Invisible vs visible watermarks
They are not competing answers to one question. They answer two different questions, and most people only ever think about the first.
The short version
A visible watermark answers “who made this?” and discourages people from casually lifting it. An invisible watermark answers “which copy did this come from?” and does nothing to discourage anyone, because they cannot see it.
If you are publishing work openly and want credit attached, you want the visible one. If you are sending work to a specific list of people before release, you want the invisible one. If you are doing both, use both — they do not interfere.
Side by side
| Visible | Invisible | |
|---|---|---|
| Who can see it | Everyone | Nobody, including the recipient |
| Same on every copy | Usually yes | No — one identifier per recipient |
| Answers “whose copy was this?” | No | Yes |
| Survives a crop | Only if the mark is not in the cropped area | Yes, until very little frame is left |
| Survives a screenshot | Yes — it is part of the picture | Yes |
| Removable by re-encoding | No | No |
| Affects how the work looks | Yes, by design | No — 47.43 dB PSNR against the original |
| Deters casual copying | Yes | Only if recipients know it is there |
Where visible watermarks win
Deterrence is real. A photograph with a name across it is meaningfully less likely to be lifted for a blog post than one without, because the effort of removing it exceeds the value of the theft. That is a genuine benefit an invisible mark cannot provide, since nobody knows it is there.
They are also self-evident. No tooling, no decoder, no service: the attribution is right there in the picture, readable by anyone, forever.
Where visible watermarks fall down
A visible watermark is the same on every copy, so it cannot tell you which recipient leaked a file. It also damages the thing it is protecting — which is fine for a portfolio thumbnail and unacceptable for a screener that a festival programmer is meant to evaluate on its craft.
And they are removable. Crop, clone-stamp, or generative fill will take one out of a still image in seconds, and the result looks clean.
Where invisible watermarks win
Per-recipient identification is the whole point, and nothing else does it. The viewing experience is untouched — Blurry's AI method measures 47.43 dB PSNR against the original, which is far beyond the point where a difference is visible.
And because the mark is spread redundantly through the image rather than sitting in one place, removing it means damaging the whole picture rather than one corner of it. You can test that claim directly on the detector.
Where invisible watermarks fall down
They deter nobody, because they are invisible. Their value depends entirely on recipients knowing — in general terms — that copies are individually traceable. Telling people is part of the mechanism, not a leak of the secret.
They also need a decoder to read, which means the attribution lives with whoever holds the tooling rather than being self-evident in the file.
See what an invisible mark actually looks like
Which is to say: see that it looks like nothing at all, then check that the ID is still in there.