What this does

Takes photos that have been scanned but not yet written up, looks at each one, and writes a single warm line naming the people recognised in it. A group photo gets one line naming everyone, not one line each.

It works through a batch per run. Run it again until the queue is empty.

What is seen by the model

Matching happened on this machine. Writing does not: each photo goes to whichever model you have configured.

What travels is a downsized copy with its embedded details removed — camera, capture time and any location are stripped before it leaves, and the copy exists only for the length of the run. The full-size original is never sent and nothing is kept afterwards. If you would rather nothing left at all, configure a local model; the step is otherwise identical.

Changing the wording

What to write is read as "produce a ⟨this⟩", so name the thing you want — "a short playful caption", "one plain line". An instruction like "write something funny" reads as a request to produce a thing called "write something funny", and the result gets stranger rather than funnier.

The names of the people recognised are added for you and cannot be moved or removed.

⚠ Changing the wording does not rewrite anything already written. It applies from now on, so a mixed batch will carry two styles. Re-running will appear to do nothing, because those photos have already been written up.

What it costs

Measured, not estimated — one photo through this recipe against a real model:

The picture is a flat 258 because a 768-pixel copy lands inside one image tile; it costs the same whether the original was a phone photo or a wall poster. That is the whole reason the copy is made — sending a 4000-pixel original tiles into roughly twenty-four of those, so the copy is about a twenty-fourth of the bill.

Photos per run is therefore your spend limit. The default 25 is about 9,000 tokens in. The most a single run will do is 200, roughly 72,000 — ask for more and the run stops rather than quietly doing part of it. On a fast model a full batch is a fraction of a penny; on a quality model, a few pence.

It asks for the fast model, because an unexpected bill is the worse surprise and a caption is not where a frontier model earns its keep. If the lines read flat, copy this recipe in the Kitchen and change the model it asks for — the choice is fixed when the recipe is written rather than each time it runs, so that the cost of a run is knowable before it starts.

The three counts

Could not read the photo means the file was not where its record says. Almost always it was moved or deleted after the scan; re-scanning repairs the records.

Model refused or failed leaves the photo in the queue, so re-running retries only those. No photo is ever written up twice.

Could not save means the model answered but the record did not take it — re-running costs a second call for that photo. Worth a look before you re-run.

Needs libvips

brew install vips, or apt install libvips-tools. Each photo is shrunk to a small stripped copy before the model sees it — the original never leaves the machine, and the copy is deleted when the run ends.

The preview width is also what this step costs. Models bill pictures by area, so the 768-pixel copy is about a twenty-fourth of the bill of sending the photo untouched. Raise it only if you can show the model is missing faces at 768.