Removing AI watermarks and metadata: what actually comes off
Written for designers, developers and everyday users who want to remove ComfyUI/Stable Diffusion workflow prompts or C2PA headers before publishing an image
Short answer
Three completely different things get called an 'AI watermark': container metadata (C2PA Content Credentials, IPTC tags and PNG prompt chunks), visible corner badges painted onto the image pixels, and imperceptible pixel watermarks such as Google SynthID. A metadata remover strips the container metadata cleanly without re-compressing the picture, and a crop tool trims off a corner badge, while imperceptible pixel watermarks are woven into the colour values across the entire image and stay behind.
On this page
When someone searches for how to “remove an AI watermark”, they are usually talking about one of three completely different technical mechanisms: hidden container metadata (like C2PA Content Credentials or ComfyUI workflow JSON), a visible corner badge painted onto the picture, or an invisible pixel watermark (like Google SynthID).
Because those three mechanisms live in different parts of the file, a tool that removes one will do nothing at all to the others. Here is an honest breakdown of what lives where, what comes off cleanly in your browser, and what stays in the pixels.
The three things people call an “AI watermark”
| Marker type | Where it lives in the file | Examples | Removed by metadata cleaner? |
|---|---|---|---|
| 1. Container metadata & provenance headers | Separate file segments outside the compressed image stream (APP1, APP11, eXIf, tEXt, iTXt, XMP, JUMBF) | C2PA Content Credentials (DALL-E, Firefly), IPTC trainedAlgorithmicMedia, ComfyUI/Stable Diffusion parameters & workflow chunks | Yes, completely and without re-compressing pixels |
| 2. Visible corner watermark | Ordinary RGB pixels in the bottom-right or bottom-left corner of the picture | Visible generator emblem or colour strip | No (requires cropping the outer margin with the Image Cropper) |
| 3. Imperceptible pixel watermark | Subtle mathematical modulations distributed across the pixel values of the whole image | Google DeepMind SynthID, frequency-domain steganography | No (designed to survive metadata stripping, cropping and JPEG compression) |
Why people strip AI container metadata from their own files
There are several practical, everyday reasons to clean container metadata before publishing or sending an image:
- Protecting ComfyUI and Stable Diffusion workflows: by default, ComfyUI embeds your entire node graph (
workflowJSON) and prompt (promptJSON) into every exported PNG file’s text chunks. If you send that PNG to a client or upload it to a site that serves original files, anyone can drag your image into their own ComfyUI window and load your exact proprietary setup, LoRA names and local folder paths. - Photoshop Generative Fill edit chains: using a minor Generative Fill touch-up to remove a power line or dust spot on a real camera photograph can attach a C2PA history block that causes automated social platforms to label the entire photograph as “Made with AI”.
- Shrinking web thumbnails: a signed C2PA manifest chain with embedded thumbnails stored across multiple JPEG
APP11segments can add 50 KB to 300 KB of non-pixel data to a 40 KB web image.
How to strip AI metadata in three steps
You can inspect and strip all container metadata, including C2PA APP11 / JUMBF blocks, XMP, IPTC and PNG tEXt / iTXt prompt chunks, directly in our free photo metadata remover:
- Open the photo metadata remover in your browser and drop one or more JPEG, PNG or WebP images onto the box.
- Check the detected metadata summary to see whether the file holds C2PA Content Credentials, Software tags, XMP Credit or PNG prompt chunks.
- Click Save clean copy to download a byte-cleaned file with all metadata and C2PA blocks removed and the original compressed pixels and colour profile preserved intact.
Because our metadata tool splices out the container segments at the byte level rather than decoding and re-saving the image on a canvas, your picture suffers zero generation loss.
What no metadata tool can remove (and platform rules to know)
If an image was generated by a model that uses Google SynthID or another imperceptible pixel-level watermark, stripping EXIF, XMP and C2PA headers will not stop a SynthID detector from recognising the pixels. SynthID modifies high-frequency colour values across the image specifically because container headers are so frequently stripped by web browsers and messaging apps.
Finally, keep platform and copyright rules in mind:
- Removing a paid stock-photo or artist watermark to avoid paying for a licence is copyright infringement in almost every jurisdiction.
- Several stock agencies, contests and newsrooms require disclosure when submitting synthetic or AI-assisted media regardless of whether the file still carries a C2PA tag. If you want to verify what labels an image currently carries before submitting or publishing it, read our guide on how to check if an image is AI generated.
Questions
Does stripping metadata remove C2PA Content Credentials?
Yes. C2PA manifests are stored as metadata blocks inside the file container (APP11 JUMBF segments in JPEG, caBX chunks in PNG and JUMBF chunks in WebP). Removing metadata with our in-browser tool cuts out those segments while leaving the image pixels untouched.
Does removing metadata erase Stable Diffusion or ComfyUI prompts from a PNG?
Yes. Local AI tools store your prompt, negative prompt, seed and node workflow inside PNG text chunks (tEXt and iTXt). Saving a clean copy in the photo metadata remover drops every text chunk so recipients only receive the image pixels.
Can a metadata remover strip Google SynthID?
No. SynthID is not stored in an EXIF, XMP or C2PA header. It subtly adjusts pixel values across the picture itself, so byte-level metadata cleaning leaves SynthID intact.
How do I remove a visible AI logo or colour bar in the corner of an image?
A visible corner badge is part of the image pixels, so metadata removal will not touch it. If the badge sits in an outer margin you do not need, trim that edge off with the in-browser Image Cropper.
Why do creators strip AI metadata from their own images?
Designers using ComfyUI, Stable Diffusion or Photoshop Generative Fill often want to publish clean artwork without exposing their proprietary node workflows, internal filenames, software versions or edit timestamps, and C2PA manifests can add 50 KB to 300 KB of overhead to small web thumbnails.
Keep reading
How to check if an image is AI generated
Skip the guessing detectors: check a photo’s C2PA Content Credentials, IPTC tags and EXIF headers in your browser first.
GuideHow to remove location from your photos
Turn off Location in the share sheet, stop the Camera recording where you are, and clean files before you email them.
GuideWhat is a WebP file and how do you open or convert it?
Why saved web images end in .webp, how WebP compares with JPG and PNG, and how to convert WebP without uploading your file.
