Local vs cloud AI culling: what actually differs

Every AI culling tool promises to find the sharp, open-eyed, non-duplicate frames. None of that tells you where your client's photos go, how long they stay there, or what happens when the venue wifi dies. This is a fair comparison of local and cloud processing on the four things that actually decide it: total time, confidentiality, offline work and how your team is shaped.

What local and cloud actually mean

Local means the analysis happens on your own machine. You copy from the cards, the app builds previews, and the model reads those previews using your processor and graphics card. Nothing has to leave the building for the picking to happen. Cloud means the analysis happens on someone else's computer. Something goes up the wire first, and the selections come back down.

The middle is where the confusion lives. A product can build previews locally and send only those. It can analyze locally but sync your decisions through an account. It can run for a week with no connection and then want one to check your license. A label on a marketing page settles none of that. The question worth putting to a vendor is narrower and harder to dodge: what leaves my computer, in what form, and when?

Here is where the tools most photographers are choosing between actually sit, taken from their own pages.

Checked August 2026 from each vendor's own pages.

Speed is the whole pipeline, not the model

Vendors publish model time, which is the part of the evening that matters least. The clock a working photographer cares about starts when the cards come out of the bag and stops when the picks are in the editor's hands. Between those two points sit six stages: copy to two backups, import, preview build, analysis, your review, and the handoff.

Cloud tools are genuinely fast at the analysis stage, because the hardware doing it is bigger than anything on your desk. The catch is that the clock starts at the upload, and it is your upload speed that governs it, which on most home plans is the small number. We did that arithmetic properly, hour by hour at five different connection speeds, in our Imagen AI review, and the honest summary is that on fiber it barely matters and on a rural line it is the entire story.

Local tools skip the transfer and pay somewhere else. The first run on a job is usually the slow one while caches and previews build, and the work competes with everything else you wanted the machine to do. On an aging laptop, a cloud service can be flatly faster, and that is worth saying plainly. Hardware minimums differ more than people expect, which we broke down in culling software on Mac versus Windows.

Judge both on your own camera files. A vendor demo built from small sample images tells you nothing about how a tool handles 45 megapixel raws from the body you actually shoot.

The client confidentiality question, asked properly

Photographers hold material other people consider private: children, the inside of homes, medical events, product prototypes, board meetings, and weddings where the couple has asked for discretion. A cloud tool adds a third party to the handling of all of it. That is not a scandal, it is a decision, and it should be a deliberate one.

Start with your own paperwork, not the vendor's. A clause that permits you to edit the images does not automatically answer whether image data may be sent to an outside processor. Corporate work, brand shoots, secure venues and anything under an NDA are where this bites, and you tend to find out late. The questions worth putting to a lawyer are in our guide to AI clauses in photography contracts.

Then read the vendor side: the retention window, whether your content trains anything, whether subcontractors touch it, which country hosts it, and how you delete a project on demand. Vendors do publish this. Imagen's support pages state that photos are stored on Amazon Web Services servers and that it will not use them for anything other than training your profile and editing your photos without your explicit consent. Read the current policy for the tool you actually use rather than assuming the worst about all of them.

Then apply the smallest amount of access that does the job. Do not upload an entire wedding because the button is there if a preview set would do, use unique passwords and two-factor authentication, and end a second shooter's access the day their part of the work ends.

Offline is a real advantage, and local is not the same as safe

The places photography actually happens are hostile to uploads. Basement ballrooms, mountain lodges, a plane home from a destination wedding, a hotel network shared with 200 guests, a corporate site that will not let a camera bag near its network. A local tool works in all of those, provided the app is licensed and the laptop is charged. Aftershoot lists offline access to the software among its plan features, which is the kind of line worth checking before you rely on it.

Test that yourself, though. Some desktop apps still want an occasional connection to verify a license, so put the machine in airplane mode on a Tuesday and try to start a job, rather than discovering the answer in a lodge with one bar of signal.

Cloud workflows can be planned around connectivity rather than abandoned. Upload overnight on a connection you trust, keep a manual rating fallback in Lightroom or Photo Mechanic, and never schedule a client reveal around a transfer that has not finished. If the whole review happens on a web link, keep an exported contact sheet so a service outage does not become your problem in front of a couple.

And be honest about the other direction. Local does not mean safe. A laptop gets stolen from a car, a drive fails in the week you meant to replace it, and a machine with no disk encryption is an unlocked filing cabinet. Turn on FileVault or BitLocker, lock the screen, keep the backups current. Privacy is a property of the whole system, not of one word on a feature list.

What happens to the files once the job is over

Retention is the part of cloud culling nobody asks about until a client does. Find out whether originals, previews and the generated scores are removed automatically, removed when you ask, or held for a stated period. Ask whether deleting a project also clears the backup copies, and whether closing the account triggers anything at all. If a client's obligations are strict, save the answers with the job file so you are not reconstructing them from memory two years later.

Local workflows have the same problem wearing different clothes. Preview and cache folders hold recognizable versions of images long after the job drive has been archived. They deserve a housekeeping rule of their own, and a machine that is being sold or sent for service deserves a wipe rather than a hopeful drag to the trash.

The fix is unglamorous: write the schedule down once. Something like raw masters kept for one year, delivered files kept for two, cloud culling projects removed within 30 days of final delivery, contractor access ended the day the job closes. The exact numbers are a business decision, and ours would not suit a school photographer, but a written schedule is far easier to follow and to explain than one living in your head. If you are still deciding what to archive at all, we worked through that in whether to keep the frames you did not pick.

Which model fits the shape of your team

This is where the cloud earns its keep, and it gets underrated by solo photographers arguing about privacy. If your editor is in another time zone, a cloud tool means you finish shooting, start an upload, and wake up to a marked selection. No drive in the post, no version confusion. For a studio running several photographers and a remote post team, that is a real advantage local software cannot match.

Local keeps a tighter loop. You analyze on the job drive, inspect the clusters yourself, and pass only the selected files onward by whatever method you already trust. For a solo photographer, or a team that shares one editing workstation, that is simpler and usually faster. It gets awkward the moment everyone is distributed.

Whichever shape you pick, the handoff is what decides whether it works. Agree on the selection language before the season starts: flags for candidates, stars for final picks, color labels for the vendor set, and a clear answer on whether the AI's picks are advisory or accepted. What actually survives the trip into Lightroom or Capture One is covered in exporting selections cleanly. A sophisticated collaboration platform cannot rescue an ambiguous rating system.

Seven questions, and one evening of testing

Run every candidate through the same short list. If a vendor cannot answer one of these from its own documentation, that is itself an answer.

• What leaves my computer: raws, JPEGs, previews, or only metadata?
• How long is it kept, and can I remove it on demand?
• Does my client agreement permit third-party processing for this job?
• What happens to my week if I have no internet for 24 hours?
• How long does the whole process take on my real job size and my real connection?
• Do the picks come back cleanly into Lightroom Classic, Capture One or Photo Mechanic?
• Who on my team has access, and how does that access end?

Then spend one evening on the only test that settles it. Take a shoot you already culled by hand, one you know well enough to argue with, ideally 300 to 500 frames including the hard sections: a dim reception, a backlit ceremony, group formals, a veil, glasses, motion. Run your local candidate and your cloud candidate over the same folder. Count the disagreements rather than the agreements, and pay attention to the frame it set aside that you would have delivered. Our method for that, including what to record while you go, is in how to test AI culling accuracy.

Where Kepla sits, said plainly

We are building a competing tool, so read this section with the skepticism it deserves.

Kepla is native on iPhone, iPad and Mac, and the point of that is timing rather than ideology. The first pass happens on the phone on the drive home, with no upload to wait for and no desk to sit down at. When the shoot needs real scrutiny, the same job opens on the Mac and you keep going: two near identical frames side by side, zoomed in, everything still in place. One shoot, two screens, nothing done twice.

Kepla clears. It never deletes, moves or renames a file. It marks which photos it chose, and from a 2,000 frame wedding you get roughly 800 picks in minutes. Everything it did not pick is still where you left it, one tap away. The worst case for a selection tool should be that you disagree with it, never that something is gone.

The honest status: the picking app is still being built, and we will keep saying so until the day it ships. The booking page is live today and free while we build. If you need AI culling this weekend, buy one of the tools in the table above. We would rather you got the work done than waited on us. If provenance and metadata are also on your mind this year, the companion piece is what Content Credentials mean for photographers.

FAQ