Model reference — Pictures
Face detection finds where the faces are in an image and returns a box around each one, with a count. It does not identify anybody — no name, no comparison against a database — and that distinction matters both technically and legally.
Because it identifies nobody, detection carries far lighter obligations than recognition, and it answers many questions on its own: how many people are in a room, whether anyone is present at all, whether a submitted photograph contains exactly one face before it is accepted. It is also the first stage of every other face task, so its accuracy sets the ceiling for recognition, verification and blurring alike. The hard cases are small faces in a wide shot, profiles, masks and heavy backlighting.
Models in this group take a single image as input: a photograph, a scan or a screenshot. 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 photograph at about 2 megapixels.
Vendor: InsightFace
What it does: finds faces reliably including small and partly hidden ones, and returns five facial landmark points that let the face be straightened for later use. The accuracy benchmark for detection.
| 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 (images/hour) | ≈ 20,000 | ≈ 80,000 | ≈ 220,000 |
Vendor: InsightFace
What it does: a faster detector with accuracy close to the best, designed to spend its computing effort where faces actually are. The usual production choice at volume.
| 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 (images/hour) | ≈ 60,000 | ≈ 240,000 | ≈ 700,000 |
Vendor: Shenzhen Institute of Advanced Technology
What it does: a very small detector that runs on almost anything, including a camera-side device with no GPU. The right answer for counting at the edge.
| 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 (images/hour) | ≈ 120,000 | ≈ 450,000 | ≈ 1,400,000 |
Vendor: Ultralytics
What it does: a general detector trained on faces, which is convenient when the same pipeline must also count people, vehicles or equipment — one model, several object types.
| 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 (images/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Google
What it does: finds faces and facial landmarks on device with no GPU, built for on-camera and in-browser use where the image should never leave the equipment.
| 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 (images/hour) | ≈ 100,000 | ≈ 380,000 | ≈ 1,100,000 |
The choice depends on face size in your images, how many faces per image, and whether detection runs on a server or on a small device. Our consultants measure detection rates on your own images, then recommend a model and the resolution it should run at.
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 detection / face count. 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.