Remove C2PA watermark & metadata remover
What Is a C2PA Watermark & How to Remove It
A free tool to detect and remove C2PA watermarks, Content Credentials, and AI metadata from JPEG and PNG files. Works as a C2PA watermark remover and metadata remover for images generated by DALL·E, Adobe Firefly, and other AI tools — stripping the provenance container without touching the image pixels.
What C2PA actually is
C2PA stands for the Coalition for Content Provenance and Authenticity, a cross-industry working group (Adobe, Microsoft, Sony, BBC, and others) that published the C2PA specification. The spec defines how to embed a cryptographically signed Content Credentials manifest inside common media file containers — JPEG, PNG, MP4, PDF, and more.
Think of it as a tamper-evident label baked into the file's metadata layer. It can record:
- The tool or generator that created or edited the file
- A timestamp and optional location claim
- References to parent files (e.g., "this image was edited from that original")
- A hash of the pixel data at the time of signing
When a platform or viewer supports C2PA verification, it fetches the manifest, checks the signature, and can display a "Made with AI" badge or a provenance chain.
Which AI generators embed C2PA
Adoption is growing fast. Known embedders include:
- DALL·E / ChatGPT image output — OpenAI embeds C2PA Content Credentials
- Adobe Firefly — a founding member of C2PA; every Firefly image is signed
- Microsoft Designer / Bing Image Creator
- Stability AI / DreamStudio (selective)
- Various camera manufacturers (Sony, Nikon, Leica) for authenticity claims on photos
Beyond the manifest itself, generators often write companion XMP tags such as
xmpRights:UsageTerms, Iptc4xmpExt:DigitalSourceType with value
trainedAlgorithmicMedia, or proprietary keys like openai:gizmo_id.
These are not C2PA-signed but travel alongside the manifest and signal AI origin to platforms.
Hard-bound vs soft-bound C2PA — what you can actually strip
The C2PA spec defines two binding modes:
| Binding type | Where it lives | Can be stripped? |
|---|---|---|
| Hard-bound | File container — JPEG APP11 / JUMBF chunk, PNG tEXt/iTXt chunk | Yes — the chunk can be removed without altering pixels |
| Soft-bound | A hash or steganographic signal baked into the pixel data itself, linking back to a remote manifest | No — removing the container metadata leaves the pixel signal intact |
Most current AI generators use hard-bound manifests. Soft binding (imperceptible steganography in the pixels) is an emerging layer being rolled out by some vendors. Stripping the container does not guarantee the image is "unmarked" if soft binding is present.
How to detect a C2PA watermark in your image
Upload your JPEG or PNG to the Images tab on this tool and click Check. The detector inspects:
- JPEG APP11 / JUMBF chunks for C2PA manifests
- PNG
caBXor text chunks for Content Credentials - XMP fields for AI-origin keys (
DigitalSourceType,trainedAlgorithmicMedia, vendor-specific keys) - EXIF fields that indicate AI generators
If C2PA or AI metadata is present, the result card explains what was found and the Clean it button becomes active. If nothing is found, the button stays disabled — the file is already clean of those container-level marks.
Check your image for C2PA watermarks →How to remove a C2PA watermark
Once Check reports a positive, click Clean it. The tool:
- Parses the JPEG or PNG binary for C2PA / JUMBF segments and Content Credentials chunks
- Drops those segments from the container
- Strips matched AI-looking XMP keys and EXIF fields
- Reassembles a valid image file without the provenance layer
- Offers the cleaned file for download
The pixel data is untouched. The resulting image is visually identical to the original; only the metadata container changes. You can re-upload the cleaned file and run Check again to confirm the manifest is gone.
The optional "keep non-AI metadata" flag preserves ordinary camera EXIF (make, model, GPS, shutter speed) while removing only the AI/C2PA-specific fields — useful when you want to keep copyright or camera data but drop the AI provenance stamp.
Remove C2PA from a PNG or JPEG →What remains in the file after C2PA removal
After the container manifest is stripped, a few things may still indicate AI origin to an informed viewer:
- Pixel-level soft binding — if the generator embedded an imperceptible steganographic signal, stripping the container does not touch it. This is rare in current mainstream tools but increasing.
- Image statistics — AI-generated images have characteristic noise profiles. A forensic classifier trained on those statistics can still flag the image regardless of metadata.
- Residual XMP/EXIF keys — if you chose to keep non-AI metadata, verify the result card shows zero AI-specific fields remaining.
A clean metadata result is not a "human-made" certificate. It means the known container-level signals are gone.
When removing a C2PA watermark is appropriate
Removing provenance metadata is legitimate in several scenarios:
- You generated the image and are archiving or publishing it in a context where the C2PA manifest would be misinterpreted or cause compatibility issues
- A tool or exporter stamped AI metadata on an image you authored without AI assistance
- You are processing a large batch of internal assets and the manifest contains API keys or endpoints you do not want embedded in distributed files
It is not appropriate to remove C2PA in order to misrepresent authorship where disclosure is required — for example, in journalism, academic submission, or platforms that mandate AI labeling.
C2PA is not SynthID
OpenAI’s current provenance table says supported images from ChatGPT, Codex, and the API carry both Content Credentials and a SynthID watermark. The credential is the manifest this page is about. SynthID, from Google DeepMind and used by OpenAI for images and audio, is a signal in the media itself. Stripping APP11 or a PNG chunk removes the first. It does not repaint the pixels, so it does not remove the second.
That split is why a file can look “clean” in this tool and still be recognized by openai.com/verify, which looks for OpenAI’s SynthID signal and for a trusted C2PA manifest. Use both surfaces when the question is “what will OpenAI’s own page say about this PNG?” Use this site when the question is “which container fields can I see and strip?” A single yes or no that blends them will be wrong half the time.
Claude’s file story, as Anthropic tells it, is C2PA on supported types such as PNG and JPEG, while Claude’s text story is a SynthID-Text variant. The text mark is not in the image manifest. A caption in a Doc and a PNG exported beside it need two checks. The lab comparison is Claude, Gemini, and ChatGPT watermarks. Text watermarks in general are not stored in C2PA, which is why pasting an essay into an image checker feels like it works and answers the wrong question.
What a screenshot does
Screenshotting is the everyday way a manifest dies. The new image is a fresh bitmap from the operating system. It usually has no APP11 segment, no caBX chunk, and no signed claim. People describe that as removing the watermark. It removed the container. If a pixel watermark was robust to a screen capture, the vendor’s detector may still fire. If it was not, both signals are gone and you also degraded the picture. None of those outcomes is a certificate that a camera made the file.
Re-encoding does a quieter version of the same thing. Saving a PNG as a JPEG through an editor that does not know C2PA drops the chunk. Saving again at a lower quality can also stress a pixel watermark. This tool’s cleaner is the explicit version: it targets the segments it recognizes and leaves pixels alone, so you can say what changed. A casual “export for web” is the implicit version, and you should re-check rather than assume. The removal guide puts this file operation next to character cleaning and next to paraphrasing so the three stop sharing a name.
Documents follow a cousin of the screenshot problem. Printing a DOCX to PDF can keep some metadata and drop the rest. Copying the paragraphs into a new file keeps the sentences, including any statistical text watermark, and drops the document properties. Inspect the object you will ship. A clean PDF and a marked DOCX can be the same article.
How to read a credential, not just detect one
A detector that says “C2PA present” has not told you what the manifest claims. Content Credentials Verify is the standard’s own viewer. Open the original file there when you need the signer, the tool name, and whether the claim is that the asset was generated, edited, or captured by a camera. Newsrooms and camera makers use the same format. “Credential found” is not a synonym for “a model made this.”
Soft binding, when a vendor uses it, is the case where the manifest can be fetched again from a pixel signal after the embedded box is gone. This tool does not do that lookup and does not claim to break it. If your obligation is to keep provenance, do not strip the box in the hope that no other signal remains. If your obligation is hygiene on a file you shot yourself that a tool falsely tagged, strip the false tag, re-check, and keep the original until you are sure the tag was wrong.
Pair the file check with a text check when the asset is a post, not a lone picture. Invisible characters in the caption are the other carrier, covered in ChatGPT hidden characters. Neither check is a style score, and a missing manifest is not proof of a human photographer. That reading is what a missing watermark means. The free-tool map, including this checker next to Content Credentials Verify and OpenAI’s verify page, is free AI watermark checkers compared. If the picture came from a writer you hired, keep the original export in the same folder as the published derivative so you can show which box was removed and which file never had one.
After you ship, spot-check the URL the public actually fetches. Some CDNs re-encode images and drop the manifest you meant to keep, or they keep a derivative you thought you had cleaned. The file on the origin server and the file at the CDN edge are not guaranteed to be the same container. Fetch the public URL and run Check on those bytes before you tell a partner the credential is still attached. A partner who downloads the CDN file and opens it in Content Credentials Verify is looking at the shipped bytes, not at the file you cleaned on your laptop.
FAQ
What is a C2PA watermark?
A cryptographically signed provenance manifest embedded in the file container (not the pixels) by AI generators and some cameras. It records who made the file and when.
Can you remove a C2PA watermark?
Yes — the hard-bound manifest in the JPEG or PNG container can be stripped without touching the image pixels. Soft-bound pixel-level signals, if present, cannot be removed this way.
Is removing a C2PA watermark legal?
On content you own or generated, yes — as long as you are not removing it to circumvent a required disclosure obligation (journalism, academic policy, platform rules).
How do I check if my image has a C2PA watermark?
Upload the JPEG or PNG to the Images tab and click Check. The detector reports any C2PA manifest, Content Credentials, or AI-looking XMP/EXIF it finds.
Does removing C2PA affect image quality?
No. The pixel data is not touched. The file may be slightly smaller (the manifest chunk is gone), but the visual content is identical.
What is the difference between C2PA and EXIF?
EXIF is an older standard for storing camera metadata (shutter speed, GPS, etc.). C2PA is a newer, signed provenance standard specifically designed for authenticity claims. Both live in the file container; both can be stripped independently.