Generative Image Cleanup Rewrote Mandatory Label Text, Retailer Reports

A German online shop discarded AI-generated product images after digits and punctuation on legally required carton text changed, and rebuilt the workflow around a deterministic chain checked against an independent reference.

A German online retailer that set out to clean up phone photos of carton back panels for its product pages found that a generative image model had silently rewritten the mandatory information printed on them, and replaced it with a deterministic processing chain checked against an independent reference. Justin Flick, writing for SOMUVA UG, the company behind the German-language shop vapechamp.de, described the case as a small one outside the factory that nevertheless reproduces a principle from industrial inspection: an automated result is only as trustworthy as the independent reference it has been checked against.

The task was narrow. Photographs of carton back panels, which carry legally required manufacturer and importer names and addresses, warning texts and contact details such as phone numbers, were to become straight, evenly lit images on a white background. Appearance could improve; the information could not change. In measurement terms, the printed text was the measurand, not the look of the image.

A plausible result that did not hold up

The first attempt used a generative model, given a rectified carton image and an explicit instruction to keep every character, digit and symbol identical. The output looked professional. Compared with the source, however, the digits "2801" had become "2001", a phone number contained wrong digits, and commas and full stops had swapped places. Because the substitutions were plausible, they were hard to catch by eye. For mandatory information the changes were not cosmetic, so the generated images were discarded and none was published. The lesson drawn was that a generative model produces what text plausibly looks like, not what was printed, which is why visual inspection cannot serve as the reference.

Deterministic processing and two checks

The replacement is a deterministic chain. Each carton face is rectified by a perspective transform from four corner points; uneven lighting is flattened by estimating and dividing out the background illumination, followed by white balancing to a target tone; denoising is gentle and edge-preserving. Text blocks are marked as protected regions whose pixels pass through the clean-up unchanged, and the result is placed on a white background with a soft shadow. Nothing is redrawn or re-typeset, and every character derives from photographed pixels.

Each image is then compared with a reference, meaning the same crop after rectification and the illumination and white-balance step only, before any denoising, using two independent principles. The first is an OCR diff: both images are read by an open-source engine with German and English language data after upscaling, and the resulting word sequences are aligned with every deviation listed. The second is an ink check that does not read text at all: pixels clearly darker than the carton background in the reference count as print, and a connected group that turns light in the processed image is flagged as a lost character.

The gate also caught classical filters. A speckle filter turned "130" into "30" and removed colons, and a threshold clean-up step erased the dot in a web address. Both stages were dropped, on the argument that any step which decides what counts as noise can decide that a full stop is noise.

When the checker itself errs

OCR is itself a measurement system with its own error. On small print it misread the reference, for example taking a "5" for an "S", so a raw diff reports disagreements the processing did not cause: false rejects if taken at face value, false accepts if ignored wholesale. Following the logic of a gauge study, the team separated the OCR's own variation from changes introduced by processing, using the ink check and a separate review step. Each disagreement is treated as a stop to be resolved rather than a similarity score to be averaged away, an approach the author links to Gleb Tsipursky's argument for measuring disagreements before scaling automated inspection.

Publication checks and takeaways

An automated review independent of the processing chain checked the images in two rounds; one composite failed because an inserted line could not be verified, and it was replaced with the genuine information block from the carton. Previous states were backed up before any replacement. After publication, the delivered files were downloaded and compared with the team's own, showing a mean absolute greyscale difference well under one grey level at reduced resolution.

The stated takeaways for quality teams are to define the measurand, keep a separate reference from the earliest reliable step, use two measurement principles, characterise the OCR's own errors before trusting its verdicts, test classical filters as rigorously as AI tools, add an independent second review, back up and pixel-compare published files, and improve inputs with frontal, unwrapped, frame-filling daylight photos. Generative models, the author adds, remain useful where nothing has to be read.

CO

Chidi Okonkwo

Chidi Okonkwo covers factory automation and motion systems, including industrial robot cells, PLC-controlled assembly lines, servo drives, and machine vision inspection stations. His coverage works from measurable parameters such as cycle time, positioning repeatability, payload and reach margins, drive sizing calculations, and safety-related control boundaries rather than unverified throughput claims. He writes for automation engineers, manufacturing engineering directors, and capital procurement managers who must compare cell designs, specify interfaces, and accept a line against quantified commissioning results.