ChatGPT Image Watermark Remover
Detect and strip invisible C2PA Content Credentials from ChatGPT / GPT-Image-2 / DALL·E 3 / Sora images. Runs entirely in your browser — no upload, no signup, no server.
C2PA Watermark Cleaner
Drop one or many images. We scan for C2PA metadata, then strip it using Canvas re-encoding. No server upload — everything stays local.
Drop images here, or click to upload
Up to 50MB per file · select multiple — Browse
JPEG · PNG · WebP
What are ChatGPT image watermarks?
Every image generated by ChatGPT, GPT-Image-2, DALL·E 3, or Sora is stamped with an invisible C2PA (Content Credentials) manifest — a cryptographically-signed metadata block embedded inside the file. You can't see it with your eyes, but platforms like X, Meta, LinkedIn, and TikTok read it on upload and auto-tag the post as 'Made with AI'. C2PA is a Coalition for Content Provenance and Authenticity standard jointly developed by Adobe, Microsoft, Intel, and OpenAI. It's stored as an APP11/JUMBF segment in JPEG, a caBX chunk in PNG, or a C2PA RIFF chunk in WebP.
How the cleaner works
- Binary scan — each image is parsed to look for C2PA markers (JPEG APP11 + JUMBF, PNG caBX chunk, WebP C2PA chunk). A badge tells you exactly which files are watermarked.
- Canvas re-encode — the pixel data is drawn onto an HTML5 canvas and exported via toBlob(). This process drops every non-pixel container field in one shot: C2PA, EXIF, XMP, ICC, and any other embedded metadata.
- Quality preservation — JPEG is re-encoded at q=0.95 (visually lossless for typical AI illustrations); PNG stays lossless; WebP re-encodes at q=0.95.
- Download — cleaned files are renamed with a _cleaned suffix so you can tell them apart from originals. No file ever touches a server.
Why strip C2PA metadata?
Avoid auto 'Made with AI' tags
X, Meta, and LinkedIn now inspect uploaded images for C2PA manifests and auto-label the post. If you've paid for GPT-Image-2 output for client work, you may not want the platform advertising it.
Privacy & professionalism
C2PA manifests record the exact model, generation time, and organization ID. For freelance / agency deliverables that's usually more disclosure than the client needs to see.
Platform compatibility
Some CMSes and optimizers handle large JUMBF payloads poorly — stripping them can shave 20–100 KB off each image and avoid weird rendering edge cases.
Workflow hygiene
If you combine AI output with your own retouching / compositing, the original manifest is stale anyway — better to republish with a clean file than to ship a signature that no longer matches the pixels.
What this tool can and can't do
- ✅ Strips C2PA metadata 100% — the signature no longer verifies on the Content Credentials Verify website after cleaning.
- ✅ Also strips EXIF, GPS, ICC profiles, XMP, and any other embedded metadata, as a side effect of re-encoding. This is normally what you want, but if you need to preserve EXIF, this tool isn't the right fit.
- ⚠️ Visible 'CR' symbols that some models overlay on the pixels themselves cannot be removed by metadata stripping — those need an image editor or inpainting tool.
- ⚠️ Model-side pixel watermarks (invisible perturbations like Google SynthID, or any future OpenAI equivalent) are not touched by this tool. Metadata stripping defeats today's platform auto-tagging; it does not defeat statistical AI-image classifiers.
Technical reference: where C2PA hides in each format
If you want to verify cleaning yourself or audit a competing tool, here's the exact byte-level layout. All three image formats use a different container, which is why generic 'remove metadata' utilities sometimes miss one or two of them.
JPEG — APP11 marker + JUMBF box
C2PA data is wrapped in a JUMBF (JPEG Universal Metadata Box Format, ISO/IEC 19566-5) container, then placed inside a JPEG APP11 marker segment (0xFFEB). Because each marker payload is capped at ~64 KB, a single C2PA manifest is usually split across multiple APP11 segments. Look for the bytes 'JP' followed by 'jumb' near the start of the file.
PNG — caBX ancillary chunk
PNG stores C2PA in a custom ancillary chunk named caBX (lowercase c=ancillary, lowercase a=private, uppercase B=reserved, lowercase X=safe-to-copy). PNG strippers that only target tEXt / iTXt / zTXt — which most do — will silently leave caBX behind.
WebP — C2PA RIFF chunk
WebP is a RIFF container, and C2PA gets its own four-character chunk type 'C2PA'. RIFF readers that only know about VP8 / VP8L / VP8X / EXIF / ICCP chunks will treat it as 'unknown' and copy it through verbatim — meaning a naive transcode can preserve the watermark unless the encoder explicitly drops unknown chunks.
Verify the cleanup with one command
After cleaning, run any of the following locally to confirm the manifest is gone:
exiftool -a -G1 cleaned.jpg | grep -i -E 'jumbf|c2pa'xxd cleaned.png | grep -i -E 'cabx|jumb|c2pa'Upload to contentcredentials.org/verify — should report 'No Content Credentials found'.
Frequently asked questions
Important
This tool is provided for educational and legitimate workflow purposes. Do not use it to impersonate human creators, commit fraud, or circumvent disclosure requirements imposed by the platforms you publish on or by law in your jurisdiction. All trademarks — ChatGPT, DALL·E, GPT-Image-2, Sora — are property of OpenAI; we are not affiliated with or endorsed by OpenAI. C2PA is a trademark of the Coalition for Content Provenance and Authenticity.