2. Pick one winner per stack, and nothing else. Do not think about the gallery yet. Do not think about variety or the final count. One stack, one frame, move on. Comparing like with like is the only comparison human eyes do quickly and reliably, and it is the decision you make most often.
3. Now lay the winners out in a row. Only here can you see the session as a whole: which looks repeat, which one is missing, whether every frame has the same head tilt.
4. Go back for seconds. Ask which stacks earned a second frame because something genuinely different happened in them. Usually two or three did.
5. Compare the finalists at full size, side by side. This is the only step that needs a big screen. The difference between two good frames of one face is in the eyes at 100 percent, and no thumbnail will show it to you.

Notice that the software is only doing work in step one. That is correct. At portrait volume its job is to save you from opening files, not to have opinions about faces. The same group first order works for family sessions, where the stacks are bigger and the tiebreakers are about who in the group is ruining the frame.

Why one weak pick costs more here

A wedding gallery typically runs 400 to 800 delivered images. One mediocre frame inside that disappears. Nobody scrolling four hundred photographs stops at number 213 and reconsiders whether they hired the right photographer.

A portrait client gets a small set, and looks at every single frame in it, several times, closely, mostly at their own face. A weak frame is not diluted, it is a noticeable fraction of everything they received. And in headshot work one frame goes on a website and a profile and stays there for years, which is why the corporate version of this job has its own rules about consistency and the approved look.

There is a second thing going on that no software knows about. Your portrait client is not judging the photograph. They are judging themselves in it. The frame you love for the light might be the frame where they see the thing they have disliked about their own face since they were fifteen. You cannot fully solve this at the desk, but you can stop assuming your criteria and theirs are the same, and you can build a little redundancy into the shortlist: where two frames are close, keep the one that flatters as well as the one that is beautiful. Some photographers go further and bring the client into the selection on purpose, which works far better for portraits than it does for weddings.

All of which means the sensible target for a portrait session is not speed. Fewer frames, more weight on each one, more time per frame. If you are borrowing time targets from wedding culling advice you will rush the only decisions that mattered.

Is it worth paying for at portrait volume

Honest answer: it depends on how many sessions you shoot, and the arithmetic is simple enough to do in your head.

Prices checked August 2026 from each vendor's own pricing page.

At six sessions a month, ten dollars is under two dollars a session and the question answers itself. At one session a month it is ten dollars a session, and you should ask honestly whether it saved you an hour. Photo Mechanic sits in the table for a different reason: it makes no decisions at all, it just makes you faster at making your own, which some portrait photographers prefer precisely because expression is the thing they do not want scored.

Every one of these has a free look with no card, which is unusually generous for software and you should use it. Run the trial on a session you have already delivered, then compare its picks against what you actually sent. Pay attention to the disagreements, not the agreements. If the trial does not clearly win you back time, do not subscribe.

The constraint nobody puts on the price page

Every tool in that table runs on a computer. Aftershoot, Narrative Select, FilterPixel and Photo Mechanic are desktop apps, and Imagen wants your files uploaded before it can look at them. That is a property of the category, not a flaw in any one product, and it is fine if your bottleneck is analysis speed.

For most portrait photographers it is not. You shoot two family sessions on a Saturday afternoon, you get home at six, and the reason the gallery goes out on Wednesday is not that a model needed four days to score the frames. It is that the first pass could not start until you sat down at the desk on Tuesday night. Where the review happens changes your turnaround far more than how quickly anything scores, which is the whole argument in picking on a phone versus a computer.

That is the gap Kepla is being built for: the first pass on an iPhone or iPad while the session is still fresh, then the same shoot open on the Mac for the finalists, side by side and zoomed in, with every frame it set aside still sitting right there. Nothing is ever deleted, moved or renamed, it only marks what it picked. To be straight with you, that app is still being built. What is live today, and free while we build, is the booking page that holds a date and takes a deposit.

FAQ