About thirty years ago, I repaired scratches in this photograph for my mother after the frame's glass broke. Later, the Photoshop file became corrupted. When I discovered that a decade or so ago, I tried every repair tool I could find, short of the really expensive specialist services. Over the past few years, I tried AI tools too.

This started with two copies of a damaged Photoshop file in Google Drive. Codex retrieved them and confirmed they were byte-identical. A second copy wasn't going to supply the missing data, so the work proceeded on duplicates while the originals stayed untouched.

Fable 5.1 Max, running in Claude, handled the diagnosis and repair. Codex coordinated the retrieval and independently checked the result. That division matters: a repair script saying it succeeded is only the beginning of a check.

The file began with sixteen zero bytes before Photoshop's 8BPS signature. Removing that prefix put the header where a reader expected it, but the image data was still truncated. Fixing the first obvious defect did not fix the file.

Fable parsed the compressed rows across the red, green and blue channels. Of 12,507 rows, 12,506 were intact. The final row, in the blue channel, stopped sixteen values short at the bottom-right edge of the photograph.

That narrowed the repair considerably. There was no reason to invent a face, rebuild the scene or alter the rest of the photograph.

The code copied those sixteen missing blue values from the row directly above. These values are an approximation; they were not recovered from the damaged file.

It then updated the final compressed row's length from 2,778 to 2,779 bytes and appended seventeen encoded bytes. Sixteen missing color values and seventeen new encoded bytes describe different parts of the job: the image data and the way it is stored.

The result was a readable, flattened RGB Photoshop image at its original 2,820 by 4,169 pixels. The file's defects were established, but their presence does not prove exactly how the corruption happened.

ImageMagick, Pillow and psd-tools all decoded the repaired file to identical pixels. An independent comparison against the damaged source confirmed that all 35,269,724 surviving color-channel values were unchanged. All fifteen metadata resources were retained as well.

Those checks apply to the recovered PSD. The portrait shown here is my later edited version of it, so the published JPG and its web derivatives are not the untouched recovery output.

For me, this is a useful example of AI tackling an unfamiliar problem: inspect the actual data, narrow the failure, make the smallest justified repair and verify it separately. The outcome was also much more personal than the numbers suggest. I could see a photograph of my grandfather that I hadn't been able to open for almost twenty years.

Getting to see my grandfather in it again was pretty damn awesome.