Invisible watermarks for AI-generated images
Mark every image you publish with a watermark people do not notice and software can find. Original Pictures embeds an invisible watermark in JPEG, PNG, WebP, TIFF and AVIF files, keeps a fingerprint of each image, and signs the result with C2PA Content Credentials.
- An invisible watermark instead of a logo in the corner
- Works on AI output and on ordinary photographs
- Published test results from the model authors, with their limits
- MANIFEST
- Embedded · valid
- WATERMARK
- In the pixels or samples
- RESULT
- Declared origin readable
Recovery is a lead back to your signed record, not cryptographic proof. Picture: AI-generated illustration.
Add a hidden watermark without a visible logo
A visible logo is the oldest way to claim an image, and it is easy to remove: one crop and it is gone, and it covers part of the picture you wanted people to see. An invisible watermark works differently. It is written into the image content itself as a very small change to many pixel values, spread across the whole picture.
The pixels do change. A watermarked image is not identical to the original, and anyone who compares the two files byte by byte will see a difference. What matters is that the change is small enough that people looking at the image at normal size should not notice it, while our detector can still read it.
Our default image watermark is built on Meta PixelSeal, an open model described in a published research paper. For teams that prefer it, several Adobe TrustMark variants are also available. The watermark documentation lists every option.
How image watermarking works
-
Submit the image. Your system uploads the file through a short-lived link. We check its size and checksum before anything else happens.
-
Embed the mark. The watermark model writes a 256-bit message into the image. Depending on the mode you choose, we also store a fingerprint of the image on our servers and sign the file with a C2PA manifest.
-
Get the output. You download the processed image. It stays in the format you uploaded, with the C2PA manifest embedded inside the file. There is no second file to keep track of.
-
Detect or recover it later. When a copy of the image appears somewhere else, you submit it for checking. If the C2PA manifest is still attached and matches, the result is valid. If the manifest has been stripped, we read the watermark and compare the image's fingerprint with the records in your account. A match is reported as recovered.
A common question is whether detection needs the original file. It does not. The watermark is read from the copy alone. The original is only needed in the sense that we match what we find against the records you created when you signed it.
Integrate watermarking into image generation or publishing
The natural place for the watermark is right after your image model produces a picture and before the file reaches your users or your CDN. The same step works for photographs coming out of an editing or upload pipeline.
For images there is an express mode that returns the watermarked, signed file in the same request. It suits generation apps where a user is waiting on screen:
curl -sS -X POST "{{API_ORIGIN}}/v1/sign" \
-H "Authorization: Bearer $OP_API_KEY" \
-H "Idempotency-Key: $(uuidgen)" \
-H "Content-Type: application/json" \
-d "{\"asset_id\": \"$ASSET_ID\", \"recipe_id\": \"rcp_system_watermark_then_sign\", \"mode\": \"express\"}"If the job finishes within the wait, the response includes a short-lived download link for the result. If it needs longer, you get a job you can poll, exactly as in the standard flow.
For publishing systems such as a CMS or a digital asset manager, the background flow usually fits better. You send each new image when it is uploaded, and a signed webhook tells your system when the watermarked file is ready. Your pipeline then swaps the original for the protected copy before it goes live. Batch requests of up to 100 images help when you are processing an existing library. The quickstart covers every step.
Image formats, dimensions and file limits
Five image formats can take a watermark. Other image formats can still be signed with Content Credentials, but without a watermark or fingerprint.
| Format | MIME type | Upload limit | Watermark | Notes |
| JPEG | image/jpeg | 25 MiB | Yes | |
| PNG | image/png | 50 MiB | Yes | |
| WebP | image/webp | 50 MiB | Yes | |
| AVIF | image/avif | 50 MiB | Yes | One 8-bit SDR still image, alpha allowed, up to 40 megapixels and 16,384 pixels per side |
| TIFF | image/tiff | 128 MiB | Yes | One uncompressed 8-bit RGB image, upright, without alpha, tiles, thumbnails or an embedded color profile |
| GIF, HEIC, HEIF, JPEG XL, SVG, DNG | various | 16 to 128 MiB | No | Signed with Content Credentials only |
A few details are worth checking before you send large batches. TIFF files with transparency or an embedded ICC color profile are refused, so export them without those first. AVIF files may keep their alpha channel. For PNG and WebP, the documentation does not yet describe how transparency and color profiles are handled, so test a few of your own files in a demo if this matters for your product. The formats documentation is the current reference, and the live capabilities endpoint shows what your account accepts.
What happens after resizing, compression or cropping?
This is the question every buyer asks, and it deserves numbers rather than adjectives. The figures below come from Meta's published tests of PixelSeal, our default image model. They are not results from our own production system.
The measure is bit accuracy: the share of the 256 hidden bits that are read back correctly. A score of 1.00 means every bit was recovered. A score of 0.50 is what you would get by guessing, so it means the mark is effectively gone.
- Bit accuracy
- 1.00
- 1.00
- 0.96
- 1.00
- 0.99
- 0.97
- 0.95
- 0.90
- 0.58
- 0.86
Source: Meta, Pixel Seal paper (arXiv 2512.16874), tests on 1,000 AI-generated images.
The pattern is clear. Normal publishing steps, such as compression, resizing, small crops and color adjustments, leave most of the message readable. A crop that keeps only about a third of the image takes the watermark close to random guessing. Screenshots are not a named test in the paper, so we do not quote a figure for them.
Two more limits should be on the table. Research outside Meta has shown that a determined attacker with the right tools can remove an invisible watermark or copy it onto a different image. And the watermark is only one of three layers. When it is damaged, the image fingerprint can still lead back to your record, since our default fingerprint is designed to hold up under crops and overlays.
What about false positives? A false positive means reporting a watermark in an image that never had one. Meta's research on this family of watermarks describes setting the detection threshold from the chance of bits matching at random, so that false alarms stay below roughly one in a million images. On our platform, a watermark reading also has to match a signed record in your account before it is reported as recovered, which adds a second check. Measured false-positive rates for our production system are not published yet.
Try it on your own images
Create an API key, upload an image and send one sign request with watermark set to true.
Image quality and watermark strength
Every watermark trades quality against strength. A stronger mark survives more editing, but it changes the image more, and at some point a careful viewer can see it. A lighter mark is harder to see and easier to lose.
You choose one of three strength levels for each image or as a default for each API key:
- Subtle uses only the lightest strength the model allows. If the mark cannot be detected at that level, the job stops instead of making it stronger.
- Balanced, the default, starts light and only raises the strength if the mark would not be detectable in the output.
- Robust allows the model's full strength range.
At the moment balanced and robust run the same range for every image model we offer.
Image quality is usually measured with PSNR and SSIM, two scores that compare the watermarked image with the original. Higher PSNR means a smaller change, and an SSIM close to 1 means the structure of the image looks the same. In Meta's tests, PixelSeal reached a PSNR of 48.9 dB and an SSIM of 0.99. Among the alternatives, TrustMark P is tuned for the highest visual quality, while TrustMark Q accepts a little more visible change in return for strength. These scores are a guide, not a promise: the same paper notes that a good PSNR can still hide small local artifacts, which is why we suggest looking at magnified samples of your own images.
Connect images to signed provenance
A watermark tells you that an image came from your system. A C2PA manifest says more: who signed it, when, and whether the file has changed since. When the manifest is intact, it is the stronger evidence, and any C2PA-aware validator can check it without contacting us.
The weakness of a manifest is that it lives in metadata, and many platforms strip metadata on upload. That is where the watermark and the fingerprint help. If a copy arrives without its manifest, they can lead back to the signed original in your account. That result is reported as recovered, a strong lead rather than cryptographic proof. You get the most from image watermarking when it runs together with signing, which is how our default image mode works.
Image watermarking pricing and FAQs
Is this a visible or invisible watermark?
Invisible. Nothing is drawn on the image. The mark is a small change spread across the pixels, meant to go unnoticed at normal viewing size.
Which image formats support watermarking?
JPEG, PNG, WebP, TIFF and AVIF. GIF, HEIC, HEIF, JPEG XL, SVG and DNG can be signed with Content Credentials but do not take a watermark.
What happens to transparency or color profiles?
AVIF files can keep their alpha channel. TIFF files must be RGB without alpha or an embedded ICC profile, or they are refused. The documentation does not yet describe PNG and WebP transparency or color profile handling, so test your own files first if this matters.
Can a watermark survive a screenshot?
We do not have published screenshot results to quote. A screenshot often scales and recompresses an image, which the model handles well in tests, but it can also crop or add other content. If the watermark is damaged, the image fingerprint gives a second route back to your record.
Can we use it on photographs as well as AI-generated images?
Yes. The watermark is added to the finished file, so it works the same way on a photo, an edited image or an AI output, whatever tool produced it.
Does detection require the original image?
No. The watermark is read from the copy alone, and the result is then matched against the records in your account.

