Restoring the family display sheet
The sheet carrying the only known portraits of all four had faded to a heavy magenta and was shot at an angle. Correcting both without inventing anything took some care about where the damage actually was.
The only known image showing all four of Thompson and Bethie's children together is a printed display sheet that has hung framed on a family wall. It is titled, in script, Bethie and Thompson's Four.
It has faded badly, and it has faded in a particular way that is worth understanding before touching it.
Take it out of the frame
The first attempt worked from a photograph of the sheet still in its frame. That photograph carried glass reflections, a band of specular glare across the top, and gilt corner brackets intruding at all four corners. Every one of those is a problem no amount of processing removes.
The sheet came out of the frame and was photographed again on a plain surface. That one decision improved the final result more than every processing step described below. If you are digitizing family material, take it out of the frame.
The cast is in the print, not the photograph
Photographed under room light, the sheet reads as almost entirely magenta. The instinct is to treat this as a white-balance problem in the phone camera and correct the whole image toward neutral.
That instinct is wrong here, and correcting on it produces a muddy green-grey sheet with purple faces still sitting in it.
The sheet is a reproduction. The portraits underneath are black-and-white photographs. The magenta is the print's dye layers failing at different rates — the cyan and yellow have gone, and the magenta is what survived. There is no color information left in the print to recover, because there was never any color in the original photographs to begin with.
So the correction is monochrome, not white balance
Once that is clear the method follows.
The image is converted to monochrome with a channel mix weighted heavily toward green. Under a magenta cast the red and blue channels are close to saturated and carry almost no tonal separation; green holds nearly all of the surviving detail.
That is followed by a percentile level stretch — mapping the 1st and 99.5th percentiles to black and white — and then a warm duotone, so the result sits as an archival print rather than a clinical grey scan.
Rectify to the printed rule, not the paper edge
The second photograph was taken by hand, so the sheet sits in it as a trapezoid rather than a rectangle. Correcting that is a standard perspective transform, and the only real question is what to correct to.
The obvious target is the edge of the paper. That turns out to be wrong. On this sheet the printed black rule is not quite parallel to the trimmed paper edge — about one degree out along the bottom. Rectifying to the paper leaves the rule visibly tilted, and cropping to the rule afterward clips the corner of Aubrey Lee's portrait.
So the rule is the target. Its four edges are located by sampling the darkest pixel across a narrow band at ninety points along each side, fitting a line through those samples, and intersecting adjacent lines to get the corners. Fitting a line rather than hunting for corners directly matters: a corner is one pixel and can be wrong, while a line through ninety samples averages out dust, a crease, and the odd bad reading.
Those corners are then mapped back through the first transform into the original photograph's coordinates, so the image is resampled exactly once rather than twice. Two resamples of a fading print is one more than it deserves.
What was not done
Nothing was added, removed, reconstructed, or retouched.
No detail was painted in. No face was sharpened into something the print does not contain. No AI upscaling was used at any point. Anna Beth's signature across the corner of her portrait is exactly where she wrote it.
The restored plates are published beside a note saying what was done to them. An archive that silently improves its own holdings is not an archive.
The pipeline is in the repository
The processing is a single script, committed alongside the website source, so the derivation from original to published image is reproducible by anyone who has the originals. It reads the untouched camera files and writes the web assets. The originals are never modified in place.
That matters more than it might appear. In twenty years the question "is this the real photograph or a cleaned-up one?" will have a definite answer, and the answer will be checkable.