The hidden problem with photo restoration
Most restoration sites do one thing: they hand your photograph to an AI and ask it to “make it better.”
The result often looks better. But look closely and it's not the same photograph anymore. A grandmother's nose has narrowed. A child's eyes have closed when they were open. A car sits in a driveway that was empty in the original. The image is cleaner, sharper, prettier — and it belongs to someone who never existed.
The AI didn't restore your memory. It replaced it with a plausible one.
This is the single hardest problem in photo restoration, and almost no consumer tool solves it. We spent months failing at it before we found a way through.
What we tried first (and why it kept breaking)
Our first instinct was the obvious one: tell the AI, in painstaking detail, don't touch the faces. We wrote a brief — over a thousand words — locking down every facial landmark: eye shape, eyelid curvature, gaze direction, expression, the natural asymmetry that makes a person recognizably them.
It made things worse. The more rules we piled on, the more the model treated our photo as a suggestion rather than a source of truth. Instead of editing the pixels in front of it, it generated a brand-new image inspired by them. Faces were redrawn. Head angles shifted. A smudge in the background resolved itself into a car that was never there.
So we tried the opposite — a gentler model, lighter instructions. The faces survived, but the restoration itself went soft and dull, as if the AI had decided the safest thing to do was almost nothing.
We tried dedicated face-restoration engines next. They hallucinated too: closing a baby's open eyes, smoothing an expression into something plastic and unfamiliar.
Every approach collided with the same wall: a single pass cannot both transform a photograph aggressively and leave every face untouched. Asking one model to do both is asking it to hold two contradictory intentions at once.
The insight that changed everything
We stopped trying to negotiate with the AI about faces.
Instead, we split the job in two — and gave each pass a single, honest instruction.
Pass one is allowed to be bold. We let the restoration model do what it does best: strip the decades of yellowing and fading, reset the colour to clean daylight, repair the scratches and creases, recover the real greens of the grass and the true fabric of the clothing. We don't burden it with a thousand rules about identity, because we're not trusting it with identity at all.
Pass two puts the original back. Before any colour work begins, we locate every face in your original photograph — every one, including infants, children, and half-hidden faces. Then, after the restoration is complete, we take the exact original pixels of each face and seat them back over the restored image. We don't ask the AI to redraw them. We don't ask it to “remember” them. We take the real pixels — your grandmother's real nose, your child's real open eyes, the real expression the shutter caught — and colour-match them to the new frame so they carry the corrected tone without a single line of geometry being invented.
The seam is feathered so finely it disappears into the background. The face you get back is, pixel for pixel, the face that was photographed. Everything around it is restored as if the picture were taken today.
We study the photo before we touch it
Before any restoration begins, a separate vision pass reads the photograph and records what is actually there — as facts, not suggestions. Garment colours and materials. Headwear (is that a fur wrap, a scarf, a hat?). Each person's apparent age, ethnicity, skin tone, eye state, and expression. Any region too damaged to read confidently.
Those observations become binding ground truth. They are injected into the restoration prompt as things the model must not contradict — so it cannot quietly swap a leopard-print collar for a crown, flatten a two-tone garment into one colour, shift a person's ethnicity, or close an open eye. The facts act as a fence around the truth of the photograph.
You can add to them yourself. An optional “Notes about this photo” field on each upload lets you correct the analysis in plain language — “the headpiece is fur, not a crown,” “the baby's eyes are open,” “that's paint flaking, not a stain” — and your note overrides the automatic analysis whenever the two disagree.
Together with the two-pass face protection above, this is the layer that stops the hallucinations every other tool eventually produces. It's why our restorations cost a little more to run — and why the people in them stay themselves.
Why this is different
The standard approach treats the face and the photograph as one problem, and the face always loses — either softened into anonymity or reinvented into a stranger. We treat them as two separate problems with two separate solutions, and we refuse to let the AI hold the pen on identity at all.
The trade-off most tools force on you is “how much do you want us to change?” Crank it up and you risk a new face; dial it back and you risk a photo that still looks like it's been in a drawer for forty years. We removed that trade-off. The restoration can be as aggressive as it needs to be, because the faces are protected by a mechanism that doesn't depend on the AI's cooperation.
There's one more principle we hold to absolutely: the model is never allowed to invent. If a region of your photograph is too faded or too blurred to read, it stays soft. We will never resolve an ambiguous smudge into a car, a sign, or a building that wasn't there. A blurry shape stays a blurry shape. We repair what the pixels support, and we honestly leave alone what they don't.
What that means for you
When your restored photograph comes back, you should be able to look at it and recognise the person — not an AI's idea of them. The colour should feel like the day the photo was taken. The background should be the same background, the same people in the same places, the same car (or no car) that was actually there.
That restraint is the whole point. It's also what makes this slow, careful, and genuinely expensive to run — each photograph is processed through a multi-stage pipeline that we tuned by hand, one failure at a time, until the faces stopped drifting and the colour stopped lying.
That's the work your support keeps alive.