Labels, logos and hidden watermarks are three separate things
| Kind | Where it lives | Survives re-encoding? |
|---|---|---|
| File labels (C2PA manifest, IPTC source type, generator text) | In the file wrapper | No. A new file built from pixels does not carry them. |
| Visible mark (logo, badge, corner symbol) | In the pixels, deliberately seen | Yes. It is part of the picture. |
| Invisible watermark (such as SynthID) | In the pixels, deliberately unseen | Designed to, yes. |
What a browser tool can strip from an AI image
The metadata viewer shows the labels in a file and saves a copy with none of them. It does that by drawing the picture on a blank canvas and encoding a new file, so the old manifest and text fields are never copied, and it then re-reads the new file so you can see what remains. It changes how the file is wrapped and leaves the picture's content as it was.
What no metadata tool can remove
An invisible watermark is part of the image data. Google says SynthID is designed to remain detectable after cropping, filters and compression, and that it's read with Google's detection tools. Re-encoding doesn't target it, and this site offers nothing that claims to erase it. A visible logo is likewise a drawn part of the picture, and painting over one is image editing, where it matters whose picture it is and what the mark was for.
What an AI label looks like inside the file
Generators and editors mark their output in a few places, and they don't all use the same one. A C2PA manifest is a signed block of provenance data that records the tool that made or edited the file. An IPTC digital source type is a short vocabulary term stored in the XMP packet, and one of its values labels media as created by a trained algorithm. Some local image tools write the prompt and sampler settings into PNG text fields, where anyone who opens the file in a viewer can read them. A phone or editing app can add its own software string to the EXIF. The viewer lists each of these under its own row, tagged Provenance, so you can see which mechanism a particular file uses before you decide anything.
That spread is why the first list matters. Two images from two services can look alike and carry completely different labels, and one of them may carry none because a messaging app or social platform already re-encoded it on the way to you. A file with no labels says nothing about how it was made, only that the labels, if there were any, didn't survive.
If the second list still shows something
The second list is the same reader run over the new file, so anything it prints is really in the clean copy. A Provenance row in the new file would be unexpected, because the canvas never carries one over. The remaining rows are almost always Technical ones: image width and height, the color model, the basic file structure a browser writes on its own. Those describe the file and say nothing about its origin. If a Provenance or Location row appears in the clean copy, treat the copy as not clean, don't send it, and run the original through again with a different output format.
- Rows tagged Provenance after cleaning: don't send the copy. Try PNG as the output and read the list again.
- Size or dimensions rows: expected. Every image has them.
- A visible mark you can see on screen: no metadata step removes it, because it's part of the picture.
- The same pixel-level reading from a detector: expected, since stripping labels doesn't touch the pixels the detector reads.
A short routine before you post a generated image
Decide first whether the platform wants a label. Several large platforms ask creators to disclose realistic AI-made images, and the wording of those rules changes, so read the current page for the place you're posting. Then open the image in the viewer and read the Provenance and Location rows. If you want to drop a software string or a stray location from a file you made, create the clean copy and confirm on the second list. Keep the original with its labels in your own storage, because a file that still carries its manifest is the easiest evidence that you made it if anyone asks later.
People who sell or license generated images have a second reason to keep the labeled original. A buyer or a marketplace review can ask for the source file, and a clean copy can't answer that. Keep both, name them so you can tell them apart, and send only the version that fits the audience.
Legitimate reasons to strip labels, and where the line falls
There are ordinary reasons to remove file metadata from your own images. The manifest can name the software or account you used, the EXIF can carry a location, and a smaller file loads faster. Those are privacy and housekeeping reasons, and they apply to AI-generated images you made as much as to photographs.
The line is intent and context. Removing labels so that other people will believe a generated image is a photograph, or so that a platform's disclosure rules don't apply, deceives the people who rely on those labels. Disclosure rules for AI-made content differ by platform, publisher and place, and they're changing, so read the ones that apply where you're posting. If you're asked to label something, label it.
Strip the metadata from your own image
- Open the metadata viewer and choose your image.
- Read the Provenance rows to see what the file says about itself.
- Choose an output format and press Create clean copy.
- Download the copy and check the second list. It shows what markers the new file still contains.
Will stripping the labels hide an image from AI detectors?
No. Detectors that analyze pixels don't need the labels, and a pixel watermark, if one was applied, is still there. Stripping the labels removes one source of information and leaves what the picture shows unchanged. For how people actually go about checking, see how to check if an image is AI generated, and for what the labels are in the first place, Content Credentials explained.
The logo in the corner of a free-tier image
If a service puts a visible mark on images from a free tier and you want to publish without it, your options are to use a plan that omits it, regenerate the image elsewhere, or keep the mark. Editing it out of someone else's work isn't something this guide covers. If your own concern is the personal data in a photo and not an AI label, removing photo metadata on an iPhone is the more relevant page.
Questions people ask
Is there a free way to remove an AI watermark?
You can strip file labels from your own image for free with a browser re-encode, and this page does that. You can't reliably remove an invisible pixel watermark with a free tool, and this site doesn't offer one. A visible logo needs image editing, which is a different job from metadata cleaning.
Does stripping AI metadata change the image?
The pixels are redrawn and re-encoded, and JPEG and WebP add a little compression. At high quality settings the picture looks the same to the eye, though it's technically a new file.
What is SynthID?
SynthID is Google's imperceptible watermark for AI-generated content, designed to survive common edits. Removing file metadata doesn't strip it, because it lives in the pixels.
Will removing metadata fool an AI detector?
No. Detectors that read pixels or watermarks don't depend on the file labels, so the same image gets the same reading with or without them.
Which AI generators does the viewer name?
None by guessing. It prints what the file says, such as a C2PA claim generator, an XMP creator tool or a PNG settings field, and it stays silent when the file says nothing.
Does the image get uploaded anywhere?
No. It's read and re-encoded in your browser, and the file stays on your device. You can confirm that in the browser's network panel while the page works.
Sources