Three thousand frames.
Ninety minutes nobody wants to spend.

Choosing between two frames of the same moment

This is what a cull actually is. Not browsing — choosing, five times a second, between frames that differ only in whether somebody blinked and whether the focus landed. Every other tool makes you answer that by looking at one frame and then the next, by which time you no longer remember the first precisely enough.

Why the eyes and not the frame

A portrait with a crisp jumper and soft eyes is a reject, and no whole-frame sharpness number can tell you that. photocull measures on the detected eye region, falling back to the face and then the frame when it has to — and it says which one it used.

Why your judgement and the model's are kept apart

Two separate rows in the table, keyed on (uid, source). A model verdict can never overwrite yours, and the interface draws the model's faintly so you can always tell what is left to answer. An earlier version keyed on uid alone and one automated pass silently destroyed sixty answered frames; the rule exists because of that.

It learns you, and it cannot learn wrong

The preference model trains on your decisions only, and on within-burst comparisons rather than absolute scores — because "is this a good photograph" is a question nobody agrees on, while "is this one better than that one, of the same moment" is a question you answer the same way every time.

Every learned model is cross-validated against the plain rule-based scorer and discarded if it loses. It can get better at predicting you. It cannot get worse than the rules it replaced. That discipline is available to a single-photographer tool and not to anything that has to ship one model to everybody.

Your Mac is the server

Every library product that works on your phone made the same bet: take custody of the files so they can be synced. Apple Photos copies them into an opaque bundle. Lightroom CC uploads them. Both are unworkable for a photographer whose archive is measured in terabytes, and both are unacceptable for anyone who signs an NDA.

photocull does the opposite. One authoritative copy stays on the Mac and the phone connects to it. That removes the hardest problem in any library product rather than solving it — and it is only possible because of what the index actually weighs.

Measured, not estimated

Every number on this page comes from running the thing on a real 435-frame shoot from a Nikon D610. Where a figure is extrapolated to a hundred thousand photographs, it says so.

What a phone would actually have to carry

A phone can browse, judge, rate and search a hundred-thousand-frame library over a connection that only ever carries kilobytes, and pull a 200 KB preview when you open one. Neither Photos nor Lightroom can do that, because both assume the original is present.

Getting a photograph on screen

Those four numbers decided the architecture: never decode an original to draw a grid. The camera's thumbnail appears at once while a proper one is made in the background, and two sizes come out of one decode because the second is free.

Nothing is trapped

A shoot culled here opens in Lightroom or Capture One with the flags and stars already on it. photocull does not need to win your whole workflow — it takes the ninety minutes nobody wants and hands the rest back.

And the cache is a published contract rather than a proprietary catalogue. Two languages already read and write it under one documented rule. A third could. Delete it and nothing irreplaceable is lost: every measurement can be recomputed, and the one table that cannot — your decisions — is the one the design protects.

The library

About these screenshots

They are of a mock shoot the project builds for itself — tools/make_sample_shoot.py — because a sample folder of stock photographs is not a shoot. What it fabricates is the structure: scenes shot in sequence, bursts within them, one good frame per burst and the rest spoiled the way frames actually get spoiled, plus the duplicate from a card copied twice and the EXIF that advances plausibly. The photographs inside it are real — scikit-image's sample set, and NASA's portrait of Eileen Collins, which is public domain. Nobody's client is on this page.

What it isn't

A tool that tells you what it cannot do is more useful than one that guesses. The same goes for its own description.

There is no develop module, and there never will be

Lightroom's raw processing is not an algorithm you can read and implement. It is dual-illuminant colour profiles measured for around a thousand camera bodies, two thousand measured lens profiles, a proprietary demosaic, highlight reconstruction that keeps skin from going magenta, and a jointly designed locally adaptive tone operator revised five times in twenty years. Three years of work would produce something that looks worse on a D610 than Adobe looks on its first day.

Culling does not need it. The camera's own embedded preview shows focus and expression, rendered by the people who made the sensor. The one real cost is that a JPEG preview clips earlier than the raw, so the clipped-highlight measure reads high — which is written down rather than papered over.

Not on an aeroplane

The Mac has to be awake and reachable. That is the direct, unbendable cost of refusing to sync, and it buys the only sentence worth saying: your photographs never leave your machine.

It says when it does not know

Orientation detection abstains rather than guess, because rotating an upright frame is corruption while leaving a sideways one is merely a nuisance. Frames it is unsure of are marked unsure rather than quietly assigned. A professional tool that says "I am not certain about these thirty-three" is worth more than one that silently decides.

Where it is in progress