Model reference — Documents · Pictures
This service hides a short piece of text inside a photograph so the picture looks unchanged to anyone viewing it, and the text can be read back later by software that knows how to look.
Two properties are traded against each other. Capacity is how much text fits — typically tens to a few hundred characters, not a document. Robustness is whether the text survives what happens to the image afterwards: recompression, resizing, screenshotting, printing and rephotographing. Simple methods hide more text but lose it the moment the file is recompressed; model-based methods hide less and survive far more, because they place the signal in the parts of the picture that transformations preserve. Neither is encryption: hidden text should be encrypted before it is embedded if it is confidential.
Models in this group embed text into page images and document scans, and read it back afterwards. 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 image of about 2 megapixels with a hundred characters embedded and read back.
Vendor: Adobe Research
What it does: embeds a short payload that survives recompression, resizing and cropping, and reads it back reliably. Designed for provenance marking, which makes it the most robust general choice.
| 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: Stanford University
What it does: a compact model that embeds and recovers a fixed-length payload, trained to resist common transformations. Small enough to run at very high 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: community-maintained open-source project
What it does: a classical signal-processing method needing no GPU at all. It survives ordinary recompression but not printing or screenshotting, and it is the cheapest option by a wide margin.
| 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 embed text into photographs and read it back after the picture has been shared, resized or reprinted. The three hardware tiers mean the same as above. The sample input here is one photograph of about 4 megapixels with a hundred characters embedded and read back.
Vendor: University of California, Berkeley
What it does: embeds text that survives being printed and photographed again, which is what allows a poster or a physical document to carry a hidden code readable from a phone picture.
| 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: Data to AI Lab, Massachusetts Institute of Technology
What it does: embeds a payload robust to heavy recompression and social-media processing, at the cost of a small loss of image quality.
| 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 projects
What it does: classical least-significant-bit tools that hide far more text than any model — kilobytes rather than characters — but lose it entirely if the image is recompressed or resized. Suited to files that will be transferred untouched.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | No GPU required | No GPU required |
| VRAM | — | — | — |
| vCPUs | 1 | 4 | 16 |
| RAM | 2 GB | 8 GB | 32 GB |
| Server | CPU instance, 1 vCPU | CPU instance, 4 vCPU | CPU instance, 16 vCPU |
| Rate (images/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 600,000 |
The right method follows from what the image will go through after embedding and how much text must survive. Our consultants review your images, your capacity requirement and the transformations you need to withstand, then recommend a method and report the measured recovery rate on your own material.
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 steganography — hide text in photo. 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.