Model reference — Documents · Pictures
Document redaction removes sensitive content from a document and produces a copy that can be released — to a customer, a court, a regulator, or the public — with the removed material genuinely gone rather than merely hidden.
The distinction between hidden and gone is the whole point. A black rectangle drawn over a PDF leaves the text underneath, recoverable by anyone who copies it; a correct redaction removes the text and, where required, replaces the page with an image so nothing survives. That makes redaction a two-part problem: finding what must go, which needs models that locate sensitive content and its position on the page, and producing the output safely, which is engineering rather than modelling. We do both, and verify the result by searching the released file for what should no longer be in it.
Models in this group take text or whole documents as input: plain text, PDFs, scanned pages and office files. 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 A4 page at 300 dpi redacted and verified.
Vendor: Microsoft
What it does: finds personal data and structured identifiers and applies your chosen masking. The usual backbone, because it covers both the detection and the replacement in one framework.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | RTX 4090 | A100 80 GB |
| VRAM | — | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | CPU instance, 4 vCPU | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (pages/hour) | ≈ 5,000 | ≈ 20,000 | ≈ 60,000 |
Vendor: Urchade Zaratiana and contributors
What it does: finds the categories specific to your case — a witness name, a supplier price, an unreleased product code — from a written list, without training data.
| 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 (pages/hour) | ≈ 3,000 | ≈ 12,000 | ≈ 34,000 |
Vendor: Microsoft
What it does: locates sensitive content and its exact position on the page, which is what allows the area to be removed rather than the whole page suppressed.
| 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 (pages/hour) | ≈ 1,600 | ≈ 6,000 | ≈ 15,000 |
Vendor: Datalab
What it does: reads scanned pages in roughly ninety languages and returns each word with its coordinates, so scanned material can be redacted as precisely as born-digital text.
| 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 (pages/hour) | ≈ 700 | ≈ 3,000 | ≈ 6,000 |
Vendor: Baidu (PaddlePaddle)
What it does: a fast character reader that returns word positions, used as the reading stage under a detection model on large scanned volumes.
| 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 (pages/hour) | ≈ 3,000 | ≈ 12,000 | ≈ 34,000 |
Models in this group take a page image or photograph as input, and return the regions to be removed. The three hardware tiers mean the same as above. The sample input here is one page image at about 2 megapixels with the regions to redact identified.
Vendor: Alibaba Cloud
What it does: reads a page image and identifies sensitive regions described in plain words — "any signature", "any photograph of a person", "any handwritten note" — which pattern-based detection cannot express.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 (7B, reduced precision) | A100 80 GB (7B, full precision) | 2× H100 80 GB (72B model) |
| VRAM | 16 GB | 80 GB | 160 GB combined |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 200 GB |
| Server | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB | 2× H100 SXM 80 GB |
| Rate (images/hour) | ≈ 400 | ≈ 1,600 | ≈ 900 (72B model, higher accuracy) |
Vendor: Microsoft
What it does: a small vision model that locates faces, signatures and stamps on a page quickly enough to run over an entire archive as a first pass.
| 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) | ≈ 1,800 | ≈ 6,000 | ≈ 14,000 |
Requirements differ sharply between a subject access request, a court exhibit and a public release. Our consultants review your obligations and a sample of your documents, then recommend the detection models, the output format, and the verification step that proves the released file is clean.
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 Document redaction. 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.