Scrub

Local, offline hygiene for media and documents you own: EXIF / C2PA / Content Credentials, hidden Unicode in text, plus research-style tools for studying invisible watermark robustness (including SynthID-class mid-band carriers discussed in public literature). Nothing leaves your machine.

Most “AI mark” tools stop at clipboard Unicode or a PDF tag. Scrub covers that plus pixels, video, and audio.

Layer	What it actually does	Honest limit
Text	Strip zero-width / bidi / Unicode tag chars and space homoglyphs	Does not paraphrase away statistical token-sampling marks
Documents	Core/app props, HTML generator meta, SVG <metadata>, PDF Info/XMP (pypdf or exiftool)	Compressed C2PA in PDF is best-effort
Images	EXIF/IPTC/XMP + best-effort C2PA/JUMBF; optional spectral/spatial disruption; corner-badge inpaint	No detector guarantee
Video / audio	Drop container tags; optional spectral pass on short video; codec/resample chain	Clips >90s skip the spectral pass

Use only on files you have rights to modify. See NOTICE.md. Transforms are best-effort. Product names below are shorthand for publicly discussed signal families.

Quick start

git clone https://github.com/HarshShah0203/Scrub.git
cd Scrub
python3.12 -m venv .venv
.venv/bin/pip install -r requirements.txt

# Inspect (JSON) then write a cleaned *copy*
.venv/bin/python cli.py inspect notes.md
.venv/bin/python cli.py clean notes.md -o ~/Desktop/Scrub
.venv/bin/python cli.py clean photo.jpg -o ~/Desktop/Scrub
.venv/bin/python cli.py clean photo.jpg -o ~/Desktop/Scrub --metadata-only

.venv/bin/python tk_app.py          # native UI
.venv/bin/python app.py             # browser UI → http://127.0.0.1:7860

macOS extras: brew install ffmpeg python@3.12 python-tk@3.12
Linux extras: sudo apt install ffmpeg python3-venv python3-tk
Optional Mac app: .venv/bin/python scripts/build_app.py --install → Spotlight Scrub.

Optional PDF helpers: pip install pypdf (already in requirements.txt) and/or exiftool.

Agent skill (Cursor / Claude Code)

This is the same distribution path people already use for local file hygiene in coding agents:

# Cursor
mkdir -p ~/.cursor/skills
ln -sfn "$(pwd)/skills/scrub" ~/.cursor/skills/scrub

# Claude Code
mkdir -p ~/.claude/skills
ln -sfn "$(pwd)/skills/scrub" ~/.claude/skills/scrub

Then ask the agent to inspect or scrub a file you own. The skill runs python cli.py locally and forbids “undetectable / guaranteed bypass” language.

Programmatic API

from watermark_remover import clean_file

out_path, detail = clean_file(
    input_path="/path/to/image.jpg",
    output_dir="/path/to/out",
    strength="medium",
    use_spectral=True,
    remove_visible=True,
    write_audit=True,
)

Documents and text go through the same clean_file / cli.py clean entry point.

Using the app

1. Add images / video / audio / documents you are allowed to modify (batch OK; HEIC/AVIF supported)
2. Pick an output folder
3. Choose Metadata + signal disruption or Metadata only (documents always get metadata + Unicode hygiene)
4. Defaults favor spectral disruption on media, badge inpaint, and an audit JSON
5. Optional: diffusion regeneration if you install torch + diffusers
6. Start — originals stay put; outputs use a _clean suffix

Signal families (public discussion)

Family	Typical transforms here
Hidden Unicode / bidi / tag chars	Deterministic delete / space fold
Mid-band pixel carriers (e.g. SynthID-class)	FFT dampening, noise, mild JPEG
Frequency / latent-style marks	Resize cycle, crop/pad, FFT
Small corner UI badges	Detect + soft inpaint
Audio spectral masks	Resample, filters, codec swap
Short video	Per-frame spectral pass (≤90s) + ffmpeg re-encode

Limitations

• No cryptographic guarantee; detectors are probabilistic and change over time
• No proprietary codebook or vendor API is used here
• Pixel-identical reconstruction is not a goal
• Statistical text watermarks need a rewrite, which this repo does not ship as a silent “remove mark” button
• Video spectral pass skips clips longer than 90s (ffmpeg chain still runs)
• Only process files you own or have explicit rights to modify

Layout

File	Role
cli.py	Unified inspect / clean
text_hygiene.py	Hidden Unicode (stdlib)
document_strip.py	PDF / DOCX / ODT / HTML / SVG / Markdown
watermark_remover.py	Image / video / audio pipelines
spectral_attack.py	Adaptive FFT carrier disruption
visible_mark.py	Corner-badge inpaint heuristic
c2pa_strip.py	Best-effort C2PA / JUMBF scrub
audit_report.py	Before/after JSON
tk_app.py / app.py	Native + Gradio UIs
skills/scrub/	Agent skill
NOTICE.md	Scope and trademarks

Research lineage

1. Hu et al. (2024), Stable Signature is Unstable — arXiv:2405.07145
2. Broader regeneration / signal-processing watermark-robustness literature
3. Community analyses of mid-band carriers (e.g. Synthid-Bypass)
4. DeepMind SynthID docs (reference only; no affiliation)

Contributing

PRs welcome — detectors, longer-video performance, Windows packaging, tests. See CONTRIBUTING.md.

If Scrub is useful, a star or fork helps others find a local offline option.

License

MIT — see LICENSE. Also read NOTICE.md.