A fold or crease can cross a section of handwritten or printed text in a scanned photo. The letters beside it are trustworthy where they remain visible. An editing model can bridge the line convincingly while guessing the words under it. Treat the retouched photo and the textual record as different outputs.
Identify enough of the original
Consider a photograph with a handwritten date on the bottom margin. A diagonal fold interrupts a digit. Save the original scan before editing and inspect it under another lighting setup if you can still access the print. Other marks can make one digit more plausible than another, but checking a related caption or original envelope is stronger evidence than interpreting a generated reconstruction.
Keep the repair local
Mask or crop around the crease when your editing workflow supports it, protecting unaffected letters. Run one version where you repair the visible paper texture and leave any ambiguous writing alone. Compare accepted output to the original at the crease and beyond it. A smooth repair that changes the slant of the nearby script, introduces a clean new digit, or replaces a shadow with flat color has gone too far.
File the known details separately
Use an image-reference workflow in the AI image generator only for the visual cleanup. Save a separate note of what you can read from the source photograph and which digit or character remains uncertain. If the text is historical evidence, publish the original alongside the restoration when possible. Readers can then see where the image is trustworthy rather than relying on a guess at missing writing.



