Restoring a damaged archival photograph, and disclosing what the machine invented
The situation
A local historical society hands students a box of water-damaged 1920s prints for a small gallery show. The instinct is to run everything through a one-click restorer and print it. What comes back is smooth, plausible and partly fiction — a face that was never in the emulsion. The painful old way was twenty hours of clone-stamping; the painful new way is twenty seconds and a quiet lie on the gallery wall.
Steps
-
Scan once at capture resolution and freeze that file
Flatbed scanner, VueScan or SilverFast, 16-bit TIFF
Scan at the highest optical resolution the scanner actually supports, 16-bit, no sharpening, no dust removal, no auto-tone. Save it read-only and never touch it again. Every later claim about what is original depends on this file existing untouched.
What you only learn by doing it: Turn OFF infrared dust removal on the master scan. On silver-gelatin prints it reads scratches and emulsion loss as dust and silently removes real physical evidence — and there is no log, so you will never know what it took.
-
Map the damage on paper before opening any AI tool
A printout, a red pen, annotation layers
Mark the scan in three colours: information that survives but is degraded (recoverable), information that is physically gone (torn, missing emulsion), and ambiguous. This map governs the whole restoration and later becomes the disclosure graphic.
What you only learn by doing it: This is the step students skip and it is the entire workflow. Once you have run generative fill you can no longer tell which was which — the output is seamless by design. Viewers in the Heritage study rated AI-inpainted historical photos 4.3 out of 5 even when told they were synthetic.
-
Recover before you invent — tonal and grain work only
Work the recoverable areas with tone curves, local dodge and burn, and conservative denoise. Set denoise and upscale far below defaults — these tools are tuned for consumer snapshots, not for keeping period grain structure. Nothing generative happens here.
What you only learn by doing it: Denoise before upscaling, always, and repair after both. Upscaling first bakes scratches into higher-resolution scratches. And students consistently over-denoise 1920s prints into plastic — period grain is information about the paper and the process.
-
Reconstruct missing areas in separate labelled layers
Photoshop Generative Fill, or manual clone from within the same image
Each red-marked region gets its own named layer so synthetic pixels stay separable forever. Prefer cloning from elsewhere within the same photograph over generative fill wherever the material exists — that is recovery of real evidence rather than invention.
What you only learn by doing it: Never let generative fill touch a face, hands, text, or any identifying detail of a real person. Mason Resnick's demonstration removed Florence Owens Thompson's children from “Migrant Mother” in about ten minutes to prove how cheap the desecration is. Students find the results beautiful, which is the lesson.
-
Build the disclosure artifact that ships with the print
Photoshop for the overlay, InDesign for the wall label
Produce a second image at the same size with synthetic regions flagged in a contrasting overlay, plus a wall label naming every tool and every reconstructed region. Give viewers the option to see the unedited original.
What you only learn by doing it: Hang the disclosure at the same scale beside the restoration, not in a binder on a pedestal. The moment it becomes supplementary material nobody reads it and the ethical work evaporates. A QR code is a compromise, not a solution.
-
Soft-proof, test-strip, and review with the donor in the room
Soft proofing to the lab ICC profile, a 4x6 test strip
Print a test strip of the most contested region at final size and review the reconstruction decisions with whoever donated the photograph before the final print. At college level the donor family or historical society gets a veto.
What you only learn by doing it: Bring the test strip and the damage map to that meeting, not the finished restoration. Donors shown a beautiful finished print approve reflexively; donors shown “here is where we guessed, is that how the porch looked?” give you real information — and sometimes a second photograph that makes the guess unnecessary.
Where this breaks down
Consumer restoration tools fail on archival material in ways the marketing omits: face-restoration models regularise period-specific features toward a contemporary average, and several lighten skin tones, which turns a technical step into a representational problem you now own.
Rights are a live trap for an exhibition. The historical society may hold the print without holding copyright, orphan-work status is not a licence, and generative output terms cover the vendor's liability, not the sitter's.
If your disclosure is a colour overlay, a red/green flag is invisible to a substantial share of viewers — use shape or texture as well as hue and write the wall label to read aloud sensibly.
Be honest about the limit case: a photograph that is 40% gone is not a restoration project, it is an illustration project, and it should be labelled that way or not shown.
Provenance: the ethical framework — separating factual restoration from speculative reconstruction, labelling synthetic content, offering a highlighted-synthesis view, involving the source community — is taken from Radek Richtr, “Photorealistic Texture Contextual Fill-In,” Heritage 2025, which documents an AI inpainting project on demolished-city photographs with user testing. The technical ordering is standard conservation-imaging practice.