Packaging artwork checks with AI before print: text, allergens, barcodes and languages

What an AI artwork check does before print

An AI artwork check compares the print-ready PDF of a carton, label or pouch with what it must say before the plates are made. A vision-language model, an AI model that reads images as well as text, reports where the artwork differs from the approved master text, where an allergen is not emphasised, where a mandatory statement or a language version is missing and what changed between two proof versions, while scripts check exact items such as the barcode number. The person who approves the proof receives a list of findings; the AI approves nothing itself.

This page is for prepress managers at packaging printers and label converters, and for packaging, regulatory and quality managers at brand owners. It describes what such a check compares, which part a script does better, what the model needs from your files and which hardware runs it in your own building, so that unreleased products and your customers' artwork stay with you.

Label text against the specification: what the check compares

Our summary; legal references from Regulation (EU) No 1169/2011 (consolidated text of 1 January 2018 on EUR-Lex) and GS1's check digit page, read 25 October 2026.

Most errors in artwork are differences: a word changed by the designer, an old ingredient list pasted into a new pack, an allergen in bold on the German panel and in plain type on the French one. The master text is therefore the most important input. It comes from your specification system, a product information database or the approved text document of the brand owner; without it, the model can only judge plausibility, which catches far fewer errors.

The PDF itself decides how much a script can do. If the text in the PDF is live text, a script extracts it exactly and compares it with the master text character by character. If the text has been converted to outlines, it is a drawing, and the model has to read it from the rendered page like a person would. The same holds for barcodes: a barcode library decodes the number from the rendered image, and GS1's method, alternating weights of 3 and 1 over the digits and the difference to the next multiple of ten, gives the check digit a script compares with the last digit.

The food information rules behind the check list

For food sold in the EU, Regulation (EU) No 1169/2011 on food information to consumers sets most of what the check list contains. This is general information; your regulatory or legal team confirms what applies to each product and market.

• Mandatory particulars. Article 9(1) lists twelve, from (a) "the name of the food" and (b) "the list of ingredients" to (l) a nutrition declaration, including allergens, net quantity, the date of minimum durability or "use by" date, storage conditions and the name and address of the responsible food business operator.
• Allergens. Annex II lists 14 substances or products causing allergies or intolerances. Article 21(1)(b) requires their names in the ingredient list to be emphasised "through a typeset that clearly distinguishes it from the rest of the list of ingredients, for example by means of the font, style or background colour".
• Legibility. Article 13(2) sets an x-height, the height of a lower-case "x", of at least 1.2 mm for the mandatory particulars; Article 13(3) lowers it to 0.9 mm for packages whose largest surface is under 80 cm². Article 13(1) requires the information to be in a conspicuous place and not hidden, obscured or interrupted by other matter.
• Languages. Article 15(1) requires the mandatory information in a language easily understood by consumers in the Member State where the food is marketed, and Article 15(2) lets Member States require one or more official EU languages on their territory.

Other product categories, such as cosmetics or detergents, fall under other EU rules, so each product category gets its own check list. Our guides to an AI server for food and beverage producers and for cosmetics manufacturers cover the specification side.

How the check runs, step by step

1. Render. Each page of the artwork PDF is rendered to an image at a fixed resolution, for example 300 dpi; the text layer, if there is one, is extracted at the same time.
2. Exact checks. Scripts compare the live text with the master text, decode the barcodes and check their digits, and mark the areas that changed against the last approved proof.
3. Model review. The model receives the page image, the master text and the check list for the product and market, and returns its findings in a fixed format: what, where on the page, and why.
4. Report. Script and model findings go into one report per artwork, with the page areas marked, for the person who approves the proof.
5. Decision. Prepress or regulatory accepts or rejects each finding; these decisions show over time where the check helps and where it raises false alarms.

The resolution matters for small text. At 300 dpi a 1.2 mm x-height is about 14 pixels, by our arithmetic. Qwen3.8-27B, an open vision-language model under Apache 2.0, turns every 32 × 32 pixel block into one image token, and its image processor accepts 65,536 to 16,777,216 pixels per image, so 64 to 16,384 tokens. A carton flat of 300 × 200 mm at 300 dpi has about 8.4 million pixels and comes to about 8,200 tokens. A large label sheet of 600 × 400 mm at the same resolution has about 33 million pixels and has to be cut into tiles first. Our guide to vision-language models for document AI explains image tokens in more detail.

What the model can judge, and what it cannot

The model is given tasks a proofreader does by eye: read the text, compare it with the master text, say whether an allergen is set in bold or underlined, and notice a missing panel or a statement in the wrong language. How reliably it does this on your artwork is what the pilot below measures. It does not measure millimetres, judge colour against a proof standard or grade the print quality of a barcode; those stay with the prepress tools and the measuring equipment the printer already uses. It can also miss an error or report one that is not there, so the check speeds up the proofreader's work and does not replace it.

Measure that in a pilot before the check runs on every job. Take artworks from the last months with errors you know about, run the check, and count what it found, what it missed and how many false alarms it raised per artwork; our guide to what to measure in a private AI pilot sets out the method. Test every language you print, since a model that reads German and French well can be weaker in Greek or Bulgarian.

Our Private AI/ML service sets up the model, the scripts and the report on your server; send us a few sample artworks with their master texts and the languages you print, and our team proposes a pilot.

Why the artwork stays in-house

Packaging artwork shows products before they reach the market: new recipes, new brands, promotions and new claims. A packaging printer holds the artwork of many brand owners, often under confidentiality terms in the supply contract, and a brand owner's packaging department holds launches that competitors must not see. Running the check on a server in your own building, or on dedicated hardware you control, keeps the files where they are today. Our article on trade secrets, NDAs and AI tools explains why that matters for contracts.

Hardware: one card in prepress or the company AI server

Artwork checking is a light load next to a company chatbot. A prepress department checks hundreds of pages a day, not tens of thousands, and each page waits a few seconds without anyone noticing. The table shows three set-ups; the first two cover most cases.

Model files on Hugging Face (Qwen3.8-27B-FP8 30.9 GB; NVIDIA's Qwen3.8-27B-NVFP4 21.9 GB), read 25 October 2026; pages by our users-per-card rule with an FP8 cache and 11,200 tokens per page; power by our estimate as on our workstation and server pages.

How we count pages at once. Our users-per-card rule takes 90 per cent of the memory the driver reports, less 3 GiB, less the model weights. On a 32 GB card that leaves 25.7 GiB, and NVIDIA's NVFP4 version of Qwen3.8-27B, published on 8 September 2026 under Apache 2.0, takes 20.4 GiB of it, leaving about 5.3 GiB. One artwork page with about 8,200 image tokens, about 2,000 tokens of master text and check list and about 1,000 tokens of findings needs about 0.48 GiB of cache with an FP8 cache, so about 10 pages run at once. On a 96 GB RTX PRO 6000 Server Edition, Qwen's FP8 version of 30.9 GB leaves 54.3 GiB, enough for about 112 pages.

How fast. StorageReview measured Qwen3.8-27B in NVFP4 on the RTX PRO 4500 Blackwell Server Edition at 153 output tokens per second with 32 requests of 8,192 input and 1,024 output tokens, with vLLM 0.27.1 (review of 14 October 2026). By our estimate that is in the order of 500 reports of 1,000 tokens an hour; the time to encode the page images comes on top and is not in that text-only figure, so measure your own artwork in the pilot.

Which set-up. A printer or brand owner that uses AI only for artwork can start with a workstation with one RTX PRO 4500 Blackwell (32 GB of GDDR7, 200 W), which needs no server room; our page to buy an RTX PRO 4500 workstation describes it. A company that already runs, or plans, an internal AI server for other departments puts the artwork check on it, with one RTX PRO 6000 Server Edition card for this and other document tasks; our guide to an AI server for small companies shows how one server serves several departments. The four-card RTX PRO 4500 Server Edition server in a rack suits several departments with smaller models each, as our page on Qwen3.8-27B on RTX PRO 4500 Server Edition works through.

Connecting it to the prepress workflow

The check fits where the proof is made today. A watched folder, or a step in the prepress workflow system that sends each new PDF version to the check service and receives the report, is usually enough. The master texts come from the specification or product information system through an export the IT team or the system's supplier sets up. The report goes back as a PDF with marked areas or as a list in the workflow system, whichever the proofreaders prefer.

Keep three things fixed from the start: one check list per product category and market, version numbers for the model and the check list in every report, and a record of each finding's decision. With them you can show a customer or an auditor how a proof was checked, and you can see whether a new model version finds more or raises more false alarms.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply workstations with RTX PRO 4500 Blackwell cards and AI servers with RTX PRO 4500 Server Edition or RTX PRO 6000 Server Edition cards, built to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the model, the check scripts and the connection to your workflow; our article on creative agencies covers the design side of artwork.

FAQ