What restoration cannot bring back

Every piece of software in this field is cheerful about what it can do and quiet about what it cannot, which leaves people spending evenings on photographs that were never going to come back and feeling that they did it wrong. They did not. Here are four kinds of loss that are final, why each one is final, and what is worth doing instead of trying.

Four kinds of loss that are final

Silvering. The blue-metallic sheen that comes up on the dark areas of an old silver-gelatin print is silver that has migrated to the surface of the emulsion. It is on top of the picture, it is reflective, and it changes with the angle of the light. A scan records a specular mess rather than the image beneath it, and there is no image beneath it in the file to recover. Photographing the print with polarised light, from an angle, gets you further than any software does.
Mould and emulsion loss. Mould digests the gelatin that holds the image. Where it has been active the picture has been eaten off the paper, and what remains is substrate. A fill can put plausible texture where the picture used to be, but it is inventing, and on a face it will invent a face that is not the person.
A print that stuck to glass. When a framed print is left in a damp room the emulsion bonds to the glazing, and separating them takes the image with it. What is left on the paper is a ghost. This is worth knowing before you try to free a stuck print: leave it in the frame, scan it through the glass at an angle that avoids reflections, and repair the scan.
A missing face. A fill invents plausible texture, not a person. If the tear crosses somebody’s features, no tool in this category will give them back, and one that claims to is generating somebody else’s face at the size and angle of the space. Cropping to what survives is more honest than filling it, and usually a better photograph.

Where this page stops short of its own ambitions

Two limits belong to this implementation rather than to physics, and both are worth stating in the same breath as the others.

Wide damage is drawn now, which is not the same as recovered

Tears, holes and missing corners want an exemplar patch search, and creases want an anisotropic fill along the fold. Neither is here. The marching fill that is here will let you mark them and will produce a smooth band where the damage was, and the taxonomy piece explains why that is a property of the method rather than a setting that can be tuned.

What has changed is that a generative model can now be downloaded on request and pointed at exactly those regions. It does not search the photograph for matching texture and it does not recover anything: it draws a continuation of what surrounds the hole, which is a guess of a more convincing kind than a smooth band. That makes it a better tool and not a different category. Everything else on this page still holds, because the thing this page is about is information that the paper no longer carries, and no model changes what a sheet of paper carries.

A file that will not open at all is a different problem

A truncated download, a half-written JPEG from a failing card, a PNG whose chunks are scrambled: those files never reach the point of being a picture, so nothing on this page can be applied to them. Recovering a broken file is byte-level work with its own engine, and it is not part of what ships here today. If a file opens and looks wrong, you are in the right place; if it does not open at all, you are not yet.

What to do with a print that is beyond repair

Scan it anyway, at 600 dpi, and keep the scan. Damage that is beyond today’s methods is not beyond every method, and the print will be worse in ten years while the file will be exactly as it is now. Write down who is in it while somebody still knows — an unrepairable photograph with names attached is worth more than a perfect one without.