I was preparing a bottle as a gift and wanted my own label on it. It is exactly the kind of job you assume AI does in five minutes, and that is half true. The image came quickly. Then the work started, because a wine label carries more than most people know, and an image model knows none of it.
The image is the easy part
Two things to know up front, and you save yourself an evening.
Don’t let the model set type. Image models produce letters that look like letters from across the room and like nothing at all up close. That holds even for the models that have got better at it: give them a name you invented and it falls apart anyway. Ask for an image with no text at all and place the letters yourself afterwards, in whatever program you like.
Ask for a technique, not a style. “Beautiful wine label” gets you the internet’s average taste: gold foil, a chateau, a flourish. “Lino print in two inks, lots of white, one shape” gets you something you can repeat on the next bottle.
Design an image for a wine label, portrait format, with no text or letters
of any kind.
Technique: [lino print / risograph / pen drawing] in [number] colours.
Subject: [one concrete thing, not a scene].
Leave the bottom third of the area empty, text goes there.
No gold foil, no chateau, no coat of arms, no flourishes.That last line is there because the model keeps falling back on those. And the empty bottom saves you the most trouble later.
The mandatory information is the real work
This is where the risk sits, and it is larger than it looks, because the rules changed twice recently. Since 8 December 2023 (Regulation 2021/2117) wine has to carry an ingredients list and a nutrition declaration, like any other foodstuff. On 18 March 2026 the conditions around the QR code were tightened by Regulation 2026/471.
What that means for a bottle sold in the Netherlands:
Always physically on the bottle. The energy value in kJ and kcal, marked with the E symbol, and all allergens: sulphites, egg and milk derivatives, lysozyme. Plus the alcohol percentage, the volume, the origin, the bottler and the lot number. For sparkling wine, the sugar content too.
May go digital, behind a QR code. The full ingredients list and the rest of the nutrition declaration. But that page may carry no advertising, set no cookies or tracking, has to land directly on the information, and has to be in the language of the market. The word “ingredients” has to sit next to the code.
Type size. At least 1.2 millimetres x-height for the mandatory information. That is bigger than most designs can absorb, and it is why you need that empty bottom third.
This is precisely the kind of subject where a model will confidently hand you the situation as it stood before December 2023, because that is what most of the written material still describes. So use it as a reading aid rather than a source: paste in the text from the RVO or the NVWA and have it turned into a checklist you verify yourself. For a bottle you hand to a friend nobody will care. The moment you sell, the rules apply.
What broke
I did it wrong on purpose to show how wrong it goes. One round with the text explicitly requested: a lino print of a pear, with “VONK & DOCHTER” in capitals underneath and the vintage 2024.
What came back was a tidy pear with the word VONKV under it, then the line ZOI SICUTR, then 2024. That last one was right, because digits go better than letters. The name I asked for no longer existed, and the second line was letter-shaped noise. From half a metre away it looks fine, which is exactly why people leave it in by accident.
The text also sat over the pear rather than under it. The model reads “underneath” as a compositional suggestion, not an instruction, and fills whichever space suits it. Ask for an empty area and the same thing happens in reverse: it reads “empty” as “empty of subject”, not “empty of ink”. Naming the proportion does help: make the image in the top two thirds and leave the bottom third as white paper.
The third problem is physical. A standard label is roughly 9 by 12 centimetres and a bottle is round, so anything important near the edges disappears from view once it is on glass. Image models think in rectangles that stay flat.
Ethics note: where the image comes from
I generate my images locally, on my own machine. Which also means there is no provenance marker in them. No C2PA manifest, no invisible watermark, nothing that lets anyone later see a machine made this.
The large services increasingly do embed that marker, and what such a marker proves is less than people think: it says the file came from that service, not that the content is true, and it does not survive a screenshot.
For a label I find the question simpler than the technical story. If the image goes on a bottle you are giving away, just say you made it with AI. Not because you have to, but because it changes how someone looks at it. I was once given a bottle with a drawing by the giver’s daughter on it. If I had thought it came out of a prompt, it would have been a different bottle.
And the other side: a small producer who has a model make their label and presents it as an artist’s work is selling a story that isn’t true. Wine leans heavily on origin and craft. That is exactly why it grates more here than on a blog image.
Do it yourself
Take a bottle you are giving away anyway. Generate an image with no text, leave a third empty, set the type yourself. Count on three rounds.
Put on it at minimum: what it is, who made it, the alcohol percentage, and the fact that it contains sulphites. That last one is the most frequently forgotten, and the only one that someone who reacts badly to it actually needs.