A German online retailer has described how a generative image model silently altered legally required text on product photos, prompting a switch to a deterministic processing chain checked against a separate reference image.
Justin Flick of SOMUVA UG, which runs the German-language shop vapechamp.de in Lüdenscheid, Germany, set out the case. The task was to turn phone photos of carton back panels into clean, straightened images on a white background for online product pages. Those panels carry manufacturer and importer names and addresses, warning texts and contact details such as phone numbers. The images could look better, but the information must not change: the measurand was the printed text, not the appearance of the image.
A convincing result that did not hold up
The first attempt used a generative image model, given a straightened carton image and an explicit instruction to keep every printed character, number and symbol identical. The output looked professional — sharp print, even light, a clean background — but comparison with the source told a different story. The digits "2801" had become "2001"; a phone number contained wrong digits; a comma had turned into a full stop and elsewhere a full stop into a comma. The changes were plausible and hard to spot at a glance, so all generated images were discarded and none was published.
A generative model produces what text plausibly looks like, not what was printed, Flick concluded, and visual inspection is too weak a check to serve as the reference.
Classical processing, pixel by pixel
The replacement was a deterministic chain. The carton face is rectified with a perspective transform from four corner points; uneven lighting is flattened by estimating the background illumination and dividing it out, then white-balanced to a target tone. Denoising is gentle and edge-preserving, and text blocks are protected so their pixels pass through the background clean-up unchanged. Nothing is redrawn or re-typeset; every character derives from the photographed pixels, placed on a white background.
Reference and gate
For each image the team kept a reference: the same crop after rectification and the illumination and white-balance step only, before denoising or clean-up. Each processed carton face was compared with it, before placement on the background, using two measurement principles.
The first is an OCR diff: both images are read by an open-source OCR engine with German and English language data after upscaling, and the word sequences are aligned with every deviation listed. The second is an ink check that does not read text: pixels clearly darker than the carton background in the reference count as print, and if they turn light while most of a connected group disappears, the group is flagged as a lost character.
The gate also showed classical filters are not automatically safe: a speckle filter turned "130" into "30" and removed colons, and a threshold clean-up erased the dot in a web address. Both stages were dropped. Any step that decides what counts as noise can decide a full stop is noise.
OCR has its own measurement uncertainty
An OCR diff is itself a measurement system with its own error. On small print it misread the reference as well, for example reading a "5" as an "S", so a raw diff reports disagreements the processing did not cause. Taken at face value these produce false rejects; ignored wholesale they hide real changes and produce false accepts. Flick likened this to a gauge study, where the measurement system's own variation is separated out before variation is attributed to the part.
The independent ink check and a separate review step did that work. He linked the approach to Gleb Tsipursky's call to measure disagreements between automated inspection and qualified human judgment before reducing human review, and to Tsipursky's third step: identify the source of the disagreement rather than assuming the person or the model is wrong. Each disagreement should be a stop to resolve, he wrote, not a similarity score to average away.
Second review and the published file
A separate automated review, independent of the processing chain, checked the images in two rounds; a composite failed because one inserted line could not be verified reliably and was replaced with the genuine information block from the carton. Previous states were backed up before any replacement, and after publication the delivered files were compared with the team's own; at reduced resolution the mean absolute greyscale difference was well under one grey level.
Practical takeaways
- Define the measurand: for images carrying mandatory text, the characters matter, not the appearance.
- Keep a reference image from the earliest reliable processing step; never compare an output only with itself.
- Use two measurement principles, such as an OCR text diff and a pixel-level ink check.
- Characterise the OCR's own errors on the reference before trusting its verdicts.
- Treat any disagreement as a stop, not a score.
- Test classical filters as rigorously as AI tools.
- Add an independent second review and remove what cannot be verified.
- Back up before every change and pixel-compare the published files.
- Improve the input: frontal, unwrapped, frame-filling photos in daylight make later steps easier.
Generative models remain useful for images where nothing has to be read, Flick concluded; for images carrying mandatory information, the shop now uses classical processing plus a text comparison against the original.
