One-to-one verification answers a single question: are these two pictures the same person? It compares a photograph taken now against one reference image — an identity document, a personnel record, a photograph taken at enrolment — and returns a similarity score with a decision.
It is a materially easier and safer problem than searching a database of faces, because there is only one comparison to make, so a false match cannot come from an unlucky resemblance elsewhere in a population. The decision threshold sets the balance between wrongly rejecting the right person and wrongly accepting the wrong one, and it is a policy decision that we set with you and then measure. Any deployment used for access also needs a liveness check, because a photograph held up to a camera will otherwise pass.
Models in this group take two images as input and return a similarity score between the faces in them. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one captured photograph compared against one reference image.
ArcFace (InsightFace)
Vendor: InsightFace
What it does: turns a face into a numeric fingerprint whose distance from another fingerprint measures whether they are the same person. The most widely deployed and best understood option, with published accuracy on standard tests.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (verifications/hour)
≈ 40,000
≈ 160,000
≈ 450,000
AdaFace
Vendor: Michigan State University
What it does: a newer fingerprinting model that handles poor-quality captures — blurred, low-light, low-resolution — considerably better, which is what most real verification failures come down to.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (verifications/hour)
≈ 36,000
≈ 140,000
≈ 400,000
FaceNet
Vendor: Google
What it does: the long-established model of this kind. Less accurate than the newer options but exceptionally well documented, which matters where a deployment must be explained to an auditor.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (verifications/hour)
≈ 50,000
≈ 190,000
≈ 520,000
RetinaFace
Vendor: InsightFace
What it does: finds and straightens the face before comparison, which is not optional: an unaligned face costs more accuracy than the difference between any two fingerprinting models.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (verifications/hour)
≈ 20,000
≈ 80,000
≈ 220,000
SCRFD
Vendor: InsightFace
What it does: the faster detection and alignment option, used where verification runs on every capture at a gate or kiosk rather than on request.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
4 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
8 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (verifications/hour)
≈ 60,000
≈ 240,000
≈ 700,000
Choosing between them
The right configuration depends on your reference image quality, the capture conditions and the relative cost of the two kinds of error. Our consultants measure both error rates on your own images, then recommend a model, a threshold, and the liveness checks the use case requires.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Face verification (1:1). We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.