Repair a damaged photoBring old photos back to life

Damage on a print is not one problem. It is a handful of different physical events that happen to a sheet of paper and an emulsion, and they leave marks with different widths, different shapes and different amounts of surviving information underneath them. A speck of platen dust is two pixels across and the picture continues underneath it. A fold has crushed the emulsion along a line. A torn corner has no picture underneath it at all, because the corner is in a different envelope.

The distinction that matters is width. Anything narrow enough that the picture either side of it still agrees about what should be there can be closed by looking at the neighbours: that is dust, hairline scratches, cat hairs on the glass, the speckle of foxing, and most of what a scanner adds on the way in. Anything wide has to be invented, and the local engine on this page does not invent — it will tell you it has nothing to work from rather than guess. There is now a second path beside it, taken deliberately and never by default: a model you choose to download, knowing its exact size, that draws a plausible continuation into a hole instead of averaging one. The word for what it produces is still invented, and nothing here calls it anything else. What this page ships is the narrow half done properly, the wide half done by something that guesses, and an account of which is which.

Three things are worth saying plainly before you spend an evening on a photograph of your grandmother:

• Wide damage is guessed at, never recovered. The fill averages intact picture from within a few pixels of each marked one. Across a torn corner or a missing face there is no intact picture in the middle to average, so what comes back is a plausible smudge rather than a face. The optional model pass draws into those regions instead of averaging them, which is a better-looking answer and exactly as much of a guess: the paper recorded nothing there, so there is nothing for anything to recover. Where the gap held a face or a hand, treat the result as a reconstruction and say so if you pass the picture on.
• Silvering and mould are not damage to the image, they are damage to the emulsion. The metallic sheen that comes up on old silver-gelatin prints is silver that has migrated to the surface, and mould has eaten the layer that held the picture. In both cases the information is gone from the paper, not merely obscured on it, and no amount of arithmetic on a scan recovers something the scan never contained.
• No colour is added to a photograph that had none. Fade recovery stretches the range each channel actually holds. On a black-and-white print all three channels hold the same thing, so the result is a grey picture with its contrast back. Putting colour into a monochrome print is a different operation and it is described further down.

The order below is the order the engine runs in, and the same five functions run on screen and on save. One thing deliberately does differ, and it is worth knowing which. The preview works on a copy no more than about two and a half thousand pixels on the long edge, because that is what redraws at the speed of a slider. When you save, detection is measured again over every pixel of the original rather than the preview mask being scaled up — a scratch three pixels wide on a full scan is one pixel wide on the preview, and enlarging that mask would fill three times what the damage occupies. So the settings you approved are the settings that run, and the mask they produce at full resolution is a closer fit to the damage than the one you were looking at.

1. 1Open the scan and let it be read. Detection runs once, in a background thread, and produces a score for every pixel rather than a yes-or-no. On a twenty-megapixel flatbed scan that is a second or two. Everything afterwards is fast because of it.
2. 2Set sensitivity by watching the overlay. The slider is a threshold on a reading already taken, so the mask redraws as you drag. Turn it up until the damage you can see is covered, then back off until the picture stops being marked. The live region announces the region count each time the mask settles.
3. 3Correct the mask by hand. Two mask brushes: one marks, one unmarks. Beside them are a Heal brush for spots and blots, a Scratch tool that closes a straight line you drag, and a Clone stamp for anything neither reaches. The plus and minus keys resize. With the picture focused the arrow keys walk the brush a pixel at a time, shift takes ten, and Enter marks where it stands — the whole tool is reachable without a mouse.
4. 4Fill, and set the reach. Two pixels of reach is right for dust. Four suits a scratch with a shadow beside it. Past six the average stops being local and the fill starts to announce itself. Turning Fill off leaves the marks drawn over an untouched picture, which is the state to edit the mask in.
5. 5Pull the fade back. Fade recovery and density both preview live, because they are three lookup tables rather than a pass of the engine. The marked pixels are excluded when the endpoints are measured, so a white scratch cannot define the white point of the photograph. Black and white points, a colour-cast correction you can measure or take from something that was grey, and a black-and-white-safe mode that curves the three channels as one sit beside it.
6. 6Compare, then save. Holding the space bar puts the scan back as it arrived, with the mask hidden, for as long as you hold it; the Split view shows both at once with a divider you can drag. Saving re-runs detection at full resolution rather than scaling the preview's mask up, and writes PNG or JPG.

Forty steps of history are kept, listed beside the picture, and any state can be kept as a snapshot to come back to. Z steps back and Shift + Z steps forward; M hides and shows the mask; A runs the automatic pass; C splits the picture into scan and repair; K shows the loupe, which puts the scan and the repair side by side under the pointer at twice their size. History stores your brushwork run-length encoded rather than as flat planes, which is what makes forty steps affordable on a phone.

JPG, PNG, WebP, GIF, BMP and ICO decode here through the browser’s own image pipeline. HEIC is decoded by a WebAssembly codec fetched the first time one is opened and not before, so nobody pays for it who does not use it. TIFF matters more on this page than on most, because TIFF is what flatbed scanning software writes by default and a scan is the usual way a damaged print arrives. Only Safari reads TIFF natively, so this site carries its own TIFF reader, fetched from this domain the first time a TIFF is opened: uncompressed, LZW, Deflate and PackBits, 8- or 16-bit, greyscale or RGB, first page. A 16-bit scan is worked at 8 bits per channel, and a TIFF that uses something else (JPEG inside TIFF, CMYK, a palette) is refused with the name of what it uses.

Orientation is settled before anything else happens. A two-pixel probe JPEG with a known rotation flag is decoded with the browser’s own orientation handling switched off, which reveals whether this particular engine applies the EXIF flag by itself. The flag is then applied by hand only if the browser did not — the alternative is a scan that arrives rotated twice on some machines and once on others.

A CMYK JPEG — the kind a print shop’s workflow produces — gets no special handling on the way in, and that is worth knowing before you open one. Nothing here reads a file for its colour-space marker; what becomes of such a scan is whatever this browser’s own decoder does with it, which on some engines is a refusal and on others is a picture whose colours are wrong. If a scan opens looking inverted or unaccountably cold, save it again as RGB before repairing it, because otherwise the detector is reading a photograph nobody meant to make. Saving offers PNG or JPG. PNG is the better end for a picture you intend to keep working on, because a repaired JPG re-encoded as a JPG is a second generation of loss on top of whatever the scanner already cost you.

A damaged print usually arrives carrying several complaints at once, and only one of them is damage. The others are separate jobs with their own methods and their own ways of going wrong, and running them from inside a restorer is how a photograph ends up looking restored instead of looking like itself. This page closes the marks and stops there, on purpose.

None of the five is forbidden and none of them is hard. They are simply other tools, run at other moments. The sequence that works is: scan flat and honest, close the damage, and only then do whatever else you meant to do to the picture. Taken the other way round, every later operation is applied to the scratches as faithfully as it is applied to the photograph.

Four of these open this same workbench with different controls already in front, because a scratched print, a faded one, a creased one and a torn one want different things within reach first. Nothing about the engine changes between them. The fifth address is the reading: four written pieces behind one index. The two file-recovery routes are listed further up the page instead, because they answer a different sentence and belong beside each other rather than beside a photograph.

• Restore an old photographOpens on fade correction, for a print that has gone pale and warm.
• Scratches and dustOpens with detection turned up and the brush already out.
• Narrow print creasesThe local fill opened at its widest reach for a thin fold with intact picture on both sides.
• AI photo restorationOpens on the model panel, for the damage a local fill cannot close.
• Written guidesEight pieces on scanning, damage, colour, loss, broken files, storage and when to ask a person.

What the written pieces cover: which scanner settings decide whether a repair is possible at all, how print damage sorts by the amount of picture surviving underneath it, the point where recovering a measurement becomes inventing one, and the four kinds of loss that live in the paper rather than in the file.