Does face recognition work on old, grainy VHS?
It's a fair thing to ask before you hand over a shelf of tapes: can a computer really pick out a face on footage that soft — a relative across the room from a camcorder, on a VHS dub that was already a copy of a copy? The honest answer isn't a clean yes or no. It's two different answers to two different questions, and keeping them apart is the whole thing.
Finding where a face is in the frame works, surprisingly well, even on a fifty-pixel VHS face. Deciding who that face is, by machine, does not — not on footage this grainy. That gap is the reason Rediscover works the way it does, and it's worth walking through, because it also answers the quieter worries underneath the question: whether this is a little creepy, and whether your family's footage is safe.
Two questions hiding inside one
When people ask whether face recognition works on old video, they're usually asking two things at once without meaning to. One is detection: can the software tell there's a face here, in this corner of this frame, and mark it. The other is identification: can it tell you that face belongs to Aunt Carol and not her sister. They sound like a single problem. They aren't, and they don't hold up equally well on tape.
Detection is a question about shape — is there something with the geometry of a face in the picture. Identification is a question about a specific person — matching this smear of pixels to that smear of pixels, and being right. On a sharp phone photo the two feel like the same easy task. On a much-copied VHS frame, one stays reliable and the other quietly falls apart. Almost every worry about face recognition on grainy footage comes from treating them as one thing.
Detection: finding a face works, even at fifty pixels
Here's the part that surprises people. Finding where faces are works remarkably well, even on bad footage. A face has a stubborn geometry — two eyes, a nose, the spacing between them — that survives a lot of grain, a lot of motion blur, a lot of generational copying. Software trained to spot that pattern will flag a face across a room, half-turned, in the corner of a frame you'd have scrubbed right past.
On the one real archive Rediscover was built on — 105 tapes, 134 hours, footage reaching back to the 1980s — that detection pass found 13,054 faces. Not names. Faces: every place a person appears, marked and ready. That number is the honest yes half of the answer. The software is genuinely good at the where. It's the who where you have to be careful.
Identification: why the machine shouldn't get the last word
Deciding who a detected face belongs to is a different animal, and on grainy VHS it's where confidence outruns accuracy. A fifty-pixel face doesn't carry enough detail to tell two relatives apart the way you can — and the machine won't hedge. It returns a name with a number beside it that looks like certainty, attaches the wrong one, and never blinks.
We know because we tried. An early version of this leaned on automatic face-matching to guess identities, and it failed — plainly enough that a $24 receipt for the experiment is still around as a reminder. The trouble isn't just one mislabeled clip. Because a search like this learns each person from the faces tied to their name, a single wrong guess teaches it to confuse two people from then on — and the risk is highest exactly where families feel it most, with cousins and siblings who already look alike. Auto-matching doesn't merely get one call wrong; it can quietly braid two people together for good.
So a person says the name
That failure is the reason Rediscover does it the other way around. The software does the tireless part — finding every face across every tape — and then stops, without ever deciding who anyone is. A person does the naming: you see someone you know on screen, put the cursor on their face, and press a key. That's the whole gesture. The name is right because someone who was there said so, not because a model rolled the dice and called it confidence.
It's less work than it sounds, and it compounds. You name the handful of people who matter, and from each confirmation the search fans out — surfacing that same face at other ages, other angles, through other cameras for you to confirm — and gathers the ones you keep onto one page, one best clip per tape. When two look-alikes do start to drift together in the math, the system flags it and lets you split them in a single pass, so no one's face is ever quietly learned from someone else's. The machine finds; the human decides. That order is the difference between a search you trust and one that lies to you with a straight face.
The quieter worry: is this creepy, and is it safe?
Underneath "does it work" there's usually a second question people don't quite say out loud: isn't this a little creepy, and where does my family's footage end up. Both are fair, and the answer to the first is the same as everything above. Nothing here is a face-scanning system passing judgment on your relatives. It's a tool that finds faces and then waits for a person to say who they are — closer to a very patient index than to surveillance.
And the footage stays yours, sealed. Each family's archive sits in its own environment, never mixed with anyone else's and never used to train anything — not our tools, not anyone's. Restoration, for a clip too grainy to bear on the big screen, sharpens what's actually there; it never paints in a face that wasn't. You keep every original master, and you can walk away with all of it, any time. The whole thing is built so the answer to "is this safe" is as plain as the answer to "does it work."
Common questions
Does face recognition work on old, grainy video?
It depends which half you mean. Detection — finding where a face is in the frame — works surprisingly well, even on a fifty-pixel VHS face; on the real archive we were built on it marked 13,054 of them. Identification by machine — deciding who that face is — is unreliable on footage that grainy, which is why a person confirms every name here rather than a model guessing it.
Can it tell two look-alike relatives apart on VHS?
A machine, no — not reliably. Two cousins on a grainy frame can sit a hair apart in the math, and an automatic guess will confidently pick wrong and then learn the mistake. So a person who knows the family confirms who's who, and if two people start to blur together the system flags it and lets you split them apart in one pass.
Is my family's footage used to train AI?
No. Each family's archive sits in its own sealed environment, never mixed with anyone else's and never used to train anything. You keep every original master and can leave with all of it, any time.
Does it ever invent a face to fill in the grain?
No. Detection only marks faces that are already in the frame, and identities come from a person, not a guess. Restoration sharpens and denoises what the tape actually recorded — it never paints in a face that wasn't in the footage.