Universal AI Watermark Remover

Universal Deep Learning AI Watermark Remover and Provenance Metadata Scrubber.
Developed and Maintained by White Way Web.

Side-by-Side Inpainting Studio

Experience seamless, zero-blur watermark removal powered by Large Mask Inpainting (LaMA) deep neural networks with interactive canvas masking:

Above: Real OpenAI Sora video frame with top-left watermark erased cleanly using LaMA neural synthesis.

Key Features

1. Neural Network Inpainting (LaMA):
   • Replaces blurred OpenCV approximations with LaMA (Large Mask Inpainting) Fourier Convolutions (big-lama.pt).
   • Synthesizes authentic image textures (sky, clouds, fur, backgrounds) with zero smudging.
2. Complete C2PA and Provenance Metadata Scrubbing:
   • Scans raw binary byte streams for b"c2pa" and b"jumb" chunks.
   • Completely strips cryptographic provenance manifests (DALL-E 3, Adobe Firefly, ChatGPT) verified on official contentcredentials.org/verify.
3. AI Text and Document Steganography Sanitizer:
   • Detects and strips invisible Unicode watermarks (zero-width spaces \u200B, joiners \u200D, BOM \uFEFF, bidi overrides \u202A-\u202E, and tag characters).
   • Sanitizes documents (Markdown, Plain Text, DOCX, ODT, HTML).
4. Interactive Side-by-Side Web Studio:
   • Draw precise masks over any watermark with an adjustable brush slider or 1-click Auto Mask.
   • Instant side-by-side preview with download capabilities.
5. Invisible Watermark Neutralization (SynthID):
   • Neutralizes frequency-domain digital watermarks via Fast Fourier Transform (FFT) sub-band filtering.
6. Full Video Inpainting Pipeline:
   • Processes frame sequences with audio pass-through via FFmpeg.

Quickstart and Installation

Option 1: Standard PyPI Installation

# Basic Core Package (includes Text, Documents, and Metadata engines)
pip install universal-ai-watermark-remover

# Full Neural Network and AI Dependencies (PyTorch + LaMA)
pip install "universal-ai-watermark-remover[ai]"

Option 2: Install Directly from GitHub into Any Project

# Basic package
pip install git+https://github.com/whitewayweb/Universal-AI-Watermark-Remover.git

# Full Neural Network and AI Dependencies (PyTorch + LaMA)
pip install "universal-ai-watermark-remover[ai] @ git+https://github.com/whitewayweb/Universal-AI-Watermark-Remover.git"

Or add to your project's requirements.txt:

universal-ai-watermark-remover[ai]

Option 3: Run / Develop from Source

git clone https://github.com/whitewayweb/Universal-AI-Watermark-Remover.git
cd Universal-AI-Watermark-Remover

# Create virtual environment and install in editable mode
python3 -m venv venv
source venv/bin/activate
pip install -e ".[all]"

Launch Interactive Web Studio

Launch the built-in side-by-side studio in your browser with a single command:

universal-ai-watermark-remover web --port 8080

Open http://127.0.0.1:8080 to paint over watermarks and remove them in real-time.

Integrate into Your Python Projects

You can easily embed Universal-AI-Watermark-Remover into your own applications, web backends, or data pipelines:

from ai_watermark_remover import WatermarkRemover

# Initialize unified remover
remover = WatermarkRemover(use_gpu=False)

# 1. Remove visible watermarks with LaMA Neural Inpainting
remover.remove_visible(
    input_path="watermarked_image.jpg",
    output_path="cleaned_image.png",
    manual_boxes=[(25, 45, 240, 150)],  # Optional: (x1, y1, x2, y2)
    method="lama"
)

# 2. Strip C2PA and EXIF provenance metadata (verified on contentcredentials.org)
remover.strip_metadata(
    input_path="dalle3_image.png",
    output_path="c2pa_free_image.png"
)

# 3. Clean AI text (strips zero-width chars, bidi overrides, and tracking tags)
clean_text = remover.clean_text("AI-generated text with hidden \u200B\u200C tracking...")
print(clean_text)

# 4. Sanitize documents (DOCX, Markdown, Plain Text, HTML)
remover.clean_document(
    input_path="ai_draft.docx",
    output_path="sanitized_draft.docx"
)

# 5. Neutralize invisible watermarks (Google SynthID / Steganography)
remover.perturb_invisible(
    input_path="synthid_image.png",
    output_path="clean_synthid.png",
    method="fft"
)

# 6. Process an entire AI video (preserving audio)
remover.process_video(
    input_path="sora_video.mp4",
    output_path="clean_sora_video.mp4",
    detection_skip=5
)

Command Line Interface (CLI)

# 1. Remove visible watermark with LaMA
universal-ai-watermark-remover remove-visible input.png output.png --method lama

# 2. Strip C2PA provenance metadata
universal-ai-watermark-remover strip-metadata input.png output.png

# 3. Clean AI text / Markdown
universal-ai-watermark-remover clean-text draft.md clean_draft.md

# 4. Sanitize documents (DOCX, ODT, HTML, TXT)
universal-ai-watermark-remover clean-doc report.docx clean_report.docx

# 5. Inspect media or text for watermarks
universal-ai-watermark-remover inspect draft.md
universal-ai-watermark-remover inspect photo.jpg

# 6. Neutralize invisible watermarks (SynthID)
universal-ai-watermark-remover perturb-invisible input.png output.png --method fft

# 7. Process video file
universal-ai-watermark-remover process-video input.mp4 output.mp4 --skip 5

Repository Structure

Universal-AI-Watermark-Remover/
├── ai_watermark_remover/
│   ├── core/
│   │   ├── visible.py      # LaMA neural network inpainting & box masking
│   │   ├── metadata.py     # Binary C2PA, JUMBF & EXIF scrubber
│   │   ├── text.py         # Unicode steganography & document cleaner
│   │   ├── invisible.py    # FFT frequency & spatial noise perturber
│   │   └── video.py        # FFmpeg video frame sequence pipeline
│   ├── gui/
│   │   └── server.py       # Side-by-side interactive canvas studio
│   ├── api.py              # Unified WatermarkRemover API class
│   └── cli.py              # Click command-line interface
├── docs/
│   └── images/
│       └── studio_demo.png # Studio screenshot
└── examples/
    ├── sample_images/      # Real Sora, Stock & C2PA samples
    └── sample_documents/   # Real AI Markdown, DOCX & Text samples

License and Attribution

Licensed under the MIT License.
Developed and maintained with pride by White Way Web.