This term points at a library's preview overlay, and the overlay belongs to the recoverable kind: it was blended into the picture by a process that runs in reverse, rather than painted on top of it. What separates this library's case from a lone badge in a corner is not the shape of the mark. It is that the same mark is applied the same way to preview after preview — which is exactly the condition that lets an overlay be established from a collection of files instead of from a sample prepared for the purpose.
That gives two routes to knowing a mark, and they are not interchangeable: one measures a mark once and exactly, the other infers one across many files and approximately. Neither of them is loaded into the tool on this site, whose single record belongs to a different generator's badge. The last section is about what actually takes an overlay off a preview.
Two routes to knowing a mark
Both routes are after the same pair of quantities — what the overlay contributes at a pixel, and how strongly it was applied there. Everything else follows from how they go about obtaining them.
- Measure it once. Composite the overlay over a backdrop whose numbers are already known, so that how much of each pixel the mark accounts for becomes something to read rather than something to guess. What comes out is exact, and it is narrow: a different size, or a different position, is a different record.
- Infer it from many. Where the same overlay has been applied the same way to a large set of pictures, the part those pictures have in common and the part that differs between them can be pulled apart. No prepared sample is needed, and no cooperation from whoever applied the overlay. What comes back is an estimate for that set — as good as the set's consistency, and no better.
The second route is the one that matters for this term, because a stock library's previews are the situation it was built for. It is also the route that a single user cannot take.
The collection this route was demonstrated on
The method was published at CVPR 2017 by four researchers, and their own description of it is worth reading rather than paraphrasing. Their premise is stated first, and it is the premise this whole page turns on:
watermarks are typically added in a consistent manner to many images … this consistency allows to automatically estimate the watermark and recover the original images with high accuracy
The instrument they built for it is described in the same abstract as
a generalized multi-image matting algorithm that takes a watermarked image collection as input and automatically estimates the “foreground” (watermark), its alpha matte, and the “background” (original) images
and the material they demonstrated it on is named in one short sentence: We demonstrate the algorithm on stock imagery available on the web.
Their supplementary material is more specific still. It lists the collections the method was run on, one entry per stock service, and gives this library's collection as 422 images drawn from it. The output for that collection is published alongside — an estimated overlay and its alpha matte, with an estimated blend factor of 0.39, 0.40 and 0.43. Three values rather than one, which is a detail worth noticing in itself: the overlay does not attenuate the three colour channels by the same amount.
None of that is a statement about any particular preview. It is a statement about a set of 422 of them, and about how alike they had to be for an estimate to be published at all. That is the useful reading of it: the consistency is not an assumption in the method, it is the thing the result depends on.
What the file holds, and how it is taken back apart
Blending a semi-transparent layer onto a picture is not stamping. Where the picture underneath is fully opaque, every pixel of the result is a mixture of two values in a proportion that changes from pixel to pixel: where the layer is solid it supplies most of the pixel, where it fades it supplies less, and where it is absent the picture arrives untouched. Nothing in that ties one pixel to its neighbour, and that is what makes the operation reversible.
Run it backwards and the whole thing is three steps:
- Start from the number stored at that pixel in the file.
- Subtract the mark's contribution there — its colour multiplied by its weight at that pixel.
- Divide what is left by the share of that pixel the mark did not cover.
The last of the three is the decisive one. What it hands back is the value from before the blending, which is why this family of techniques is called reversible rather than restorative: the mark is withdrawn and the numbers it displaced come back. A tool that paints over a mark is doing something else, and the difference is not one of degree.
Turn the same three steps around and they describe a tool that holds no record. With the mark's colour and its weight both missing, step two has nothing to subtract and step three has nothing to divide by. What such a tool does instead is take its cue from the pixels around the mark — indistinguishable from the truth across an empty sky, and visibly made up across a face, a texture, or a line of lettering.
An estimate is not a measurement
The two routes return different kinds of thing, and the difference is worth being exact about.
- A measurement pins down one mark. It says nothing at all about a second one, and it is exact: a different size, or a different position, is a different record, and neither stands in for the other.
- An estimate belongs to a collection. Approximate, collective, and conditional on the files agreeing with each other. It is not a stand-in for a measurement of the mark sitting in front of you.
The authors drew that consequence themselves and made it the headline point of the work:
visible watermarks should be designed to not only be robust against removal from a single image, but to be more resistant to mass-scale removal from image collections as well
Read in the direction of this query, that says something plain. The weakness they identified is not in one preview; it is in the fact that previews resemble one another. And the same sentence settles the practical question for anyone holding a single file: a collection is the input the second route takes, and one preview is not a collection. At the scale of one file, the route that works is the one that needs an exact record — and the exact record is a thing somebody had to prepare.
Scope, stated plainly
What is loaded here is one record — the visible badge of one specific generator, in the two layouts it has shipped — and nothing else. PNG, JPEG and WebP are the three it reads, and one image is handled per run. No library overlay is among those records, and no estimate of one either — both routes described above end at the same word here, which is absent.
Bring it a watermarked preview and there is nothing on the other side to align the overlay with; no slider or setting on the page alters that. It is not a matter of trying harder.
This term is a narrow member of a broader phrase, and that broader phrase gets a page of its own, written specifically to keep its two senses from merging: what “AI watermark remover” can mean. Getty Images takes the same reasoning from a different angle. At Shutterstock the overlay is one mark stamped repeatedly over the frame rather than a single instance. Where the thing to be identified is printed with a library's own name, the naming itself becomes the subject: a name is not a measurement. And where a mark belongs to a platform rather than to a file, one of them is not in the file at all.
What takes the overlay off a preview
The overlay is the only difference between a preview and the asset it advertises — which is what a preview is for in the first place. So what takes the overlay off is a licence and not an eraser; cleaning the preview leaves you with a tidier copy of an asset you have no licence for.
One detail in the material quoted above makes that point better than any argument could. The same page that publishes the estimated overlay for these previews also records that every stock image used in the work was licensed from the corresponding service. The study that showed a catalogue's overlay can be estimated was itself working from licensed copies.
The tool this site is built around sits on the front page, and it is for pictures of your own that carry somebody else's mark.
Sources
- Dekel, Tali; Rubinstein, Michael; Liu, Ce; Freeman, William T., On the Effectiveness of Visible Watermarks, CVPR 2017, read from the authors' project page on 7 October 2026. Source of the quoted premise that watermarks are typically added in a consistent manner to many images, and that this consistency allows the watermark to be estimated automatically; of the quoted description of the method as a generalized multi-image matting algorithm taking a watermarked image collection as input; of the sentence naming stock imagery available on the web as what the method was demonstrated on; and of the stated takeaway about mass-scale removal from image collections.
- The same authors' supplementary material, read 7 October 2026. Source of the composition of the collections the method was run on — one entry per stock service, with this library's collection given as 422 images — and of the note that all stock images shown in the paper and its supplementary material were licensed from the corresponding stock services.
- The same authors' supplementary page for this library's collection, read 7 October 2026. Source of the output published for it: an estimated watermark image and alpha matte, an estimated blend factor of 0.39, 0.40 and 0.43 for the three colour channels, and the section references attached to the intermediate stages.
- Porter, Thomas; Duff, Tom, Compositing Digital Images, SIGGRAPH '84. Where the operator for placing one image over another comes from, together with its reduced form when the backdrop is opaque: two values combined per pixel, in a proportion set by the foreground's coverage.
- Our own implementation, read on this site. Source of the scope stated above: one badge, two layouts, still images only, and no library overlay or estimate of one behind it.
- Our own observation of the results returned for this term, checked 7 October 2026. Recorded as a pattern only: no site is named, no ranking position is claimed, and no page's wording is reproduced.
- Not used: the library's own pages. They returned an access error when checked, so nothing from them is quoted here and no second-hand paraphrase was put in their place.