What is AI metadata?
AI metadata is the provenance record that an image generator writes into the file next to the picture you can see. It lives in the metadata sections of the file, not in the pixels. A camera writes EXIF, and a generator writes its own record: a signed C2PA manifest, an IPTC source type, generator fields in XMP and, in some workflows, the prompt itself. None of it changes how the image looks. All of it tells a reader where the image came from.
C2PA is the signed layer. The Content Credentials technical specification describes a manifest that can name the tool that created the image, the actions taken, the timestamps involved and a hash tied to the pixels. In JPEG, PNG and WebP files the manifest sits in a JUMBF box or an XMP packet. Because the record is signed, a reader can tell whether it was altered. Because it sits in metadata, an operation that rewrites the metadata sections can drop it.
IPTC is the descriptive layer. The Digital Source Type vocabulary defines values that name the origin of media, including trainedAlgorithmicMedia for output from a generative model and compositeSynthetic for a mix of generated and captured material. The IPTC Photo Metadata standard carries those values in the file, and it also defines a field for AI prompt information. XMP is where much of this lands in practice: a packet of tags that records the creating software, the edit history and the generator fields the tool chose to write. This page removes those records from files you own.
How it works
How to remove AI metadata from an image
Step 1: Upload your image
Drop a PNG, JPEG or WebP file into the panel. The upload goes straight from your browser to processing storage, and nothing is published. The original file stays untouched until the strip runs.
Tip
The file the generator exported carries the richest record. Screenshots often carry almost none.
Step 2: Run the metadata strip
The metadata pass is selected by default and costs 5 credits. It reads the file, removes the C2PA manifest store and its wrapper, the IPTC source type, the XMP generator fields and any prompt data, then writes the image back out. The pixels are copied through, so the picture keeps its resolution and color depth. The job runs server-side and returns a cleaned file in seconds.
Tip
Credits are consumed only when a strip job actually runs.
Step 3: Download the clean version
Compare the result with the original, then download. A metadata viewer will show the provenance fields gone. The file keeps its original dimensions and framing. If you are signed out, the page asks you to sign in first, then restores the panel when you return.
Tip
Platforms can write their own metadata when they re-encode an upload. Clean the file at the last step before you publish.
The AI metadata this page clears
C2PA / Content Credentials
Generators sign a manifest that can name the model, the actions taken and the time of creation. The strip removes the manifest store, its JUMBF wrapper and the XMP references that point at it. The C2PA technical specification at https://spec.c2pa.org/specifications/specifications/2.5/index.html describes what the record contains.
IPTC DigitalSourceType
The IPTC extension records the source of an image in a controlled vocabulary. Values such as trainedAlgorithmicMedia mark model output. The strip removes the field and the rest of the AI-related IPTC block. The vocabulary is published at https://cv.iptc.org/newscodes/digitalsourcetype/ and defines each value.
XMP generator fields
XMP carries the software name, the tool version and the edit history that a generator or editor wrote. Fields such as CreatorTool and the generating application's history sit here. The strip removes the packet, so the file no longer names the tool that made it.
Prompt data
Some tools store the prompt, the negative prompt or model settings in the file. The IPTC Photo Metadata standard defines an AI Prompt Information field for exactly this data, documented at https://www.iptc.org/std/photometadata/documentation/userguide/. The strip takes those fields out with the rest of the metadata.
What stays after a metadata strip
A metadata strip changes the record inside the file, not the picture. Invisible watermarks embedded in the pixels, such as SynthID, stay where they are. Visible marks such as a corner logo or a text badge stay too. OpenAI’s help center describes both layers for its images: Content Credentials and a SynthID watermark. Stripping the credentials leaves the watermark readable, so a detector that scans pixels can still find a signal.
That division is deliberate. This page is a metadata remover. It clears the provenance record that platforms and verifiers read from the file. If you need the pixel layer handled, the AI Watermark Remover runs a separate pass over the image data. If a visible logo is the problem, that same page is the right starting point. The layers fail and succeed independently, so matching the pass to the layer you want gone is the whole decision.
The strip also leaves image quality alone. It works on the metadata sections of the file and does not re-encode the picture in a lossy format, so resolution, color depth and framing come back as they went in. PNG and WebP files keep their original encoding. A JPEG keeps the dimensions of the source file, and the metadata sections are gone from the copy you download.
Limitations and honest expectations
Where to go next
Which page fits depends on the layer you want gone. For the platform label that appears on an uploaded photo, the Remove AI Label page covers what platforms read and how to strip it. For a pixel watermark, the AI Watermark Remover runs the image pass. For a closer look at the credential standard itself, the C2PA Remover page covers the signed manifest in depth. For the personal fields in a phone photo, such as GPS, device model and timestamps, the Metadata Remover page handles that privacy job.
The details on this page come from official documentation. The C2PA technical specification is at https://spec.c2pa.org/specifications/specifications/2.5/index.html. The IPTC Digital Source Type vocabulary is at https://cv.iptc.org/newscodes/digitalsourcetype/, and the IPTC Photo Metadata User Guide at https://www.iptc.org/std/photometadata/documentation/userguide/ documents the prompt fields. OpenAI’s help center at https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content explains the Content Credentials and SynthID layers in its generated images.
FAQ
Frequently asked questions
Short answers about this AI metadata remover, from C2PA manifests to the pixel watermarks it leaves alone.
Explore more tools
Remove personal photo metadata
The privacy page for phone photos: EXIF, GPS, device model and editing history stripped from the same metadata layer.
Remove AI Label
What Instagram, Facebook and other platforms read from metadata before they apply an AI label, and how to clear it before upload.
AI Watermark Remover
For the marks in the pixels: SynthID patterns, visible logos and credential layers cleared in one full-image pass.
C2PA Remover
The credential specialist: what Content Credentials record, where they live in the file and how a strip removes them.
Apple Image Playground Metadata Remover
The Apple side of the same file layer: the metadata Image Playground writes into an export, and what a strip leaves behind.
SynthID Detector
Check whether the pixel layer still carries a readable SynthID signal after the metadata strip.
ChatGPT Watermark Remover
For ChatGPT exports, where the C2PA record and the visible mark arrive in the same file.
Grok Metadata Remover
The metadata-layer clean for Grok images, with the page explaining the visible watermark xAI documents.
Stable Diffusion Metadata Remover
Strips the prompts, seeds and workflow data written by Stable Diffusion, ComfyUI and Automatic1111 exports.
All tools
Every remover and detector on one page, grouped by what each one cleans or checks.
Strip the AI metadata before you share
One metadata pass clears the C2PA manifest, the IPTC source type, the XMP generator fields and any prompt data. The picture comes back at its original dimensions.