Culling a shoot
without uploading it

01 The short answer Before the detail

A browser can read a folder of RAW files off your disk and decode them locally, which means culling needs no upload, no account, and no waiting for an ingest to finish. The cost is that it only works in a desktop Chromium browser, and that no local pass will judge an expression as well as a model trained on millions of frames.

What the upload actually costs

A wedding comes home as roughly two thousand frames. On a 24-megapixel body that is somewhere near forty gigabytes, and on a domestic upload of 20 Mbit/s it is more than four hours before the first decision can be made. The work has not started; the files are still moving.

The second cost is not time. Once a client's shoot is on somebody else's disk, its retention is somebody else's policy, and for commissioned work that can be a contractual question rather than a preference. It is a short conversation when the answer is that the files never left the machine, and a long one otherwise.

The third is that the tool stops when the subscription does. A culling pass is not a one-off: it is the thing you do on every job, including the ones you take three years from now on a laptop with no connection.

Why a browser can do this now

Two pieces landed that were not there five years ago. The File System Access API lets a page ask for a folder and receive a handle to it, so the page reads the originals in place rather than receiving copies. WebAssembly then decodes those RAW files at close to native speed, in the tab, with no round trip.

The consequence is worth stating plainly, because it is the whole argument: a page built this way is a viewer, not a client for a service. There is no server holding your frames because there is no server in the path at all.

• The folder is picked by you, and the page sees nothing else on the disk.
• Decoding, previews and measurements all run on your machine.
• Decisions come back as XMP sidecar files written next to the originals, which is what Lightroom and Capture One already read.

Where cloud culling genuinely wins

This is the honest half. The funded tools in this category — Aftershoot, FilterPixel and Excire among them — put real machine learning behind the pass, trained on labelled sets far larger than anything that fits in a tab. On the questions that are about a face rather than about pixels, they are better, and it is not close. Closed eyes, a held expression, which of two near-identical frames a client will prefer: those are their questions.

They also carry a session between machines, and they scale past what one computer can hold. If you shoot three thousand frames a week and bill for the hours, an upload that buys back an afternoon is a good trade.

Vendor behaviour described here was checked on 7 September 2026 and will change.

What a local pass is actually good at

The things that are arithmetic rather than judgement. Sharpness is measurable. So is clipping, so is exposure, and so is the fact that eleven frames were taken inside two seconds and are therefore one burst with one keeper in it. None of that needs a model, and all of it is the bulk of a cull.

What you get in exchange for giving up the expression call is that the first frame is on screen immediately, and the last one is too. There is no ingest, so there is no point at which you are waiting on the tool instead of looking at photographs.

The trade-off, stated once

Neither column is the right answer for everybody. The column on the left is the right answer more often than the search results for this subject would suggest, because the search results are written by the companies in the column on the right.

99 Questions Asked more than once

More from the blog