Remove Google SynthID-Text watermark without changing visible text
FuckMark is a small retokenizer for SynthID-Text. It weaves U+034F (combining grapheme joiner) and U+FE00 (variation selectors) into English text so Google's watermark detector stops firing, while the letters you see stay the same.
On the public Google synthid-text 30-key GPT-2 eval, marked samples score 188/192 before and 0/192 after. Stripping Mn combining marks restores the watermark. Cap in that test is 192 tokens.
- Visible text unchanged
- English ASCII remains English ASCII
- Not paraphrase, not translation
Site: https://mark.q1z.org
GitHub: https://github.com/byte271/FuckMark