Model reference — Pictures · Video
Watermarking embeds an invisible, machine-readable mark in an image, video or audio track that records where it came from — which organisation produced it, when, under which licence, and whether it was generated by a model.
It answers a question that metadata cannot, because metadata is stripped the moment a file is uploaded, screenshotted or recompressed. A watermark is carried in the content itself and survives those operations, so a picture found later can still be traced to its source. The two applications are rights — proving an asset is yours and detecting unlicensed use — and disclosure, marking synthetic media as synthetic, which several jurisdictions now require. A watermark is a signal, not a lock: it identifies content, it does not prevent copying.
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 2-megapixel image watermarked and the mark read back after recompression.
Vendor: Adobe Research
What it does: embeds a provenance payload that survives recompression, resizing and cropping, and is designed to work alongside content credentials standards. The default for image provenance.
| 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) | ≈ 3,000 | ≈ 12,000 | ≈ 24,000 |
Vendor: University of California, Berkeley
What it does: embeds a mark that survives printing and rephotographing, which extends provenance to physical prints and screen captures.
| 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) | ≈ 2,000 | ≈ 8,000 | ≈ 18,000 |
Vendor: Stanford University
What it does: a compact model embedding a fixed-length mark with good robustness, small enough to watermark every asset at publication volume.
| 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) | ≈ 6,000 | ≈ 20,000 | ≈ 50,000 |
Vendor: Data to AI Lab, Massachusetts Institute of Technology
What it does: robust to heavy recompression and social-media processing, which is the path most leaked images actually take.
| 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) | ≈ 2,400 | ≈ 9,000 | ≈ 20,000 |
Vendor: community-maintained open-source project
What it does: a classical method needing no GPU, adequate where files are not reprinted or screenshotted and cost per asset must be near zero.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | No GPU required | No GPU required |
| VRAM | — | — | — |
| vCPUs | 2 | 8 | 32 |
| RAM | 4 GB | 16 GB | 64 GB |
| Server | CPU instance, 2 vCPU | CPU instance, 8 vCPU | CPU instance, 32 vCPU |
| Rate (images/hour) | ≈ 20,000 | ≈ 80,000 | ≈ 300,000 |
Models in this group take a recorded video file as input and are applied frame by frame. The three hardware tiers mean the same as above. The sample input here is one minute of 1080p video watermarked and the mark read back after re-encoding.
Vendor: Meta
What it does: embeds a mark across video frames that survives re-encoding, cropping and frame rate changes, and can be read from a short excerpt rather than the whole file.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 | A100 80 GB | 2× A100 80 GB |
| VRAM | 20 GB | 80 GB | 160 GB combined |
| vCPUs | 8 | 16 | 32 |
| RAM | 32 GB | 64 GB | 128 GB |
| Server | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB | 2× A100 SXM 80 GB |
| Rate (video minutes watermarked per hour) | ≈ 400 | ≈ 1,400 | ≈ 3,200 |
Vendor: Meta
What it does: marks the audio track so provenance survives even when the video is stripped to sound or the picture is replaced.
| 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 (video minutes watermarked per hour) | ≈ 3,000 | ≈ 10,000 | ≈ 26,000 |
Vendor: Adobe Research
What it does: applied frame by frame where a per-frame mark is wanted, so a single screenshot taken from the video still carries the provenance.
| 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 (video minutes watermarked per hour) | ≈ 60 | ≈ 220 | ≈ 500 |
Vendor: FFmpeg project
What it does: handles the decoding and re-encoding around the watermarking step, and is where a poorly configured pipeline destroys the mark it just embedded.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | No GPU required | GPU-accelerated decode (NVENC/NVDEC) |
| VRAM | — | — | 8 GB |
| vCPUs | 2 | 8 | 16 |
| RAM | 4 GB | 16 GB | 32 GB |
| Server | CPU instance, 2 vCPU | CPU instance, 8 vCPU | 1× RTX 4090 24 GB |
| Rate (video minutes watermarked per hour) | ≈ 1,200 | ≈ 4,000 | ≈ 12,000 |
The right method depends on the medium, the payload you need to carry and what the file will go through afterwards. Our consultants review your assets and distribution channels, then recommend a method and report the measured survival rate through your own pipeline.
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 Digital watermark / provenance embedding. 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.