A small product-imaging project at SOMUVA UG illustrates a risk in retouching photos that carry legally required text. Justin Flick, who writes for the German online shop vapechamp.de, explains that the team wanted to turn phone photos of carton back panels into straightened, white-background images for product pages. The panels showed manufacturer and importer names and addresses, warning statements and phone numbers. The images could be improved, but the printed information had to remain exact; in measurement terms, the text was the quantity of interest, not the picture’s appearance.
The first route used a generative image model. The team supplied a straightened carton image and instructed the model to preserve every printed character, number and symbol. The result looked professional, with crisp print, even light and a clean background. Comparison with the source revealed changes: “2801” became “2001,” a phone number gained wrong digits, and commas and full stops were exchanged. These errors were plausible enough to escape a quick look. The team discarded the outputs and published none. Flick’s lesson is that such a model renders text that looks plausible rather than reproducing what was printed, and that human visual checking is a weak control against plausible substitutions.
A deterministic pipeline with a protected reference
The team replaced the model with a deterministic sequence. A perspective transform from four corner points rectifies the carton face. Background illumination is estimated and divided out to reduce uneven lighting, followed by white balancing to a target tone. Denoising is kept gentle and edge-preserving. Text areas are marked as protected regions, so their pixels pass through background cleanup unchanged. The face is then placed on a white background with a soft shadow. No character is redrawn or re-typeset; the final printed panel is derived from photographed pixels.
Each image also had a reference: the same crop after rectification and illumination/white-balance only, before denoising or cleanup. Processed carton faces were compared with that reference before background placement. The team used two independent principles. One was an OCR diff: both images were read by an open-source OCR engine with German and English data after upscaling, and the word sequences were aligned to list deviations. The other was an ink check that did not rely on reading. Pixels clearly darker than the carton background counted as print; if such a connected group became light in the processed image and largely disappeared, it was flagged as a possible lost character.
Classical filters also failed the gate
The checks caught problems in conventional processing as well. A speckle filter changed “130” to “30” and removed colons, while a threshold cleanup erased the dot in a web address. Both stages were removed. Flick observes that a rule for deciding what is noise can also misclassify punctuation.
OCR introduced its own error. On small print, it misread the reference, for example taking a “5” for an “S.” A raw OCR diff therefore reports mismatches that processing did not create. Accepting those at face value produces false rejects; dismissing them wholesale hides real changes and produces false accepts. The ink check and a separate review helped distinguish OCR error from processing error. Flick links this to Tsipursky’s point that the source of a disagreement should be identified rather than assuming the person or model is wrong. Each mismatch was handled as a blocker requiring resolution, not as a similarity score to be averaged away.
Independent review and post-publication checks
A separate automated review, independent of the processing chain, examined the images in two rounds. In the first, a composite image failed because one inserted line could not be verified reliably. It was removed, and the genuine information block from the carton was used. Previous versions were backed up before replacement. After publication, the delivered files were downloaded and compared with the team’s own copies; at reduced resolution, the mean absolute greyscale difference was well under one grey level.
The practical lessons include defining the text as the thing that must not change, keeping a separate reference from the earliest reliable stage, and using two measurement principles. Flick also advises characterising OCR errors on the reference before trusting its findings, applying the same rigor to classical filters as to AI tools, adding independent review, and comparing published files pixel by pixel. Generative models remain useful where no text needs to be read; for mandatory information, the team now relies on classical processing plus a text comparison with the original.
