Synthetic biometric data and biometric models
Synthetic humans are indistinguishable from real. TessLabs builds the synthetic biometric data that defeats frontier AI models, and the biometric detection stack that catches what is missed. Data and models built on eight years of live fraud detection experience. We provide the biometric stack for hyper-scalers, frontier AI models, sovereign identity programs, and defence.
1.1M
synthetic identities, each from a fixed, repeatable seed
23
independently controlled attributes per identity
1013
latent state space, infinite variations on demand
3.74
clean-FID realism score,
as good as real
Origin
Training a production biometric model on real customer faces — at the scale and adversarial diversity that production models need — creates consent, privacy, and regulatory exposure most teams can't clear. TessLabs was built to solve that problem. TessLabs' synthetic-face generator was developed to train a live identity-verification business. The purpose: to provide our identity fraud models with safe and realistic biometric data to train against.
The data and models have been hardened against real fraud over eight years and c.25 million real identity checks. The synthetic faces TessLabs now builds are so good that frontier models cannot distinguish them from real. That's why we decided to make our data and models available to help train and close those gaps.
Read the full story on AboutWhat we offer
Synthetic identities from a fixed, repeatable seed, for training and stress-testing your own vision models. Scoped to your population, your failure modes, and the specific edge cases your model is missing. No fixed dataset. Every run is quoted against what you're actually trying to fix.
See how a run is scopedFour production models: verification, identification, liveness, deepfake detection, age, and gender. Built and hardened over eight years of live fraud. Licensed to run on your own infrastructure. No capture ever leaves your environment.
See the full model spec sheetWho this is for
Verify who's signing up, and catch the synthetic identities your liveness checks let through.
Generate synthetic identity data and harden your models against it, at the volume your compute already serves.
Ship a checksummed manifest with every delivery, so auditors verify the data instead of a description of it.
Score attack presentations at generation time, so your liveness model tunes to the exact mix your network sees.
Specify pose, occlusion, lighting, and age as dials, so hard captures are covered by design, not by chance.
Train your detector against identities that pass six flagship frontier models 94 to 100% of the time.
Calibrate thresholds to the population your programme serves, and rehearse deduplication before a citizen enrolls.
Red-team your systems against synthetics with no real face at risk, and keep the pool exclusive to you.
The data lab
Most synthetic-face generators create a new, unrepeatable image each time. TessLabs is different: 1.1 million unique, mathematically locked seed identities that generate near-infinite edge-case variations. Same maths, same face, down to the pixel, every time.
That's why we generate the precise edge-case data to train your models. Each seed carries 23 independent attribute dials — age, expression, headwear, skin tone, lighting etc— each one independently adjustable, and every output hashed to its original seed.
Our synthetics pass 12 core quality benchmarks covering realism, diversity, and structural correctness. The headline number: a clean-FID realism score of 3.74, and the acid test: Multiple tier 1 vision-language models classified our synthetic identities as real 94% - 100% of the time.
yaw −31° · pitch 0°
render only
yaw 0° · pitch −22°
box + 5 landmarks
yaw 0° · pitch 0°
mesh solving
yaw 0° · pitch +18°
mesh complete
face_085· synthetic: true
seed_locked · region: menat
How it works
01
No cost, no commitment. A small batch, scored against your own detectors.
02
The first production-shaped run, against thresholds you agree before generation starts.
03
Reweighted to exactly what the baseline exposed.
04
New identities every quarter, aimed at whatever moved.
Biometric models
1:1 face verification
Confirms that two faces belong to the same person — a selfie against a document photo at onboarding, a login against an enrolment. Calibrated on MENAT-heavy evaluation sets, so thresholds hold on the populations most vendors underserve.
1:N face identification
Searches one face against an enrolled gallery and returns ranked candidates — watchlist screening, duplicate-account detection, access control. Galleries to 10M identities; larger deployments scoped individually.
Passive presentation-attack detection
Scores whether the face in front of the camera is a live person or a presentation attack — print, screen replay, mask — from the capture itself, with no challenge gestures.
Deepfake & synthetic-face detection
Separates generated faces from camera-captured ones — deepfakes, face swaps and fully synthetic images. This is the discriminator side of the foundry, and the reason we can state that our own synthetics defeat frontier models while this engine still separates them.
Age estimation
Estimates age from a single capture, reported against the product's own bins — 18–24, 25–34, 35–44, 45–59, 60+ — so a result maps onto the same bands the training data is built in.
Gender classification
Classifies apparent gender from a single capture, evaluated per demographic group rather than on a pooled average, because a pooled number hides exactly the failures that matter.
Questions
Start with a sample you score yourself. Nothing's charged and nothing's reserved by asking.
Book a call
