Invisible text markers
Zero-width characters, Unicode tags, bidi controls and unusual spaces can ride along in AI chat or pasted text. Text Clean reveals and removes the supported set locally.
Inspect before you clean
Text markers, metadata cleanup and pixel-level Deep Clean
Learn the difference between invisible text markers in AI output, visible watermarks, file metadata, C2PA Content Credentials and pixel-level invisible watermarks.
The layer model
Files, pixels and text each store watermarks differently
File-level metadata
JPG, PNG and WebP containers can carry camera settings, coordinates, dates, XMP, IPTC, AI parameters and C2PA manifests alongside the encoded image.
Visible overlays
A logo, caption or date stamp that you can see is already represented by image pixels. Deleting EXIF or XMP cannot reconstruct what is behind it.
Invisible pixel signals
A watermarking system may alter pixel patterns while keeping the change hard to see. A file-container scan cannot confirm whether such a signal exists.
Scan
What the browser actually checks
The scanner parses container structures rather than searching arbitrary text across the whole file. JPEG markers, PNG chunks and WebP RIFF chunks are bounds-checked before their metadata is classified. This lets the report distinguish EXIF, XMP, IPTC, PNG text and recognized C2PA storage without sending the image to a server.
A C2PA container can be located without claiming that every signature is valid. The result therefore distinguishes a recognized container, no supported container, and an unreadable or unsupported payload. It never turns “not found” into a claim that the image has no AI watermark.
Quick Clean
Lossless container editing with a rights-preserving default
- 01
Classify each block
The tool separates privacy fields, AI-generation fields, Content Credential containers and recognized rights information.
- 02
Preserve image-critical data
Orientation, ICC profiles, transparency and animation chunks stay in place. Pixel-bearing JPEG scan data, PNG IDAT data and WebP frame data are copied unchanged.
- 03
Fail conservatively
If a mixed XMP or IPTC block contains recognized rights information, the default mode keeps the whole block and reports that limitation instead of risking an unauthorized deletion.
- 04
Rescan the output
The clean report compares the actual output with the original and lists remaining metadata, file-size change, dimensions and whether pixels were re-encoded.
Deep Clean
Why pixel regeneration is a different paid task
Deep Clean is a separate server task
After sign-in, the server verifies the image dimensions and quote. Free center-crop previews are limited to 3 attempts per account per UTC day and one per file, subject to the site’s daily processing budget. Failed or unconfirmed attempts count toward this allowance. A paid task debits the displayed quote before processing, refunds technical failures, validates the output, and keeps the same source available for one free retry for 24 hours. Deep Clean is unavailable in the EEA while the required provenance review is completed.
Deep Clean is an AI-assisted regeneration rather than a metadata deletion. It may reduce supported embedded signals, but output can change texture, fine detail, typography or identity. Results vary by source image and by the third-party method used to inspect it. A completed regeneration is not proof that any detector changed its result.
FAQ
Technical boundaries
Why is metadata cleanup different from pixel processing?
Metadata is stored in file containers and can be removed without decoding image pixels. Pixel-level signals are part of the image data and require a separate image transformation.
Does the scan verify an invisible pixel watermark?
No. The scan reports supported file metadata and C2PA containers. It cannot infer a pixel-level watermark from metadata alone.
Can text cleanup remove an AI text watermark?
Text Clean removes supported hidden characters such as zero-width markers, Unicode tags and bidi controls that can appear in AI chat output. A statistical token-level watermark such as the one described for Claude is not an inserted character, so character cleanup cannot detect or prove its removal.
Why are rights fields kept by default?
Author, copyright, licence, attribution and rights-holder information can carry legal meaning. The default cleaner separates them from privacy and AI-generation fields.
Does Quick Clean change image quality?
The implemented cleaner rewrites container metadata without re-encoding pixel payloads. The report explicitly says that pixels were not re-encoded.