AI metadata remover
Image generators label their output. This removes the labels that live in the file, and tells you plainly which ones it cannot reach.
What this looks for
Add a PNG, JPEG or WebP and every metadata block in it is listed here by name and byte count, before anything is removed. It looks for EXIF, XMP, IPTC and Photoshop resource blocks, free-text comments, modification timestamps, and C2PA Content Credentials. The colour profile is left alone on purpose, because deleting it changes how the image looks and it says nothing about you.
A watermark encoded into the pixels themselves, such as SynthID, is a different mechanism and is not something this or any metadata tool can remove.
Drag images in, paste, or click to choose
A whole folder works too · Ctrl/Cmd+V pastes · on phone this opens your photo library · up to 60 MB per file
What an AI metadata remover can and cannot take out
There are two mechanisms and they are not related. The first is metadata: a generator writes its name, and often the prompt, into a text chunk, an XMP packet or a C2PA manifest attached to the file. That is a labelled block with a length, and deleting it removes it completely. The second is a statistical watermark, where the signal is encoded into the pixel values themselves. Google's SynthID works this way, and it is designed to survive cropping, resizing and re-compression. This tool removes the first kind and has no effect at all on the second. Everything it found is listed by name before anything is deleted, so you can see which kind you are dealing with.
What it cannot do
- SynthID and other pixel-domain watermarks are not removed. They are not in the metadata, so there is nothing here for this tool to delete, and Google's own detector will still flag the image afterwards.
- A clean file is not proof that nothing was AI-generated. It only means the labels that were attached to the file are gone.
- Some platforms re-attach provenance data when you upload, so a file cleaned here can arrive at the other end labelled again.
Questions people actually ask
- Will this make an AI image undetectable?
- No, and it would be dishonest to imply otherwise. It removes the labels stored in the file. Detectors that examine the pixels, including Google's SynthID Detector, work on the image content and are unaffected. If a tool tells you it removes SynthID by stripping metadata, that claim does not hold up to the first test you run.
- What do generators actually write into the file?
- It varies by tool. Some write only their name into a text chunk. Others include the model version, the settings, and in a few cases the full prompt. C2PA manifests go further and record a signed chain of what created and edited the file. Add an image above and the listing names what yours is carrying.
- Why remove AI metadata at all?
- The usual reason is that the prompt is in there. People treat a generated image as a finished piece of work and do not expect the file to carry their working notes, the model they used or their account's fingerprints into every place they post it.
- Does an image with no AI metadata mean it was made by a human?
- No. Metadata is trivially removable, which is exactly what this page does, so its absence proves nothing in either direction. Provenance systems are built on the assumption that the presence of a valid signature means something, not that its absence does.
- How can I tell which kind of watermark my image has?
- Add it above and read the list. If blocks appear, those are metadata and they are removable, and the list names each one so you can see whether it is a generator tag, an XMP packet or a signed manifest. If nothing appears, the file carries no metadata, and that is all the listing can tell you. It does not mean the image is unmarked, because a pixel-domain watermark leaves nothing in the container for any tool to find. Checking for that means running the image through the detector belonging to whoever generated it.
- Does removing the metadata change the image the model produced?
- Not by a single pixel. The image data is copied straight from the original and never decoded, so what comes out is the same render at the same resolution. This is worth stating because the common workaround, pushing the file through a converter or a compressor to shed its tags, does re-encode it, and on the flat gradients and fine detail that generators like to produce, that re-encode is visible.