Does an AI watermark remover erase invisible provenance?
No visible cleanup can establish that. Cleanmark changes selected image pixels. In our controlled browser test, it made an artificial corner badge unreadable, while the repaired corner still showed a patch. The exported PNG also lacked a plain text field placed in the input file. That field was ordinary metadata, not an invisible AI watermark. We did not run a SynthID detector, so this experiment says nothing about whether SynthID survives or disappears.
The input is an abstract graphic we made for this test, with our own artificial TEST badge. It is not a real photograph, a Gemini export, or proof that the tool can reconstruct covered detail. Edit only images you own or have permission to change.
What this browser run measured
Our input and downloaded output were both 640 × 400 pixels. The brush selection covered 8,700 pixels. Comparing the decoded images found 6,724 changed pixels, all inside the tested corner region. There were no changed pixels elsewhere in this fixture. These counts describe this one run, not a quality score or a general device limit.
We added a PNG tEXt field containing a controlled test note to the input. A parser found that field before the edit and did not find it in the downloaded PNG. This is a file-export observation. The tEXt field is a standard place for plain text in a PNG; it is not an invisible pixel watermark. The output image still has a visible smear, so removing one sort of file data did not restore the art beneath the badge.
The run was automated against the unchanged homepage tool: upload the owned graphic, paint over the badge, press Remove, then download the PNG. The browser logged no page errors. We saved the exact input, brush view, downloaded PNG and machine-readable run report. We did not upload an image to a model service for this repair.
Why the distinction matters
Visible marks, file metadata, and invisible image signals are different kinds of evidence. Google describes SynthID as an invisible watermark embedded in AI-generated media and checked with a detector. The PNG specification defines ordinary tEXt chunks for text associated with a PNG file. Our check covered only the latter and the visible pixels. No responsible conclusion about SynthID follows from either observation.
For a small corner badge on a simple background, a manual brush may make the mark unreadable without cropping the frame. A larger selection, an object edge, a face, or a detailed texture can leave an obvious patch. This tool fills from nearby pixels; it does not uncover the original scene or run a generative reconstruction model. Keep the original and inspect the result at full size.
Questions
Did this test erase an invisible AI watermark?
Unknown. We did not place or detect SynthID or any other invisible provenance signal. We only tested an artificial visible mark and one ordinary PNG text field.
Does the downloaded image keep its dimensions?
In this run, yes: the uploaded file and downloaded PNG were both 640 × 400 pixels. That says nothing about a larger input, which the editor may resize before repair.
Can I use this on a private image?
Cleanmark's brush processing runs in your browser without uploading the image to our server. It handles one PNG, JPG/JPEG, or WebP image at a time and exports PNG. You choose the area to repair. Permission to edit the image still matters.