0 photos
is all it takesFrom roughly six clear images of your face, an open-source model can re-identify you on public camera footage with about 70% confidence — no special hardware required.
Open facial-recognition research
NanoFID
An open lab for seeing how facial recognition actually works — and where it quietly fails.
scroll
Identifiable.
Your face isn't a password — it's a permanent key you can never change. And the math to read it is already open source.
0 numbers
that's your faceRecognition models don't keep a photo. They compress your face into a vector of 512 floating-point values — a fingerprint that fits on a single line of text.
10–0×
uneven error ratesGovernment testing has found many algorithms misidentify some demographic groups up to 10–100 times more often than others. The bias lives in the data.
0B+
faces scrapedAt least one company has built a searchable index of more than 30 billion facial images pulled from the open web — most people in it never agreed to be there.
0
times you can reset itLeak a password and you change it in seconds. Leak your faceprint and it's compromised for the rest of your life. You only get one face.
Real-time
no consent neededNo tap, no badge, no opt-in. Modern systems pick a face out of a moving crowd, from across a street, the moment it appears on camera.
Train it on yourself. Then watch it fail.
NanoFID is a research bench, not a product pitch. You are both the subject and the auditor.
01
Sit for portraits
Upload a handful of clear photos of your own face. Six is enough to begin.
02
Watch it compress you
The model turns those pixels into a 512-number embedding — the same primitive production FaceID systems use.
03
Read the failure
Live tests and bias metrics show where the model is confident, brittle, or quietly unfair.