Model reference — Pictures
Image classification assigns a whole image to one of your categories — which product it shows, which of four damage grades applies, whether a photograph is acceptable for a claim, which document type a scan is.
It is the simplest computer vision task and the one most often the right answer. Where a business asks "which of these is it?", classification is cheaper, faster and more accurate than detection or segmentation, which answer harder questions nobody asked. The work in a classification project is almost entirely in the category definitions and the labelled examples: categories that overlap or that people label inconsistently will limit accuracy no matter which model is used, which is why we measure human agreement on your categories before choosing one.
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: Google
What it does: a small, quick classifier that is accurate on most business tasks and cheap to retrain as categories change. The default starting point.
| 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: Meta
What it does: a stronger classifier for visually similar categories — distinguishing near-identical parts or damage grades — at moderate cost.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 12,000 | ≈ 48,000 | ≈ 140,000 |
Vendor: Google
What it does: a transformer-based classifier that performs best when a large labelled set is available, and provides features reusable for search and duplicate detection.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 10,000 | ≈ 40,000 | ≈ 110,000 |
Vendor: Microsoft
What it does: accurate on images where fine detail decides the category and the image must be examined at high resolution, such as textiles or printed circuit boards.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 8,000 | ≈ 32,000 | ≈ 90,000 |
Vendor: Microsoft
What it does: the long-established classifier: less accurate than newer models but supported on effectively any hardware, which matters for equipment that must run unchanged for years.
| 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) | ≈ 30,000 | ≈ 120,000 | ≈ 320,000 |
Vendor: Meta
What it does: produces image features that let a classifier be trained on very few labelled examples per category, which shortens a project when labelling is the bottleneck.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 8,000 | ≈ 30,000 | ≈ 85,000 |
Vendor: OpenAI
What it does: classifies against written category names with no training examples at all, which is how a new category is covered the day it is defined. Less accurate than a fitted model.
| 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) | ≈ 15,000 | ≈ 60,000 | ≈ 170,000 |
Vendor: Google
What it does: a stronger model of the same kind, better at distinguishing categories described in similar words, and a good middle option before committing to labelling.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 12,000 | ≈ 48,000 | ≈ 140,000 |
Model choice depends on your number of categories, how visually similar they are, how many labelled examples you have and where it must run. Our consultants review your categories and examples, then recommend a model and the accuracy it should reach on your data.
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 Image classification. 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.