Why this exists
AI generators increasingly embed a signed record of how a file was made. C2PA Content Credentials, EXIF software tags, XMP packets and PNG text chunks all travel with the file. Often that text includes the prompt you wrote, an internal job identifier, or a timestamp you would rather not hand to a client.
Most tools that remove this do it the crude way: open the image, re-encode it, and let the metadata fall away as a side effect. That works, but it costs you a generation of quality on every pass and quietly discards your colour profile.
The other common approach is to upload your file to someone else's server. For a designer handling client work under NDA, that is often simply not an option — and it is a strange trade to make for a task that is fundamentally just rewriting a container.
How the engine works
PurifyAI parses the actual file structure. For PNG it walks the chunk list and drops the metadata chunks while copying the image chunks through byte for byte. For JPEG it walks the marker segment chain and removes the APP segments that carry EXIF, XMP, IPTC and C2PA data — while checking APP2 payloads so an ICC colour profile survives even though it shares a marker with C2PA manifests.
The result is a file whose pixels are bit-identical to the original, minus the metadata. And because it is all just byte manipulation, it runs comfortably in JavaScript in your browser, which is exactly why no server is involved.
What we will not claim
Removing a C2PA manifest removes one declared signal. It does not make an image undetectable as AI-generated: classifier-based detection, pixel-level watermarking and platform-side provenance records are separate mechanisms that this tool does not touch. Anyone telling you otherwise is selling something.
We also cannot re-encode video in the browser at any sensible speed, so video support is limited to inspection and frame export. When that changes, we will say so — and not before.
How we build it
Privacy by architecture
We did not write a policy promising not to look at your files. We built a product that cannot, because there is no upload endpoint and no server-side processing. A promise can be broken; a missing endpoint cannot.
Show, do not assert
Before removing anything, the tool lists every metadata segment it found, with sizes and readable previews. You should be able to see what is in your file, not take our word for what we took out.
Lossless where it matters
Metadata scrubbing rewrites the container and copies pixel data through untouched. No decode, no re-encode, no generation loss. ICC colour profiles are preserved so nothing shifts.
Honest about limits
We say plainly what the engine does not do: it does not defeat pixel-level or steganographic watermarking, and browsers cannot re-encode video. Overstating capability would just waste your time.
Check it yourself
Open your browser network tab, scan a file, and watch nothing get sent. That is the whole pitch.