INITIALISATION DU SYSTÈME...
Google Imagen and Gemini image outputs ship with two provenance layers: a C2PA Content Credentials manifest in the JPEG APP11 segment, and an invisible SynthID watermark embedded at pixel level. GhostMeta strips the C2PA layer — what most viewers and platforms read — via canvas re-encode. SynthID is designed to survive re-encoding by Google's research and would require destructive editing or a dedicated remover to defeat. Beyond the APP11 manifest, Imagen and Gemini exports often write an IPTC "Digital Source Type" value set to trainedAlgorithmicMedia into the image's XMP / IPTC photo-metadata — a plain-text AI flag that stock libraries and newsrooms increasingly scan on ingest, independent of C2PA. Because Imagen ships as JPEG, the file also carries a standard EXIF APP1 segment alongside the C2PA APP11 segment, so a clean strip has to clear both. GhostMeta wipes the EXIF, the IPTC AI marker and the C2PA manifest in a single canvas pass; the JPEG is re-encoded at a fixed high quality (92%), and the pixels — and the SynthID baked into them — are left untouched.
SynthID is written into the pixels and survives a re-encode: GhostMeta does not remove it and does not claim to. What the tool does remove is the file's metadata — EXIF, XMP, IPTC and the provenance manifest when present. The operation is local and the file is sent nowhere.
Google marks its generated images with SynthID, a watermark written into the pixels themselves. It is designed to stay detectable after cropping, format changes or recompression. Cleaning metadata does not reach it, and the canvas re-encode used here is exactly the kind of transformation SynthID resists.
So the limit has to be stated before anything else: an Imagen image cleaned here still carries its pixel marking. This page does not sell you the opposite.
What goes is the file's metadata layers: EXIF, XMP, IPTC, and the provenance manifest when present. On a JPEG, those blocks disappear in the re-encode.
The benefit is real but bounded: you remove tool history, timestamps and export information. You do not remove the signature in the pixels. To see exactly what your file holds before deciding, drop it into the metadata viewer: the list of fields appears without any data leaving the device.
Even with the pixel marking in place, removing metadata has a concrete effect: the file no longer declares the tool that produced it, nor the date, nor the export identifiers. A metadata reader finds nothing left.
That is a reduction in surface, not a disappearance. It holds for everyday use, where nobody runs a specialised detector, and it does not hold against a properly equipped verifier.
| Marking | Carrier | Removed |
|---|---|---|
| EXIF, XMP, IPTC | File metadata | Yes |
| C2PA manifest | File metadata | Yes, when present |
| SynthID | Pixels | No, by design |
| Server-side fingerprint | Outside the file | No |
No. SynthID is a pixel-level steganographic watermark designed by Google DeepMind to survive standard re-encoding. GhostMeta strips the C2PA metadata layer only.
Yes. Gemini's image generator pipes through the same Imagen backend and produces the same C2PA + SynthID combination.
The C2PA manifest names Google as producer. Account-level identifiers are not stored in clear text but the assertion list may include hashed session refs.
GhostMeta decodes the JPEG to a canvas at full resolution, then re-exports at a fixed 92% quality, so any loss is a single generation and visually negligible for web or marketplace use. It is not lossless: each canvas re-encode is one JPEG pass. For a print master, keep your original and strip a working copy.
The same Imagen backend powers Gemini image generation, ImageFX in AI Test Kitchen and the Whisk tool, so their outputs carry the APP11 C2PA manifest and, when present, the IPTC trainedAlgorithmicMedia source-type value. GhostMeta treats them identically — it strips by reading the actual JPEG segments and metadata blocks, not by trusting a generator name.
The metadata-based signals it reads — the C2PA APP11 manifest and the IPTC AI source-type field — are gone after GhostMeta, so a metadata scan comes back clean. But Google's SynthID lives in the pixels and survives the re-encode, and pixel-based AI classifiers analyze the image content directly. Stripping metadata removes the declared provenance, not the statistical fingerprint of the pixels.
Partner : M.E Expert Serrurier