What photo restoration actually repairs
An old photograph degrades along several axes at once, and it helps to separate them. The print itself ages: colours fade or shift, contrast flattens, and the surface picks up dust, light scratches and handling marks. The capture was limited to begin with: consumer film and early digital cameras resolved far less detail than we are used to, so even a pristine 1970s snapshot is soft by modern standards. And the copy you are working from adds its own losses, because a scan or a phone photo of a print is a second-generation image.
Restoration addresses all three at once. A face-restoration model rebuilds structure in detected faces — eyes, lashes, the boundary between lips and skin — where the original has gone soft. A detail-enhancement pass works over the whole frame, removing grain and film noise while sharpening edges, and outputs at 2× the input dimensions so the recovered detail has somewhere to live.
The two run together as a single job rather than as stages you assemble. That is why the cost is a flat 3 credits: for the photographs this tool exists to serve, you almost always want both, and splitting them would mostly create a way to get a worse result for slightly less.
What it can and cannot fix
The honest boundary runs between degradation and absence. Where the original still contains a weak version of the information — a face that is soft rather than gone, a texture buried in grain, an edge that has lost its crispness — the model has something to work from, and the result is genuinely a recovery. Fine scratches, dust, film grain, fading and general age-softness all fall on this side.
Where the information is simply not there, the model must invent, and invention is where restoration stops being reliable. A crease that runs through a face has destroyed the pixels underneath it; a torn corner has no data at all; a large water stain has replaced the image rather than obscured it. In each case you will get a confident, plausible-looking fill that is a guess. Sometimes the guess is excellent. It is still a guess, and for a photograph of a real person that distinction matters.
For major damage, the realistic path is restoration first to clean up everything recoverable, then manual retouching in an editor for the specific area that is destroyed. The tool gets you most of the way in one pass and leaves a much easier job behind.
One more limit worth stating plainly: restoration does not colourise. A black and white photograph comes back as a sharper black and white photograph. Adding colour to a monochrome image is a guess about the world — what colour was that dress — rather than a guess about the picture, and it should be a deliberate choice rather than something that happens to your grandmother's wedding photo without being asked.
Getting a good scan in the first place
The single biggest factor in the final result is the quality of what you feed in, and it is the part entirely under your control. A model can only work with what reaches it, so twenty minutes spent on the scan pays back more than any setting on this page.
If you have a flatbed scanner, scan at 600 DPI. That is high enough to capture everything a consumer print holds and gives the detail pass real information to sharpen. Save as PNG or maximum-quality JPEG — a compressed scan bakes in artifacts that the model will then faithfully enlarge. Critically, turn off the scanner's own enhancement features: automatic dust removal, sharpening and colour restoration all pre-process the image in ways that conflict with what happens next, often smearing exactly the fine detail restoration wants.
If you are photographing prints with a phone, which is a perfectly reasonable approach for a shoebox of photos, the rules are about light and geometry. Use bright indirect daylight rather than flash, which reflects off the print surface and blows out a patch of the image. Shoot straight down rather than at an angle so the print is not keystoned. Fill the frame. And if the print is glossy, tilt it slightly to move reflections off the subject rather than lighting it more.
Photographs behind glass are worth removing from the frame first if you safely can. Glass adds reflections and a layer of dust that the model will treat as part of the image.
Restore, upscale or enhance: which one
The three AI repair tools overlap enough to be confusing, and picking the wrong one wastes credits on a worse result.
Choose restore when the source is old or damaged — faded prints, grainy scans, soft snapshots, anything where age is the problem. It repairs and enlarges at 2× in one 3-credit pass, and it is the only one of the three with a face-restoration model tuned for degraded sources.
Choose upscale when the image is clean but small. A modern photo that simply lacks the pixels for a print does not need repair; it needs resolution. Upscale costs 2 credits at 2× and 4 at 4×, and gives you the 4× option restore does not.
Choose enhance when the image is sharp and undamaged but badly lit — dark, flat, or colour-cast. The correction is free, which makes it the sensible first thing to try on any modern photo before spending credits elsewhere.
They stack in a sensible order for a difficult photograph: restore first to rebuild what age took, then enhance to correct the colour of the restored result. Doing it the other way round means correcting an image that is about to be substantially rewritten.
Restoring photographs of people, carefully
Family photographs are the main reason this tool exists, and they carry a consideration that a product shot does not: the output is a claim about what someone looked like.
Face restoration works by reconstructing plausible facial detail from a degraded source. When the source preserves most of the face, the reconstruction is well-constrained and the result is a genuinely better rendering of that person. When the face occupies very few pixels, the model has far more freedom, and it will still produce something sharp and convincing — because producing sharp, convincing faces is exactly what it was trained to do. It does not signal uncertainty.
The practical habit worth adopting is simple: keep the original, and compare. Look at the restored face beside the scan at the same size and ask whether the model recovered detail you can also find hints of in the original, or whether it added structure that was not there. For most photographs it is the former and the restoration is a gift. For the most damaged ones, treat the result as a beautiful interpretation rather than a document.
This is not a caveat unique to OpusImg. It is a property of every face-restoration model on the market, and anyone offering you a restored face without mentioning it is telling you less than they should.
How your photo is handled
Restoration runs on our servers, not in your browser. The models are far too large to ship to a web page, so the image is uploaded over HTTPS, processed on GPUs, and returned as a PNG. OpusImg labels every AI tool as server-processed for this reason, and the label is a real distinction — the client-side tools genuinely never transmit anything.
Restored images are not used to train models and are not shared. If a particular photograph is sensitive enough that you would rather it stayed on your device, the local tools remain fully available; that is a decision worth making per image rather than once for the whole account.
The cost is 3 credits per photograph, shown before the job starts, and failed jobs are not charged. Pro includes 500 credits a month and Business 2,000 per seat — enough for well over a hundred restorations a month, which is a larger shoebox than most people have.