[{"data":1,"prerenderedAt":24},["ShallowReactive",2],{"blog-body-restore-old-photo-limits-en":3},{"content":4,"toc":5},"\u003Cp>Photo restoration is two operations wearing one name. Part of it recovers information that is genuinely present in the scan but hard to see. The other part invents information that is not there at all. The tool does not distinguish between them, and neither does the result — which is why knowing where the line falls is more important here than with almost any other image operation.\u003C\u002Fp>\n\n\u003Ch2 id=\"what-can-genuinely-be-recovered\">What can genuinely be recovered\u003C\u002Fh2>\n\n\u003Cp>These are cases where the information survived, degraded rather than destroyed, and a good restoration genuinely brings it back.\u003C\u002Fp>\n\n\u003Cul>\n  \u003Cli>\u003Cstrong>Fading and colour shift.\u003C\u002Fstrong> Old prints lose their dyes unevenly — cyan first, which is why decades-old colour photographs go yellow-orange. The luminance detail is usually intact, so correcting the colour casts recovers a natural-looking image rather than inventing one.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Dust, specks and light scratches.\u003C\u002Fstrong> Small defects sitting on top of intact underlying detail. Removing them restores what was photographed.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Noise and grain.\u003C\u002Fstrong> Reducible, particularly from a high-resolution scan where the grain is well sampled.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Mild overall softness and low contrast.\u003C\u002Fstrong> Correctable to a meaningful degree, especially when the scan retains detail that the print's tonal range compressed.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Tears and cracks with context on both sides.\u003C\u002Fstrong> Where the surrounding image indicates what the missing sliver contained, the repair is a reasonable inference rather than a fabrication.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Skew and perspective.\u003C\u002Fstrong> A photograph of a photograph taken at an angle can be squared up, recovering the correct geometry.\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Ch2 id=\"what-cannot-be-recovered\">What cannot be recovered\u003C\u002Fh2>\n\n\u003Cp>Five cases where the information is gone and any result is a fabrication, no matter how convincing.\u003C\u002Fp>\n\n\u003Cul>\n  \u003Cli>\u003Cstrong>A region that was never captured.\u003C\u002Fstrong> If someone was cropped out of the frame, or a corner of the print is missing entirely, there is nothing to recover. Anything that appears there was generated.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Highlights burned to pure white.\u003C\u002Fstrong> A blown-out window behind a face has no detail recorded in it. Tonal correction cannot bring back what the sensor or the film never registered — the information is not dark, it is absent.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Motion blur.\u003C\u002Fstrong> Where subject or camera movement smeared detail across many pixels, the blur is a many-to-one transformation. It cannot be inverted reliably, and a deconvolution pass will produce ringing artefacts rather than real detail.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Severe out-of-focus blur.\u003C\u002Fstrong> The same problem with a different cause. Fixing the focus is not a computational task; it is a reshoot that is usually impossible.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Heavy JPEG compression blocks.\u003C\u002Fstrong> Once an image has been compressed aggressively, the blocking and ringing are baked into the file. A restoration pass can smooth them, but smoothing is not recovery, and the original texture does not come back.\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Ch2 id=\"where-restoration-fabricates\">Where restoration fabricates\u003C\u002Fh2>\n\n\u003Cp>Modern restoration tools increasingly use learned models, which is what makes the convincing results possible — and also what makes the failure modes hard to see.\u003C\u002Fp>\n\n\u003Cp>Any region the model fills is a guess informed by training data, not by your photograph. Faces are the clearest case: a face at low resolution will be reconstructed into a plausible face, and it will not be the person's face in any measurable sense. Text and numbers get the same treatment, which is why a restored document should never be read as evidence. Logos and signage are reconstructed from the model's memory of such marks rather than the specific one in the frame.\u003C\u002Fp>\n\n\u003Cp>The practical dividing line is whether the content is statistically generic or specifically yours. Film grain, fabric weave and foliage are generic, and reconstruction of them is usually acceptable. A face, a signature, a licence plate or an invoice total is specific, and reconstruction of those is a change of meaning rather than a restoration.\u003C\u002Fp>\n\n\u003Ch2 id=\"restored-is-not-archival\">Restored is not archival\u003C\u002Fh2>\n\n\u003Cp>This is the point that gets lost, and it matters most for exactly the photographs people care about most.\u003C\u002Fp>\n\n\u003Cul>\n  \u003Cli>\u003Cstrong>Keep the original scan, untouched.\u003C\u002Fstrong> The restored version is a derived artefact. If the technique improves in three years — and it will — all the work can be redone from the scan. It cannot be redone from the restored file.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Label the restored copy.\u003C\u002Fstrong> A filename convention is enough. In ten years, nobody will remember which version was which, and a family photograph that has quietly acquired invented features is a small, real harm.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Do not use a restored image as evidence.\u003C\u002Fstrong> For insurance claims, legal matters, genealogy records or anything with a factual consequence, the unretouched scan is the document.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Be cautious with faces of people who cannot consent.\u003C\u002Fstrong> Making a deceased relative's face sharper by inventing features is a decision worth thinking about before making it.\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Ch2 id=\"scanning-well-matters-more-than-restoring-well\">Scanning well matters more than restoring well\u003C\u002Fh2>\n\n\u003Cp>Restoration has a ceiling set entirely by the scan, and a better scan raises it far more than a better tool does.\u003C\u002Fp>\n\n\u003Cul>\n  \u003Cli>\u003Cstrong>600 to 1200 dpi\u003C\u002Fstrong> for prints. Higher than that mostly captures paper texture; lower than 600 discards detail a restoration could have used.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>PNG or TIFF, not JPEG.\u003C\u002Fstrong> Compressing the scan before restoring it bakes in artefacts you spend the next step fighting.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Turn off automatic enhancement.\u003C\u002Fstrong> Scanner software will happily apply sharpening and colour correction that removes information irreversibly and before you have seen the original.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Keep the highest bit depth offered.\u003C\u002Fstrong> 16 bits per channel preserves more tonal information for the correction step than 8.\u003C\u002Fli>\n  \u003Cli>\u003Cstrong>Clean the glass and the print first.\u003C\u002Fstrong> Dust on the scanner platen is recorded as part of the image and then has to be removed computationally, badly.\u003C\u002Fli>\n\u003C\u002Ful>\n\n\u003Ch2 id=\"a-realistic-expectation\">A realistic expectation\u003C\u002Fh2>\n\n\u003Cp>A well-restored photograph from a good scan of a mildly faded print can look genuinely new. A severely damaged print, a badly out-of-focus shot, or a heavily compressed digital file will produce a result that looks cleaner and is less true — smooth, plausible, and quietly incorrect in the places that mattered. Both outcomes are useful, but only one of them should be mistaken for the original.\u003C\u002Fp>\n\n\u003Cp>The \u003Ca href=\"\u002Ftools\u002Fphoto-restore\">photo restoration tool\u003C\u002Fa> works on your device, which is the right property for family photographs that you may not want to send to an unknown server. If the scan itself is the limiting factor, the \u003Ca href=\"\u002Ftools\u002Fimage-upscaler\">upscaler\u003C\u002Fa> can enlarge a low-resolution scan — but read \u003Ca href=\"\u002Fblog\u002Fhow-does-image-upscaling-work\">what upscaling actually does\u003C\u002Fa> first, because the same fabrication problem applies at a larger scale, and it applies to faces most of all.\u003C\u002Fp>",[6,9,12,15,18,21],{"id":7,"text":8},"what-can-genuinely-be-recovered","What can genuinely be recovered",{"id":10,"text":11},"what-cannot-be-recovered","What cannot be recovered",{"id":13,"text":14},"where-restoration-fabricates","Where restoration fabricates",{"id":16,"text":17},"restored-is-not-archival","Restored is not archival",{"id":19,"text":20},"scanning-well-matters-more-than-restoring-well","Scanning well matters more than restoring well",{"id":22,"text":23},"a-realistic-expectation","A realistic expectation",1791511468599]