One question, two tools. First: can you mark an image so no model strips the
watermark without wrecking the asset? Second: given a suspect file, can you prove it
was scaled, edited, or scrubbed? We put six frontier editors to the test.
See it live
The watermarked asset vs. the AI removal result. Pick a version on the
left, an attack model on the right.
Drag the slider. The watermark is baked into the box — the AI eraser can only strip it by rebuilding the asset.
What we found
Two tools
Research
In Search of an Unbreakable Watermark
Four versions of a protected-preview watermark, each attacked with six image-editing
models. See exactly what survives, and what the forensic verdict was.
Open the lab →
Tool
Detector — Image Forensics
Drop in an original and a copy. Find out if it's a clean resize, an edited composite,
or a watermark-scrubbed manipulation — with the evidence.
Open the detector →
In Search of an Unbreakable Watermark
Each version is rendered on the same box. We then ask image-editing models to
remove the watermark, and run the result through the Detector. A verdict of
manipulated means the eraser had to damage the asset — the watermark won.
Detector — Image Forensics
Click an area to select it, then paste (⌘/Ctrl-V) an image into it — or use
browse / drag-and-drop. Beyond clean-resize vs manipulated, it harvests forgery
evidence: the registry serial, our invisible watermark, and the suspect's own C2PA /
AI-tool traces.
browse
Original (reference)
Selected — paste an image here, or browse / drop
browse
Under review
Click to select, then paste — or browse / drop
Example detections
3 real icons8 3D-Stickle tests — click any to load & re-run
