584 photos, one Monday

On a Monday in June I photographed a friend's registry-office wedding. By the evening, 584 RAW shots sat on the memory card, and I handed the AI the task that usually eats many hours: sift everything, rate everything, lay down a first editing pass — as sidecar files for Lightroom, so every decision stays visible and reversible there. With one instruction that mattered more than any of the tech: rate generously, emotion counts. “Because of emotional value … a blurry picture can absolutely have its charm.”

One picture first, then all of them

The AI didn't start with the 584 — it started with one. And the rules for the file format didn't come from documentation but from reality: I edited a single photo by hand in Lightroom — three stars, a touch of exposure, a small crop, a slight rotation — and had the AI read the resulting file as the template. What Lightroom itself writes is the truth; everything else would have been guessing. The first trap surfaced immediately: on a German Windows, the numbers would have carried a decimal comma instead of a point — and Lightroom would have silently discarded the values.

Then the work scaled: the AI split the shots into 24 blocks and sent a team of helper AIs sifting in parallel — each got the rating rules, each returned structured verdicts. A calibration block ran first, and it caught the day's most dangerous mistake before it happened: one helper saw portrait shots “lying sideways” and wanted to rotate them upright by 90 degrees. In truth they were only sideways in the preview — Lightroom already rights them via a flag in the file. That rotation would have wrecked perfectly good pictures. The sign convention for straightening, too, was only settled by a counter-test in Lightroom; no documentation would give it up.

Halfway through came a correction of a different kind. The first editing draft leaned bold — I pulled it back: “What matters to me in this round is natural skin tones; the picture doesn't have to pop.” Taste can't be guessed into existence; it has to be said. After that, the line held.

584 out of 584

By the end of the evening every single shot had its sidecar: 55 five-star highlights, 182 four-star keepers, a solid middle field, 16 clear misses — and ten pictures carrying a label invented for this job, “Emotional”: technically flawed, humanly indispensable. Cropping and straightening stayed deliberately conservative; across all 55 highlights, exactly two micro-corrections of 0.4 degrees were needed. The highlights themselves I edit by hand — and on the final selection the human kept the last word: mostly I went along with the AI's ratings, on a few pictures I diverged sharply, in both directions.

Two side finds made the run worth more than planned: three shots from one narrow time window turned out to be corrupted — the cause stayed unresolved — but they were discovered that same evening instead of months later while hunting for exactly that one picture. And when Lightroom seemingly received nothing, the pipeline wasn't to blame but an interface trap: Lightroom reads metadata only for selected photos. Select all first, then read.

The pushback belongs on record too: one helper AI had simply skipped two of the 55 highlight frames — caught only because the AI checked its own numbers against the list. And my most-typed sentence of the evening concerned administration, not photos: I nearly clicked my fingers sore on permission prompts for the helper AIs before asking for batching. Batched; ran.

What the AI can genuinely see while sifting — and what it invents once you ask it to describe — is a story of its own: AI sorts my wedding photos — and invents what it doesn't see.