What AI culling misses: four kinds of frame

Every culling tool is good at the same things and weak at the same things, and the weaknesses are not random. They fall into four categories you can name, predict, and check in a few minutes per shoot. Here is the taxonomy, why each one happens, and the counter-move for each.

A taxonomy, not a benchmark

This article does not contain a lab test, an accuracy percentage, or a hit rate. We did not run one, and if we had, the number would be about our folders and not yours. The only test worth trusting uses a gallery you already delivered as the answer key, and we wrote the half hour version of that in how to test AI culling accuracy on your own shoot.

What follows is the other half of that job: knowing what to look for when you run the test, and what to check on every shoot afterward. Automated selection fails in four recognizable ways, and once you can name them, an audit stops being "review everything again just in case" and becomes ten minutes in four specific places. The broader argument about whether these models deserve your trust at all is a separate piece, can you trust AI to pick your photos. This one assumes you have already decided to use one and want to know where to stand guard.

What the machine is actually scoring

The clearest public statement of what these tools measure comes from Adobe, because Adobe documents it. Lightroom's Assisted Culling chooses its selects on three signals: subject focus, eye focus, and whether eyes are open. It rejects on exposure problems, misfires, and documents or receipts. Imagen's culling describes detecting blur, closed eyes and blinks, with a special exception carved out for kissing, and grouping subjects and duplicates by face recognition. Different companies, near identical list.

Look at what every item on that list has in common. Each one is a property of the pixels in a single file. None of them is a property of the day you photographed. A model can measure whether an eye is sharp. It cannot know the eye belongs to the grandmother who will not be at the next wedding.

Adobe is also unusually direct about the fit: it says the current feature set is optimized for portraits and headshots. That is not a criticism of Adobe, it is the honest shape of the whole category. The four failures below are what you get when a frame-level technical score is asked to stand in for a judgment about meaning. They are not bugs anyone is about to fix.

One piece of vocabulary before the list. Adobe's eye detection has a third state beside open and closed, a "Can't Tell" bucket for faces it will not commit to. That bucket is the honest one, and everything below lives in the same territory without having a bucket of its own.

The emotionally decisive but technically flawed

This is the famous one, and it is famous for a reason. The father of the bride finishes his speech and puts his hand over his eyes. His eyes are shut. The frame reads as a blink to anything scoring eyes open, and it scores as a reject, because closed eyes because you are crying and closed eyes because your nervous system fired are identical at the pixel level.

It shows up in a dozen shapes. Focus lands on the microphone stand rather than the face, and the face is still the photograph. A guest lunges into the foreground and the composition is ruined, except the ruin is the reaction. Half a face is behind a shoulder, and the visible half is doing all the work. The technical score is genuinely correct in each case. It is just answering a question you did not ask.

The tell: these frames cluster around the two or three minutes of a shoot with the highest emotional temperature, which means they are findable. You do not need to review the whole reject pile. You need to review the frames within thirty seconds of a peak.

The counter-move: decide before you filter which minutes of the day are human-required, and look at every frame in those windows regardless of what the software thought. More on that below.

The intentional

The second category is the one photographers find most insulting, because the software marks your best decision as a mistake. A dragged shutter on the dance floor, so the lights streak and the couple stays sharp. A panned frame of the exit car. A silhouette against the window where the faces are deliberately unreadable. Shooting through a veil, a doorway, a crowd. A frame that is soft everywhere on purpose because the memory is soft.

Every one of those is, measured honestly, a focus failure, an exposure problem, or both. The model is not wrong about the measurement. It has no access to the fact that you chose 1/15 of a second, and there is no field in the metadata that says "on purpose".

It helps to separate blur into three working kinds, because they deserve different treatment. Focus failure, where the plane landed in the wrong place, is the safest thing in the world to let a machine push down. Subject movement, where a hand or a veil moves and the face is sharp, needs a look, because it is often the frame with life in it. Camera movement, where the whole frame is soft, needs the most care, because that is where intent hides. We take that distinction further in when a soft photo is worth keeping.

The tell: this failure is style-shaped, not moment-shaped. It concentrates in whatever you are known for. If your portfolio is full of motion, this category will cost you the most, and it will cost you on every single job.

The context dependent

The third category is invisible in the file. Two frames are technically identical, and one of them matters, for reasons that exist only in your head and in the couple's family.

The examples repeat across every genre. The scruffy guest at the back of the group shot is the bride's brother who flew in from overseas. One wide frame of the room includes the florist's arch and the band, and you owe both of them a usable photograph. Two frames of the same toast look interchangeable until you notice that in one, the grandmother in the background is watching, and in the other she is looking at her plate.

Near-duplicate grouping is where this bites hardest. Similarity detection is genuinely good, and it earns its keep on a burst shot from a tripod. It gets unreliable the moment the scene moves: you step sideways, a subject rotates, someone walks into the edge of the frame. Photos that were interchangeable a second ago stop being interchangeable, and the group still looks like one thing to the machine. Treat a group as a stack to open, not a stack to accept, and look at the first and last frame of the run, because that is where the scene changed.

The counter-move: the software genuinely cannot help here, so stop expecting it to. Capture the context while you still have it. A four line note on your phone at the end of the night is worth an hour of squinting on Tuesday.

The one irreplaceable frame

The fourth category is not a different kind of mistake. It is the same mistakes, weighted by whether a substitute exists. It is the one that should actually worry you.

If a tool down-ranks one frame out of nine of the same hug, nothing happened. You had eight others and you were only ever delivering one. If it down-ranks the only frame of the ring going onto the finger, the moment is gone from the gallery, and no amount of overall accuracy makes up for it. The risk in a cull is never the miss rate. It is the miss rate multiplied by how replaceable the frame was.

Redundancy is your safety net and it is not evenly spread across a day. Bursts, formals and portraits are thickly protected: dozens of near copies, so a wrong call costs you a slightly worse version of the same photograph. Singletons are naked. The first look reaction, the ring, the one time the flower girl looked up, the guest who left at nine: those often exist once.

The counter-move: when you review, spend your attention in inverse proportion to frame count. The scene with 40 frames needs the least of you. The scene with two needs the most. Most photographers do exactly the opposite, because the big pile feels like the big job. The systematic version of this problem, and how to log it, is in AI culling false negatives.

The four categories on one screen

Notice what is not in this table: blinks in a family formal, a genuinely missed focus, the fifteenth frame of the same handshake, a picture of your lens cap. Machines are excellent at all of those and always will be, which is why this article is a list of exceptions rather than an argument against the tools. The categories a machine should never be handed at all are a shorter list, and we made it in what AI should never choose.

Human-required zones, decided before you filter

The practical answer to all four categories is the same, and it costs about ten minutes a job. Decide which parts of a shoot you will review by hand before you look at any suggestion, so the software does not get to shape what you notice first.

For a wedding, the usual list is processional, first look, vows, ring exchange, the first embrace after, speeches and the reaction shots during them, first dance, and the exit. For a family session it is the first genuine laugh after the stiff setup, and the in-between frames where a child goes back to a parent. For a portrait day it is the two minutes after someone stops performing. If you have never written down what you select for, that exercise comes first, and we walked through it in building your culling criteria.

Inside those zones, look at what the tool did not pick. Outside them, trust the picks and move on. That is the entire discipline, and it is affordable in a way that "check everything" never was.

Then keep a rescue label and a short log. When you pull a frame back from the unpicked pile, mark it, so you end up with a compact list to weigh against the gallery instead of a memory. Three columns is plenty: file number, scene, and why you think it was passed over. After three or four shoots the pattern is yours and specific: profiles, or backlight, or the angle your second shooter covers. That is worth more than any vendor's accuracy claim. For where the biggest tool in the category actually stands today, checked against Adobe's own pages, see Adobe's AI for photographers.

FAQ