What Assisted Culling does not do: four honest limits

Adobe's Assisted Culling is a genuine feature and for a lot of photographers it is enough. It is also narrower than the marketing around AI culling suggests, and Adobe says so itself. Here are the four jobs it is not built to do, read from Adobe's own documentation rather than from a benchmark we did not run, and a half hour test that tells you whether any of them matter for the work you shoot.

What Adobe actually promises

The limits of this feature are not a secret and they are not a scandal. Most of them are visible in Adobe's own help pages, which is more than can be said for the rest of this category.

On the select side, Adobe's Lightroom Classic documentation lists three criteria: Subject Focus, Eye Focus and Eyes Open, each with a slider for the threshold, plus options to reject photos without people's eyes and to decide what happens to frames it cannot call either way. On the reject side there are three rules: exposure issues, misfires, and documents and receipts. Results come back as Selects and Rejects, and batch actions turn them into flags, ratings, labels or collection membership. The cloud version is the same feature with two controls renamed, Subject Sharpness and Eyes Sharpness. Adobe labels the whole thing Early Access.

Then there is the sentence that decides most of this article. In Adobe's own tutorial for the feature, it states that "the current feature set is optimized for portraits and headshots." In Adobe's early access forum thread, the Lightroom product manager puts the target a little more broadly, as individual portraits and multi-person scenes including weddings, events and group photos. Read together, the shape is clear enough: this was built around people, and specifically around faces.

What follows is only the edges. The setting by setting walkthrough is in our Assisted Culling review, the money question of whether a paid app still earns its place is in Lightroom culling versus dedicated tools, and the buying decision by genre and volume is in do you still need a culling app. This one answers a narrower question: which jobs is it not built for.

Limit one: the shoots with no eyes in them

Two of the three select criteria need a face. Adobe even gives you a checkbox to reject photos without people's eyes in them, which tells you plainly what the feature expects to find in front of the lens.

Point it at a shoot with no people and Subject Focus is the only select criterion with anything to grip, while the reject side carries on doing useful work, because misfires, exposure problems and photographs of documents happen on every card ever filled. That is a real but modest job. It is card hygiene, not a first pass.

None of that is a failure on Adobe's part. It is a portrait feature, described as a portrait feature by the company that ships it, and using it outside that scope is a choice you make rather than a promise it broke. If you shoot the bottom half of that table, the useful comparison is with the tools built for near identical frames rather than for faces, which is the job we go through for product sets.

Limit two: it judges frames, it does not thin a burst

This is the gap that costs the most at wedding volume, and it is a gap in design rather than in quality. Nothing in Adobe's documented feature set groups near identical frames. Three select criteria, three reject rules, and not one of them is about the relationship between two photographs.

Compare that with what the specialists document. Imagen's culling page describes using face recognition to group subjects and duplicates. The desktop culling apps cluster a burst so you see the six versions of one moment side by side instead of scattered through a timeline. That clustering is the feature people actually mean when they say a tool saved them an evening.

Picture what that means on a real card. You shoot the kiss at nine frames a second and come home with twelve frames of it. Eleven are sharp, eyes are open in nine, and Assisted Culling will faithfully report that nine of them are Selects. It is not wrong. It has also not reduced the work, because the question in front of you was never which of these is technically fine. It was which one of these nine.

Lightroom does have Auto-Stack by capture time, which collapses frames shot within a few seconds of each other into stacks. It knows nothing about what is in them, but it makes the pile easier to walk, and we cover how to combine it with the culling panel in the Assisted Culling setup guide. If you shoot in short bursts, weigh this limit rather than the accuracy of the eye detection. Eye detection is the part that works.

Limit three: one threshold, and a day that changes every hour

The select criteria are sliders, and you set them for the run you are about to make. That is fine on a headshot day, where every frame was made in one light on one lens at one distance. A wedding is not that.

A ceremony in a dark church, an hour of portraits in open shade, and a reception lit by two uplighters and a phone torch are three different standards of sharpness, and any honest photographer applies three different bars to them. Set the threshold where it belongs for the portrait hour and it will bury most of the reception. Loosen it enough for the reception and almost everything from the portrait hour comes back as a Select, which is another way of saying nothing was decided.

The workable answer is to run it in sections rather than once across the card, which Adobe's design allows since the analysis is pointed at a folder, an album, or a folder chosen in the import window. It works. It is also manual bookkeeping, and the bookkeeping grows with the job: 3,000 frames, two bodies, three cards, and a second shooter whose lens is a stop softer than yours at the same aperture. The feature has no idea which part of the day it is looking at, because nothing in it models a day.

Limit four: the RAW files have to be home first

Assisted Culling runs where Lightroom runs. In Classic that is your computer, after the import or during it. In the cloud app it runs inside an album, and files sitting on a local drive get a Copy to Cloud step before they are album files at all.

So the analysis is fast and free, and it still starts at the moment the cards are in the machine and the RAW frames have been read. That is the honest cost of any catalog based approach: nothing about the first pass happens on the drive home, in the hotel, or on the sofa while the coffee lands. Running it at import is the best version, because it uses the one stretch of the evening you were going to spend waiting anyway. If previews rather than decisions are your actual bottleneck, that is a different problem with different fixes, and we take it apart in why Lightroom feels slow when you are culling.

The other half of this is the feature's best property, so it deserves saying: the result lands in your catalog as flags, ratings, labels or a collection, with no export, no round trip and no second database claiming to hold the truth. Every dedicated app has to hand its answer back across a boundary, and that boundary is where workflows break. Assisted Culling has no boundary. For a lot of photographers that single fact outweighs everything above it.

And the limit nobody can engineer away

Three technical criteria and three reject rules, and not one of them is about what happened in the room. Adobe does not pretend otherwise, which is more than the category usually manages.

The Eyes Open rule is the clearest case. A blink and a father with his eyes shut in the middle of a toast are the same measurement. So are a focus miss and a deliberately dragged shutter on the dance floor, a blown highlight and a window you exposed for the faces, a misfire and the frame you shot from the hip while walking backwards that turned out to be the picture of the night. The score is correct every time. It is answering a question about the file, and you are asking a question about the evening.

This is not a limit a future release fixes, in Lightroom or anywhere else, because it is not a measurement problem. The argument for which decisions should stay with a person even when the software is excellent is in what AI should never choose for you.

Half an hour that settles it for your own work

Do not take anyone's word for how these limits land on your shoots, including ours. You already own the only test that matters, because you have delivered galleries and they are the answer key.

1. Pick a job you delivered at least a month ago, ideally in the genre you shoot most.
2. Run Assisted Culling on the original folder, with the settings you would genuinely use, section by section if the light changed.
3. Compare its Selects against the frames you actually delivered.
4. Write down every delivered frame that did not make its Selects, and note where in the day each one sits.
5. Then look at the other direction: what did it put in Rejects that you kept, and why did you keep it?

You are not looking for a percentage, and any percentage you calculate is about your folders rather than anyone else's. You are looking for a pattern. If the misses scatter randomly, the feature suits your work and you can stop shopping. If they cluster in the ceremony, on the dance floor, in the frames with no faces, or in the bursts, you have just found the exact shape of what a second tool would have to do for you, which is a much better basis for spending money than a comparison table.

Two honest closing notes. First, this feature is free with a plan you were keeping anyway, so for portrait, headshot, school and family work the limits above may never once bite. Second, the one limit nobody in this category has solved is the last one: the first pass still waits for you to get home. That is the bet Kepla is making, with native picking on iPhone, iPad and Mac, and it is still being built, so there is nothing to compare yet and we will keep saying so until it ships. The booking page is live and free today.

FAQ