Model reference — Pictures
Face restoration repairs faces in damaged, blurred, compressed or very small photographs, rebuilding features so the picture becomes clear — the usual applications being archive and family photograph collections, old identity records, and video stills.
The same caveat as upscaling applies, more strongly. A restored face is a plausible reconstruction, not a recovery of the original: the model draws a face consistent with the blurred evidence, and a different model would draw a slightly different one. Restored faces must therefore never be used to identify anyone, and should not be fed into face recognition, where they produce confident matches to people who were never in the picture. For presentation, publication and archive work they are excellent, and the original should always be kept alongside.
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 face region from a 1-megapixel photograph.
Vendor: Tencent ARC Lab
What it does: restores blurred and compressed faces quickly, with a natural result. The general default for archive and photo-library work.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 1,600 | ≈ 5,500 | ≈ 13,000 |
Vendor: Nanyang Technological University
What it does: offers a control that trades faithfulness against clarity, so the same photograph can be restored gently for a record or strongly for display. The most useful option where both are needed.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 1,200 | ≈ 4,200 | ≈ 10,000 |
Vendor: Wuhan University
What it does: recovers heavily degraded faces where the others produce a smeared result, at a higher cost per face.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (faces/hour) | ≈ 700 | ≈ 2,400 | ≈ 5,800 |
Vendor: Tencent ARC Lab
What it does: restores the rest of the photograph around the face, which matters because a sharp face on a soft background looks wrong. Usually run alongside a face model.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 1,200 | ≈ 4,000 | ≈ 9,000 |
Vendor: InsightFace
What it does: finds and straightens each face before restoration, which is what allows a group photograph to be restored face by face rather than as one blurred whole.
| 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 |
The right model depends on how damaged your images are and whether faithfulness or clarity matters more — the two trade directly against each other. Our consultants test the candidates on your own photographs and recommend a setting, with the disclosure the output should carry.
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 restoration. 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.