INITIALISATION DU SYSTÈME...
If you reuse AI-generated images for client work, marketing collateral, dropshipping listings or social posts, every embedded provenance signal is a leak vector. C2PA manifests announce the generator. EXIF reveals the editing chain. XMP carries the prompt. The cumulative fingerprint is enough for an automated platform to flag the image as AI-generated and demote, label, or reject it. GhostMeta is a one-click anonymizer: drop any image, get back a clean copy stripped of every metadata layer, all in your browser. A mixed batch is the hard case: a JPEG from Imagen hides its claim in an APP11 segment, a PNG from Stable Diffusion scatters it across tEXt and iTXt chunks, and a Firefly export wraps a JUMBF box — three different containers, one pass each. GhostMeta normalizes them all by decoding to a canvas and re-encoding, so the output carries no APP11, no caBX, no text chunk, regardless of which model made it. On a stock or print-on-demand marketplace, a single leftover Content Credentials manifest is enough for an automated intake filter to reject the listing or tag it as AI before a human ever reviews it.
Anonymising here means removing what travels with the file and leads back to a device, an account, a piece of software or a moment: EXIF, GPS, IPTC, XMP, PNG chunks, C2PA manifest. The scope stops at the edge of the pixels — a face stays a face. Processing happens in the tab, with no transmission.
Anonymising an image, in this tool's sense, means removing the data travelling with the file that could lead back to a device, an account, a piece of software or a precise moment. That is a clear scope, and it stops at the edge of the pixels.
What was photographed is not anonymised: a face stays a face, a plate stays readable, an interior stays recognisable. Confusing the two senses of the word leads to publishing an image you believed was neutral.
Processing happens inside the tab: the image is decoded in memory, redrawn onto a canvas, then exported again. A canvas carries pixels only, so no metadata layer survives the operation.
You can verify this without taking our word for it. Open your browser's Network tab during cleaning: no request leaves with the file. Then read the cleaned copy back in the viewer to see which fields are gone.
The bar is not the same for everyone. Posting a holiday photo on a forum and submitting a document to a sensitive case do not call for the same care, even though the technical operation is identical.
If your safety depends on the outcome, do not rely on metadata cleaning alone: deal with the image content, the filename and the channel it travels through as well. This tool covers part of the problem, not all of it.
| Item | Covered |
|---|---|
| Device identifiers and serial numbers | Yes |
| GPS position and place names | Yes |
| Capture and editing timestamps | Yes |
| Prompt, seed, model name | Yes |
| C2PA provenance manifest | Yes, on JPEG and PNG |
| Faces, plates, recognisable places | No, those are pixels |
| Visible or pixel-level watermark | No |
Some platforms now read C2PA manifests and label images. After stripping with GhostMeta the C2PA signal is gone, but content-based AI detectors (which analyze pixels, not metadata) can still flag the image.
Most providers' terms allow you to strip their metadata for downstream use. Check the specific TOS — Adobe, OpenAI and Google generally permit it for non-deceptive use.
Any valid image at least 12 bytes. Below 10 MB processes instantly. Above that the canvas re-encode may exceed mobile memory limits — use desktop.
The cleaning is identical in result but targets different containers. In the JPEG, the provenance lives in an APP1 segment (EXIF/XMP) and an APP11 segment (the C2PA/JUMBF manifest). In the PNG, the same information sits in tEXt or iTXt chunks plus a caBX chunk for C2PA. GhostMeta decodes either format to a canvas and re-encodes, so both outputs come out with those segments and chunks absent.
No. GhostMeta does not add a signature, a comment chunk, or its own XMP block. The re-encoded file simply lacks the original APP11/caBX manifest and any tEXt/iTXt entries — it looks like an ordinary canvas export. There is no GhostMeta marker to trace back, because nothing is written in; metadata is only removed.
The pixel grid is preserved at its original resolution up to the 4096px cap, so dimensions are unchanged for typical generator outputs. Re-encoding to JPEG applies lossy compression like any save, while PNG stays lossless. Only the metadata layers — APP11, caBX, EXIF, XMP, IPTC, tEXt, iTXt — are dropped; the visible content is the same image.
Partner : M.E Expert Serrurier