AI photo culling for Mac: what runs on the machine, and why that is the whole point

A Mac with Apple silicon has a Neural Engine, a GPU and a handful of performance cores that sit idle for most of a photographer's evening. An AI culling app that runs on the machine puts them to work on the one job nobody wants: looking at every frame from a shoot and finding the plain misses. This guide explains what that kind of app does, what it should never do, and how to judge one before you hand it a wedding.

What a culling app for Mac is actually doing

Culling has two halves. The first is mechanical: find the frames where the eyes are shut, the focus missed, the exposure is gone, the burst produced nine copies of one instant, or nobody is in the frame. The second is judgment: of the frames that survive, which ones tell this couple's story. A good app for Mac takes the first half and leaves the second alone.

That split matters more than any feature list. The mechanical half is where the hours go. A 4,000 to 5,000 frame wedding is four to eight hours of looking by hand, and most of that time is spent confirming that a frame is unusable, not deciding that it is great. The judgment half is the part of the job worth keeping, and it is the part a model trained on other people's taste gets wrong in ways you cannot see.

Kepla is built on that split. It clears the obvious misses, names the reason on every frame it sets aside, and hands you the clean frames to choose from. On the reference wedding we use for every number on the site, that is 4,812 frames in, 2,836 set aside with a reason, 1,880 clean frames left, 96 it would not judge, and 0 deleted. The choosing stays yours.

Why local beats a queue

Cloud cullers ask for your card first. A wedding is 100 to 240 GB of RAW files, and the upload alone is the long pole: their own pages quote forty minutes to three hours for a shoot this size, most of it upload and queue behind other customers. A Mac app skips that step because there is nothing to send. Every frame is read where it already sits, on the card or on the drive you just copied it to.

Speed is the obvious win, but two others come with it. Privacy: a client's wedding never leaves the machine, which is a sentence you can say to the client without a footnote. And independence: with the wifi off, at a venue with no signal, the app runs at exactly the same speed. Nothing about the job depends on someone else's server being up.

Measured on a MacBook Pro M3 Max on battery, one 128 GB card of 4,812 frames, copying to two drives at 214 MB a second. Competitor ranges are the times their own pages quote for a shoot this size.

The four decisions that make it fast

Four minutes is not a faster computer. It is four decisions about where the work happens.

• It reads what the camera already made. Every RAW holds a full size JPEG the camera wrote at capture. Reading that is a fraction of the work of decoding the RAW, and it gives the same answer for blinks, blur and duplicates. The app opens the real RAW only where the answer needs it: 214 frames of 4,812 on our reference wedding.
• It works in the gaps of your own copy. The card and the drives are the bottleneck. While bytes move, the Mac is idle. Kepla works in that gap, so the sort finished 19 minutes 48 seconds before the copy did.
• It never leaves the Mac. Neural Engine and GPU, sixty four frames at a time, on battery at 21 W.
• It does the work once. Every decision lands in an XMP sidecar beside the file. Add a card, change your mind, run it again, and it touches only what changed.

The embedded JPEG versus RAW decode question gets its own article, because it is the one most people have not heard of and the one that saves the most time.

The three lines a Mac culler should not cross

Because the app has your files in front of it, the rules about what it touches matter more than on a web service. Three lines are worth checking before you trust any local culler with a client's card.

It never deletes and never writes a reject flag. Deleting is obvious. A reject flag is subtler: it is one keystroke from gone inside Lightroom, and you have to trust it blind. Kepla writes neither. Everything it sets aside stays on the card and on both drives, with one of five plain reasons attached.

It never writes to your card. Sidecars go beside the copies on your drive. The card is read and nothing else.

It never judges taste. No opinion on the pose, the moment or the look. Only what a loupe can settle. When a frame is blurred in a way that might be deliberate, the app asks instead of deciding: those are the 96 uncertain frames on the reference wedding.

How to test one on your own Mac

Take a shoot you already culled by hand. Copy the card, run the app during the copy, and compare two numbers: how many of its set aside frames you would have kept, and how many of its clean frames you would have thrown out. The first number is the one that matters, because a wrongly cleared frame is the failure you would never notice on a tool that hides its rejects. On a tool that prints the reason on the frame, you can check all of them in one filtered pass.

Kepla for Mac requires macOS 14 Sonoma or later and runs on Apple silicon and Intel as a universal binary, notarised by Apple. It is free through the private preview, and the first hundred photographers keep it at $99 a year for as long as they keep the app. Request Founding Access and we email you the short form and your place in the founding 100.

FAQ