A few years ago I scanned 214 photographs from a shoebox in my aunt’s attic. About a third had water stains creeping in from one corner. A quarter had creases running straight through someone’s face. The worst was a 1961 wedding portrait so faded that the bride’s dress and the tablecloth behind her had settled into the same flat grey.
Doing that by hand in Photoshop would have cost me a month of evenings. Running the batch through AI photo restoration tools took an afternoon, and the best results were genuinely startling. Cracks closed up. Faces came back. The dress separated from the tablecloth again.
The failures were just as instructive. A handful of portraits came back looking like wax figures, and one great-uncle’s nose had been quietly redesigned. That gap between startling and wrong is the whole story of this technology, so here’s how to stay on the right side of it.
What the software is actually doing
Restoration tools get sold as a single magic button, but they’re really running three different jobs, each trained on different data and each with its own way of failing.
Damage repair removes scratches, dust, tears and creases. It’s a close cousin of object removal: the model studies the pixels around the damage and reconstructs whatever probably belonged there. If you understand AI image inpainting and how it replaces parts of a photo, you already get the mechanic. The difference is that a restorer is filling a scratch two pixels wide, not deleting a stranger from a beach.
Enhancement handles tone, contrast, noise and softness. It’s the least risky of the three because it’s mostly redistributing information that’s already in the file.
Reconstruction is the ambitious one. When a face is blurred past recovery, the model doesn’t sharpen it. It substitutes a plausible face assembled from thousands of training examples. That’s where noses get redesigned.
The four kinds of damage, and how each one is handled
Cracks, tears and dust
This is the cleanest win in the whole field. Dust specks, hairline cracks and emulsion scratches sit on top of the image rather than inside it, so the surrounding pixels give the model plenty to work with. A scanner with a decent dust-removal setting handles most of it before the AI even runs.
Fading and colour casts
Old prints lose contrast unevenly. A 1970s colour photo often carries a magenta shift, while black-and-white prints from the 1940s go yellow along the edges. Automatic correction fixes the global cast in seconds. Fixing a photo that’s faded in one corner and not another takes manual masking, and no consumer tool does that well.
Grain, blur and low resolution
You can’t restore detail that was never captured, but you can rebuild a convincing surface. Scanning at 600 dpi minimum is the real lever here: a 4×6 print at 600 dpi lands around 2400 x 3600 pixels, which gives any model far more to work with than a 300 dpi pass. If your scan is already small, an AI image upscaler that adds pixels without adding detail will resize it, and the honest ones don’t pretend otherwise. For a fuller picture of what tone and noise tools really change, there’s a good breakdown of what an AI photo enhancer fixes versus what it invents.
Missing corners and cropped heads
Half a face at the edge of a torn print is the hardest case, because the software has to guess. Extending the frame outward is a separate skill, covered well in this guide to AI outpainting and extending a photo past its original border. Use it for a torn corner, not for a missing person.
What AI restores well, and where it starts inventing
- Reliable: dust, scratches, creases, scanner noise, lens vignetting, global colour casts, small tears.
- Usually fine: moderate blur on faces, gently restoring contrast in a faded print, filling a corner that’s been torn away.
- Risky: eyes in a badly degraded portrait, text on signs and gravestones, uniforms, jewellery, anything with fine repeating patterns.
- Don’t bother: a face smaller than about 64 pixels across. There’s nothing there to restore, only something to invent.
That third category is where people get burned. A model that repairs a blurry eye has to decide on a pupil shape, an iris colour and a catchlight position. It will decide confidently. It will also be making it up.
A workflow that gets clean results
Scan generously. 600 dpi for prints, 1200 dpi for anything smaller than a postcard. Save as TIFF, not JPEG, so the first round of compression doesn’t eat the detail you’re about to restore.
Work on a copy. Keep the untouched scan in a separate folder. This sounds obvious until you’ve spent a weekend improving a file and then decided you preferred the damaged version.
Do the boring fixes first. Rotate, crop, fix the colour cast, then let the AI handle damage. Every model performs better on a level, neutral-toned image.
Run one pass, not five. Stacking three enhancers on the same file produces the plastic look everyone recognises. If the first pass is decent, stop.
Check at 100%. Zoom in on hands, eyes and text. That’s where artefacts hide, and they’re invisible in a thumbnail.
Colourisation: the part people get wrong
Adding colour to a black-and-white portrait is technically impressive and historically unreliable. The model has no idea what colour your grandmother’s coat was; it picks something statistically common for the era and the apparent fabric.
Two habits help. First, ask an older relative before you trust the output, or dig out a written description. Second, keep the colourised file as a separate version and never overwrite the black-and-white original. A sepia-toned 1910 portrait colourised into a bright red dress looks like a costume drama rather than a document.
Restoration or rewriting history?
There’s a line worth drawing, and it moves depending on what the photo is for. Removing a coffee stain from a family snapshot is repair. Erasing a relative nobody likes is revision. Adding a person who wasn’t standing there, or smoothing someone’s face into a generic twenty-year-old, changes the record rather than recovering it.
For heirlooms and anything going to a historical society, apply the lightest touch that makes the image readable and note in the filename what you did: 1958-wedding-restored. For a birthday print you’re framing, go as far as you like. Just label it.
Where to start, and what it costs
Free options cover more than people expect. Google Photos and most phone galleries include a basic enhance that handles dust and contrast. MyHeritage runs a solid web-based restorer with a free monthly allowance, and it’s particularly good on faces from the 1900s to the 1950s.
Paid desktop tools like Topaz Photo AI and the current Photoshop build give you control over how aggressive the repair is, which matters once you’re dealing with more than a handful of images. Expect somewhere between $10 and $25 a month. For a one-off box of 200 photos, a month’s subscription is usually enough to get through the lot, so don’t sign an annual plan out of habit.
Archiving the results so they survive
The restored file is now the most fragile version you own, because it lives on a drive that will die. Save a full-resolution copy as TIFF alongside a compressed JPEG for sharing, and name both files with the date and subject so they’re findable in ten years.
Write down what you know while you still can. Who’s in the photo, roughly when it was taken, and which details were reconstructed rather than recovered. A short text file in the same folder costs nothing and doubles the value of the image to whoever opens that folder in 2050. The pixels are only half the restoration. The context is the other half, and no model can generate it for you.

