“AI label” can mean several different things. A prompt saved in a PNG text chunk, a signed Content Credential, an invisible watermark in the pixels, and a detector’s probability score are separate technologies with different limits. Treating them as interchangeable leads to bad privacy decisions and overconfident claims about whether an image was generated.
For a small business or independent creator, the useful first question is what you need the file to do. Do you want to remove personal workflow details before sharing? Preserve a provenance record? Check for a watermark from a particular generator? Unmetaimage can create a local browser re-encode for common formats, but it does not promise undetectability or pixel-watermark removal.
1. File metadata is attached information
Metadata is structured information stored in or alongside an image container. Examples include EXIF camera and GPS fields, IPTC captions, XMP editing information, PNG text chunks, ICC color profiles, and application-specific prompt or generation fields. A metadata viewer can read these values without making a judgment about what the pixels depict.
Metadata is convenient for a workflow. A prompt and seed can help an artist reproduce a draft. A color profile can help a designer manage a handoff. A timestamp or GPS coordinate can be a serious privacy concern. Metadata can also be lost when a file is edited, copied, or exported, so its absence is weak evidence about how the image was created. Google DeepMind makes the same distinction in its SynthID overview: metadata is stored with the file and can be removed or lost during editing.
2. An invisible watermark is part of the pixels
An invisible watermark changes image data in a way intended to be imperceptible during ordinary viewing. A detector that knows the watermarking method looks for the embedded pattern. It is therefore different from an EXIF field: deleting a text block does not directly address a signal distributed through pixel values.
The C2PA specification uses “soft binding” as a technical term for a fingerprint or an invisible watermark that can help discover a manifest. Separately, Google describes SynthID as an example that embeds a digital watermark directly into AI-generated images and is designed to remain detectable after some common modifications. Those statements describe particular systems, not a guarantee that every AI image has a watermark or that every watermark survives every edit.
A watermark can be used for provenance, platform labeling, ownership, or detection. The purpose and detector matter. Without knowing which system created it, a generic metadata cleaner cannot identify or safely remove it.
3. AI detection is a model judgment
An AI detector analyzes an image and returns a score, confidence, or state based on patterns it learned. That output is neither metadata nor a watermark. It may use pixels, compression traces, a known watermark, a reference database, or a combination of signals. Results can be uncertain, and a detector designed for one watermark or generator should not be treated as a universal test for every image.
Provenance is a different kind of evidence. A valid C2PA signature can show that a signer made a particular claim about an asset and that the covered content matches the signed binding. The C2PA standard explicitly separates validation from a value judgment about whether the claim is true. A missing credential does not prove that a file is human-made, and a detector score does not prove who created it.
What a local re-encode changes
When a browser decodes an image into a canvas and exports a new JPEG, PNG, or WebP, it creates a new file from rendered pixels. It may omit attached EXIF, XMP, prompt text, or a manifest. The MDN toBlob() documentation explains that the browser’s encoder accepts a requested type and an optional quality value for lossy JPEG and WebP.
That process can be useful when your goal is to reduce attached metadata before sharing. It does not certify that all metadata disappeared. It does not guarantee that a pixel-level watermark was removed. A lossy export may weaken, preserve, or otherwise change a watermark signal depending on the method and settings. Treat the output as a new derivative and verify it for the purpose you actually care about.
Choose the action that matches the question
If the goal is privacy, inspect and remove attached fields, then check GPS, software, timestamp, and prompt values in the output. Read EXIF and GPS privacy for images for a checklist.
If the goal is reproducibility, keep the original and save the prompt, seed, model, and edit recipe separately. Do not assume that a public copy’s metadata is a reliable archive; the Stable Diffusion parameters guide explains why the seed alone is incomplete.
If the goal is provenance, preserve the credentialed source and use a C2PA-aware validator. Cleaning its metadata can break the embedded or referenced manifest, so read C2PA Content Credentials explained first.
If the goal is watermark or detector research, use the specific provider’s documentation and an authorized analysis tool. Unmetaimage is a local metadata and format re-encoding workflow; it is not a detector, and it does not promise to remove invisible or visible watermarks.