How AI picks your best photos: what it actually measures

Photographers trust a tool more once they know what it is looking at, and they stop expecting things it was never built to do. So here is the plain version of how AI culling works: the four checks that do almost all the work, what each one is genuinely good at, where each one falls over, and the parts of your wedding day that no software can see.

It measures. It does not understand.

An AI culling tool is not looking at your photo the way you do. It is running a short list of measurements over a grid of numbers. Four checks do almost all the work:

• Sharpness. How fast does brightness change across the edges in this frame?
• Faces and eyes. Where are the faces, and how open are the eyelids?
• Duplicates. Which frames were shot seconds apart and look alike?
• Exposure. How much of this frame is pure white or pure black?

Everything else is built on those four numbers, and none of them is a judgment about whether the photo is good, because none of them can be. They are physical properties of the file. A tool measuring edges and eyelids will be superb at catching the frames where somebody blinked, and useless at knowing that the woman blinking is the bride's mother. Hold both facts at once and the category makes sense. We covered the trust side in can you trust AI culling. This is the mechanics underneath.

How software sees blur

To a computer a photo is a grid of numbers, one per pixel, each saying how bright that spot is. Now think about an edge, the line where a dark suit meets a white shirt. In a sharp photo you cross that boundary and the brightness jumps hard within a pixel or two: dark, dark, bright, bright. In a blurred photo the same boundary slides up gently over fifteen pixels.

That rate of change is the whole measurement. The software sweeps a small window across the frame, asks at every point how fast brightness is changing, and adds it up. Fast changes everywhere means sharp, slow means soft. It is arithmetic, not opinion, which is why it is the most reliable check in the box. It also explains two things you have noticed.

Motion blur reads as soft, even when it is beautiful

A veil caught mid air, a first dance at a slow shutter, a bouquet toss with intentional streak. Their edges change slowly, so the math says soft and a strict tool scores them down. The frame you are proudest of can be the frame the tool likes least.

A photo focused on the wrong person can still score sharp

If the tool measures the whole frame, all it needs is something crisp somewhere. Focus that landed on the groomsman behind the groom still produces fast brightness changes, so the frame passes. Better tools narrow the check to the face they found. The same quirk means a lace detail shot, full of tiny edges, can outscore a sharp portrait against a plain wall.

Faces first, then the eyelids

Eye detection happens in two steps, and the order tells you where it breaks.

Step one: find the faces. A face detector has been trained on a huge number of labeled images until it learned the pattern of a face from many angles. Hand it your frame and it returns boxes, plus a number saying how sure it is.

Step two: look inside the box. Within each face it finds landmark points, the corners of the eyes and the line of the lids, then works out how far apart the lids are relative to the width of the eye. Wide gap, open. No gap, closed. That ratio is the score you see.

So the failure cases are not mysterious. They are situations where the gap is small for a reason the software cannot understand.

• A real laugh. Genuine laughter squeezes the eyes almost shut. The ratio says closed. The photograph is the best one of the speeches.
• A kiss. People close their eyes when they kiss. That is the point of the photo.
• Looking down. Reading vows, looking at the ring, a child looking at their shoes.
• Sunglasses. No eye region to measure, so the tool skips it or guesses.
• Facing away, in profile, or thirty feet back. No face box, no eye score, so the frame is judged on sharpness alone.

The vendors know this. Imagen publishes a kiss recognition feature noting that some closed eyes are still keepers, a carve out that exists purely because the eyelid measurement gets kisses wrong. Narrative Select reports closed eyes and lists blinking, kissing and looking down together, giving you a measurement rather than a verdict.

Why it groups your bursts

This check is the least glamorous and probably saves you the most time. The software reads the capture time written into each file and finds frames taken within a few seconds of each other. Then it builds a small visual fingerprint of each, a compressed summary of light and shape, and compares fingerprints inside that window. Frames shot close together that also look alike get grouped as one moment.

The grouping is the useful part. Instead of six frames of the ring exchange sitting in a list with forty other files between them, they arrive as one set, side by side, at the same size.

That matters more than it sounds. People are poor at absolute judgment and good at comparison. Asked "is this a good photo", you hesitate. Asked "is this one better than that one", you answer instantly. Grouping turns thousands of hard questions into a few hundred easy ones, which is why a decent tool feels lighter rather than just faster. More in why picking photos is so hard and our guide to duplicate photos.

Where it struggles: a slow scene shot one frame every thirty seconds may not group at all, and if you run two bodies, clock drift scatters a moment that belongs together. Sync your clocks first.

The two weakest signals

Exposure checking is simple. The software builds a histogram, a count of how many pixels sit at each brightness level, and looks at the ends. Pixels pinned at pure white mean blown highlights, pixels at pure black mean blocked shadows. Both get flagged.

The trouble is that the machine cannot tell a mistake from a choice. A low key portrait lit by one window is measurably identical to an underexposed accident. A backlit couple at golden hour is measurably identical to a metering error. If you shoot dark and moody, a strict exposure filter will argue with you all day.

Composition checks are weaker still. A tool can measure whether a horizon is level and roughly where the subject sits, but those are rules of thumb, not taste. Treat both as hints. Over trusting them is one of the most common culling mistakes we see.

What each check is good at, and where it fails

Read that as a division of labor. The top rows are mechanical, and the machine is better at them than you are at 1 AM. The bottom rows are editorial, and you are better at any hour.

What a confidence number actually means

When a tool shows 0.94 next to "eyes open", people read it as a 94 percent chance the photo is good. It is not that. Confidence is the model's estimate of how well this frame matches its training examples: of all the eye regions that looked like this one, this is how consistently they were labeled open. It is pattern similarity, nothing more. A frame can score 0.99 on eyes open, 0.98 on sharpness, and still be the worst photo of the wedding.

The number is most useful when it is low. Low confidence means the frame is unusual: harsh backlight, a partial face, an odd angle, an expression the model has not seen much of. That is a flag saying a human should look, and it deserves more attention than the high scorers, which are the easy cases.

So judge a tool by how fast you can overrule it. If seeing the frames it passed over takes four clicks and a filter change, it has quietly become the decider. If they are one tap away, it is still an assistant. That interaction, more than any accuracy claim, is what you are buying, and we compare it across products in the best photo culling software.

What it cannot know about your day

Every measurement above is about the file. None is about the wedding, and that gap is permanent.

• It does not know who the bride is. It can group faces and tell you face number four appears in 380 frames. It cannot tell you face four is the person paying you.
• It does not know the man crying at the back is the father. To the software he is a small, partly obscured face in poor light. To you he is the photograph the family will frame.
• It does not know the couple asked you to get their dog. There is no measurement for a request made in an email in March.
• It does not know what the light meant, or that you are heavy on the ceremony and thin on the getting ready hour. Gallery balance is invisible to it.

So let the tool do the mechanical pass and keep the editorial pass. A UK survey of more than 300 wedding photographers, reported by PetaPixel, put culling at 11 percent of working time against 4 percent spent shooting. Most of that 11 percent is mechanical. Almost none of the value is. And a frame the tool passed over is not rejected, it just scored lower on four measurements, which is why we wrote separately about the photos you do not pick.

How we are approaching this at Kepla

We are building a picking app, so treat this as interested rather than neutral.

A shortlist with reasons attached

From a wedding of around 2,000 frames you get roughly 800 picks, each carrying the reason it made the list: sharpest of its group, eyes open, the only clean frame of that moment. Reasons tell you when to argue. A pick labeled sharpest in group, where you wanted the softer version, is a disagreement you spot in half a second.

Nothing is ever deleted, moved or renamed

Kepla marks which photos it picked. Every original stays where it was, under the same name, in the same folder, and anything it passed over is one tap from being added back. If disagreeing is expensive, you stop disagreeing, and then the software is making your editorial decisions by attrition.

Phone first, then the Mac

The first pass is a run of quick yes and no judgments, which work fine on a phone on the drive home. Comparing two nearly identical frames at full size does not, so the same shoot opens on the Mac and you carry on. More in phone versus computer culling.

Status, plainly: the picking app for iPhone, iPad and Mac is still being built and has not shipped. What is live today is the Kepla booking page, free while we build.

FAQ