This is a different problem from the text one
The two halves of this tool share an engine and almost nothing else. A document leaks through characters nobody can see. An image leaks through structured records deliberately written into the file by the camera, the editor or the generator that made it.
C2PA Content Credentials are the sharpest example: a cryptographically signed record naming the tool that produced the image and, often, the chain of edits since. That is a feature — it exists so provenance can be proven — and it is also a disclosure you may not have meant to make when you send a file to a client.
EXIF is the older version of the same exposure and is frequently worse, because it can carry GPS coordinates, a camera serial number and a timestamp. XMP blocks carry editing history and sometimes the prompt.
What comes out, and what stays
Everything in the container comes out completely: C2PA manifests, EXIF, XMP, ICC comments, generator tags, stray text chunks. We rewrite the file with the metadata gone and the pixel data copied through byte for byte, so the picture is unchanged and the check afterwards comes back empty.
A watermark in the pixels does not come out, and this is the part most tools in this category will not tell you. Google's SynthID for images works by altering the pixels themselves in a way that survives cropping, compression and recolouring. Stripping metadata does nothing to it. When a file looks like it came from a generator in that family, the report says so in as many words rather than letting a clean metadata result imply something it does not mean.
That warning is the reason this tool exists in the shape it does. A metadata cleaner that quietly reports success on a SynthID image has told its customer they are safe when they are not.
Nothing is stored
Images are processed in memory and discarded with the response. The file is never written to disk, never kept, and never used to train anything. The only thing that outlives the request is a counter saying a check happened.