AI Watermark Remover

Q: Why are some invisible characters marked "kept"?

Because they are doing a job. 👨👩👧 relies on zero-width joiners to fuse three people into one family, ❤️ needs a variation selector to render in red, and Persian and Hindi depend on zero-width non-joiners for correct word shaping. Stripping those blind genuinely corrupts content, so they are preserved by default. You can turn the protection off.

Find and remove the invisible marks and metadata hiding in your text, images and documents. Everything runs in your own browser — no file is ever uploaded.

Nothing is uploaded — everything runs in your browser

AI text watermark & invisible character remover

Find and strip zero-width characters, tag characters, bidi controls and odd spaces. A visual preview lets you actually see what was invisible.

AI text watermark · zero-width detection

Image metadata & C2PA remover

Strip EXIF, XMP, C2PA content credentials and generator fields from PNG and JPEG, while optionally keeping the ICC profile so colours don't shift.

C2PA removal · EXIF stripper

Document metadata remover

Clear author, company, last-modified-by and editing time from DOCX, ODT and PDF. Worth a pass before sending a CV or a deliverable.

Word metadata · PDF properties

AI watermark detector

Read-only. Lists the metadata, C2PA manifests and AI traces a file actually contains, so you can decide before changing anything.

AI watermark detection · metadata viewer

"AI watermark" covers several unrelated technologies with wildly different strength. Some are invisible characters injected into text — delete them and they are gone. Some are generator fields and C2PA content credentials written into file metadata — strip them and you can verify the result field by field. Others are embedded in image pixels or in the probability distribution of token sampling, where no ordinary tool can reach.

These tools handle only the parts that can actually be done and verified: invisible characters, homoglyph spaces, bidirectional controls, and the metadata layer of images and documents. Every run lists exactly what changed, so you can check for yourself rather than take our word for it.

Where the limits are, we say so plainly. The table below draws the line. Any tool claiming to remove "all AI watermarks" tends not to survive follow-up questions.

What this can and cannot do

Stated up front, so you don't waste a run finding out.

Mark type	Detect	Remove	Verifiable	Notes
Invisible characters (zero-width, tag chars, bidi controls)	✓	✓	✓	Character-level. Removed is removed, and you can check it codepoint by codepoint.
Homoglyph spaces and lookalike letters	✓	✓	✓	No-break space, ideographic space, Cyrillic letters posing as Latin.
Image metadata (EXIF, XMP, generator fields)	✓	✓	✓	Removed chunk by chunk; rescan afterwards to confirm.
C2PA content credentials (the manifest itself)	✓	✓	partial	The in-file manifest goes; a service may still re-link it by content fingerprint. See below.
Document metadata (author, company, editing time)	✓	✓	✓	Property fields in DOCX, ODT and PDF.
Pixel-domain image watermarks (e.g. SynthID for images)	✗	✗	—	Embedded in the pixels, survives cropping and compression. Out of scope, and we don't pretend otherwise.
Statistical text watermarks (e.g. SynthID-Text)	✗	✗	—	Hidden in token sampling probabilities. Only heavy rewriting weakens it, and no public detector exists to verify the result.

When it's worth a pass

Before sending a CVDOCX properties routinely still carry your previous employer's name, a colleague's account, and an "editing time: 3 minutes" that is hard to explain away.

Before posting photosEXIF can hold GPS coordinates accurate to a few metres, camera and lens serial numbers, and exact capture time. Not every platform strips them for you.

Before an academic submissionIn double-blind review, an author name left in the document properties can void the submission outright. It happens every year.

After copying from a web page or PDFCopy-paste drags in zero-width characters and odd spaces that break code, fail form validation and defeat search — none of them visible in your editor.

FAQ

Will this get my text past an AI detector?

No, and we won't claim it will. Removing invisible characters is definite, but that is largely unrelated to what detectors measure — they mostly look at statistical features of writing style. As for genuine statistical watermarks such as SynthID-Text, only heavy rewriting can weaken them, and since vendors publish no detector, nobody can verify removal. Any tool promising to "guarantee passing detection" is not being straight with you.

Are my files uploaded to your server?

No. Everything runs in your browser, and the page keeps working offline once loaded. img.diy has no upload endpoint anywhere on the site — the server only sends static files.

Once C2PA is stripped, is the image anonymous?

Not necessarily. Beyond the in-file manifest, C2PA also allows "soft binding", where a service re-links content to its original record by fingerprint. We can remove the manifest from the file; the record in someone else's database is beyond any local tool.

Why are some invisible characters marked "kept"?

Because they are doing a job. 👨‍👩‍👧 relies on zero-width joiners to fuse three people into one family, ❤️ needs a variation selector to render in red, and Persian and Hindi depend on zero-width non-joiners for correct word shaping. Stripping those blind genuinely corrupts content, so they are preserved by default. You can turn the protection off.

Could cleaning damage my file?

Images and documents only lose metadata blocks; image data and body text are untouched byte for byte. DOCX cleaning also updates content-type and relationship declarations so Word is not left with dangling references, and ODT keeps its mimetype entry first and uncompressed. Options marked "destructive" (homoglyph folding, NFKC) are off by default because they do change visible content.

Which formats are supported?

Text can be pasted directly. Images: PNG and JPEG. Documents: DOCX, ODT, PDF, SVG, HTML, Markdown and plain text. PDF cleaning rewrites the document info dictionary and drops the XMP metadata stream.