Read the label off a photo

September 9, 2026

Photograph a carton and its name and barcode fill themselves in for you to check. You can now also choose which reader looks at the picture, and ask for a description of what the thing appears to be.

Open a line to read the whole story behind it.

• New Photograph a carton on the phone and the name and barcode fill themselves in from what is printed on it.

  Cataloguing something with no barcode, or one too scuffed to scan, meant typing the name in by hand. Where your copy has this switched on, the form for adding something new now has a Read the label button above it: photograph the product and its label, and what can be read off the picture drops into the fields below. A barcode printed on the box in ordinary digits is the best case — it is looked up the same way a scanned one is, so what you get is the catalogued name rather than a guess. Everything it offers is only a starting point in an ordinary editable field, and it says underneath each one where the words came from, because a machine reading a photo of a shelf gets things wrong. You can also open what it actually read, line by line, to copy something out yourself. Nothing is saved until you press Save, and the photo you took is kept as the picture for that product, so you are never asked to take a second one. There is a button on any picture you have already added, too, for reading a photo taken months ago. Your own part number is never filled in from a photo: that is your reference for the thing, and nothing printed on somebody else's packaging is it. It is marked experimental while it settles in: it works as described, and what it offers is a starting point rather than an answer.

• New Try a product photo now lets you pick which reader looks at the picture, so you can see for yourself which one reads your own labels best.

  How well a photo is read depends on which reader is looking at it, and the honest answer to "which one is best" is that it depends on your labels — a reader that wins on clean printed boxes can lose on a scuffed sticker in a dim shed. The Try a product photo pane in your settings now has a picker for each step, listing whatever your copy has available, so you can send the same photo through each one and compare what came back and how long it took. Every part of the answer now names the step that produced it — which reader found each line, which model guessed, what a detector saw — with its own timing beside it, and each field it offers says whether it was read off the photo or guessed. The whole answer is there to open at the bottom, for when you want the numbers themselves. This only changes the photo you are trying: everywhere else in the app keeps using whatever your copy is set up with, so nothing about adding a product changes until somebody deliberately changes that setting.

• New A photo can now come back with a plain-language guess at what the thing is, and a list of everyday objects it spotted in the picture.

  Reading a label tells you what is printed on it, which is the useful part and sometimes the only part — a plain brown box says nothing at all. Where your copy has them switched on, two extra opinions now come back with the reading: a short description of what the thing appears to be, and a list of ordinary objects visible in the photo with a score for each. Both are shown as plain words for you to read, correct or ignore. Neither is ever allowed to fill in a part number, a barcode or an SKU: those come only from what is actually printed on the label, because a description that sounds right and is wrong is worse than no description at all. You can turn either of them off for a single photo on the Try a product photo pane, which is the easiest way to see what each one is adding.

• Improved Photos up to 20 MB are accepted, which covers a picture straight off a recent phone without shrinking it first.

  The limit was 16 MB, and a phone that writes big HDR pictures could just exceed it — with a refusal that arrived as a bare error page rather than as a message on the form, because a file too large never reaches the part of the app that could explain itself. Twenty covers what current handsets produce with room left over. It costs nothing in storage: a photo is scaled and re-encoded as it is added, so a picture that arrives at 18 MB is a few hundred kilobytes once it is saved. What your copy actually allows can still be lower, and every screen that takes a picture says the real figure rather than a number it cannot honour.

• Fixed A photo from a phone shooting at its highest resolution can now be read; before, those came back with nothing at all.

  Phones with 48 and 50 megapixel cameras produce photos far larger than the ones this reads best, and anything above about 30 megapixels was quietly impossible to read: it came back as "the reader did not answer", the same words you would get if nothing were running at all, no matter how many times you tried. A photo taken by the same phone in its ordinary mode worked, which made it look as though a particular picture was at fault. Those photos are now read normally, and quickly — a 35 megapixel one takes well under a second. If a photo still cannot be read for some other reason, what went wrong is now written down where whoever looks after this copy can find it.

• Fixed Reading a photo now says it is working while it works, instead of leaving the button looking as though nothing happened.

  Reading a photo takes a few seconds — the picture has to be sent, read, and sometimes looked at by a model that may need a moment to wake up — and until now the page sat there unchanged the whole time. That reads as a button that did not do anything, so the natural thing is to press it again, and the second attempt usually answers quickly because everything it needs is warm by then. The button now says it is reading and stops taking presses until the answer arrives. Nothing else about the page changed, and it still works with scripting turned off.

• Improved The name suggested from a photo is now what the thing appears to be — "canned peas" rather than the longest words on the packaging.

  The name offered when you photograph something used to be the longest line of words the reader found on the box, and the longest line on a box is as often marketing as it is the name: you got "extra fijn" for a tin of peas. Where your copy has a model switched on, the name is now what that model says the thing is, and it answers in the language you are working in. Underneath the box it says where the words came from, because the three possible sources are worth trusting differently: a name looked up from the barcode was written by somebody holding the article, a name read off the packaging is what a reader made of the picture, and a name from a model is a description of a photograph. Every one of them is a starting point in a field you can edit. Nothing changed about the part number, the barcode or the SKU: those still come only from what is actually printed, because a plausible wrong code is worse than an empty box.