Watermark Studio · Gemini, Veo, NotebookLM & SynthID Cleaner

Free, air-gapped, and open-source studio for eradicating Google Gemini and Google Veo visible watermarks, scrubbing NotebookLM slide deck logos from PDFs and presentation slides, removing Gemini Notebook logos from videos, disrupting DeepMind SynthID latent watermarks, and stripping C2PA Content Credentials from photos and videos with lossless audio preservation.

---

Key Highlights

Most generic watermark tools rely on heavy neural diffusion inpainting that smears background artwork, leaves visible blur halos, strips video audio tracks, and ignores imperceptible AI provenance tracking.

Watermark Studio solves this with a cleanroom, mathematically calibrated pipeline:

1. Zero-Artifact Watermark Eradication:

   • Discrete Resolution Catalogs: Exact canonical pixel priors for Google Gemini images (0.5k, 1k, 2k, 4k) and Google Veo videos (1080p, 720p portrait and landscape).
   • Veo Text Logo Templates: Multi-scale binary template matching for rectangular "Veo" marks (23x10, 68x30, 99x43).
   • Safety Dilation Inpainting: Morphological (7, 7) dilation plus a 6px safety perimeter completely eliminates watermark star tips and anti-aliasing glow without edge bleeding.

2. NotebookLM Presentation & Video Cleaner (PDFs, Slides & Video):

   • Calibrated Geometry Engine: Precision bottom-right detection targeting NotebookLM exported slides and portrait overview videos (widthRatio: 0.0825 for landscape slides, 0.28 for portrait videos, with configurable scan regions).
   • Boundary-Feathered Gradient Patching: Mathematical source patch sampling with smooth gradient blending — completely eliminates blurring and preserves crisp slide borders, cards, and solid gradients.
   • Triple Execution Workbenches: In-browser client-side presentation cleaner powered by pdf.js and jspdf for zero-upload multi-page PDF processing, alongside a forensic Python laboratory for slide images and audio overview MP4 videos.

3. Google DeepMind SynthID Disruption & Phase Scrambling:

   • SynthID embeds imperceptible pseudo-random frequency perturbations into generated pixels.
   • Nuclear Mode: Applies 0.997 asymmetric Lanczos rescaling, 2px border uncoupling crop, subtle channel bias, and spatial-frequency micro-dithering to de-synchronize the sub-LSB phase trees required by SynthID verification models while maintaining visually pristine fidelity.

4. Lossless C2PA & Provenance Sanitization:

   • Low-level byte chunk walkers for PNG (iter_png_chunks) and JPEG (iter_jpeg_segments).
   • Audits and scrubs 12 container fingerprint categories: C2PA manifests (caBX chunks, APP11 JUMBF boxes), AI prompts (Stable Diffusion Automatic1111, ComfyUI node graphs, Midjourney flags, DALL-E/OpenAI tags), IPTC DigitalSourceType = trainedAlgorithmicMedia, camera EXIF/GPS, Photoshop IRB blocks, and trailing payloads.
   • Safe Mode: 100% lossless container sanitation — input and output pixels are bit-identical (pixel delta = 0, verified by SHA-256).

5. Lossless Audio Passthrough:

   • Uses FFmpeg stream copying (-c:v libx264 -c:a copy) to preserve original AAC, MP3, and PCM audio tracks without re-encoding, volume loss, or sync drift. Tested and verified on NotebookLM overview videos like How_to_Scale_PSI_for_Data_Drift.mp4.

6. 100% Air-Gapped & Private:

   • Zero telemetry, zero external APIs, zero database storage. All processing occurs in temporary, ephemeral memory on your local machine.

---

Cleaning & Sanitization Tiers

Tier	Name	Target Vectors	Visual Impact	Pixel Guarantee
none	Off	Visible watermark only	Imperceptible	Inpainting restricted strictly to logo bounding box
safe	Lossless C2PA & Metadata Strip	C2PA manifests, AI prompts (A1111, ComfyUI, MJ), EXIF, GPS	None	100% Bit-Identical Pixels ($\Delta = 0$)
paranoid	Quantization & Sensor Reset	safe + JPEG quantization table fingerprints, camera PRNU	Imperceptible	Standardized sRGB re-encode + Gaussian micro-dither ($\sigma=0.5$)
nuclear	SynthID & Latent Frequency Disruption	paranoid + Google DeepMind SynthID sub-LSB phase trees	Imperceptible	0.997 Lanczos scale + 2px border uncoupling crop + spatial frequency dither

---

Technical Comparison Matrix

Capability	Generic Online Removers	Reference Catalogs	Watermark Studio
Watermark Matching	Continuous float heuristics	Discrete size catalogs	Discrete 0.5k-4k catalogs + continuous fallbacks
Veo Video Presets	Static fixed offsets	Discrete video catalog	Veo Inset (144), Standard (108), Compact (40), Corner, Text Logo
Veo Text Watermark	Unsupported	Binary base64 templates	Multi-scale template matching (23x10, 68x30, 99x43)
NotebookLM Slide & Video	Unsupported	External third-party scripts	Triple engine: In-browser PDF cleaner + Python image patcher + Video cleaner
Star Tip Clipping	Halo / smudged tips remain	Alpha unblending	Morphological (7, 7) dilation + 6px safety padding (100% clean)
Audio Pipeline	Stripped or lossy re-encode	OpenCV standard	Lossless FFmpeg stream copying (-c:a copy)
Google SynthID Disruption	Not supported	Concept only	0.997 Lanczos phase scrambling + frequency dithering
C2PA / Provenance Stripping	Incomplete / Re-compressed	Chunk walker	Low-level byte parser + 12 finding categories + Bit-identical Safe tier
Privacy & Telemetry	Cloud upload / data retention	Local script	100% local, ephemeral RAM execution, zero tracking

---

Quick Start (Local Run)

1. Clone & Install Dependencies

git clone https://github.com/developerOSindia/watermark-remover-studio.git
cd watermark-remover-studio/developeros-watermark-studio

# Install Python requirements
python3 -m pip install -r requirements.txt

2. Launch Streamlit Studio

streamlit run app.py

Navigate to http://localhost:8501 in your browser.

Use the Workspace Selector in the sidebar to toggle between:

• ✦ Gemini & Veo (Images & Video): Eradicate 4-pointed stars, Veo text badges, disrupt SynthID, and sanitize C2PA manifests with lossless audio passthrough.
• 📑 NotebookLM Studio (PDFs, Slides & Video): Clean entire multi-page PDF slide decks directly in your browser with zero upload, test slide images, or clean full NotebookLM audio overview videos with lossless audio.

Tip

The studio uses system ffmpeg when available and falls back to bundled imageio-ffmpeg. Verify FFmpeg is active with ffmpeg -version.

---

Command-Line Interface (CLI)

The underlying processing engine (remove_watermark.py) can be executed directly from your terminal or integrated into automated shell pipelines.

1. Basic Watermark Removal

# Clean an image
python3 remove_watermark.py input.png output.png

# Clean a video with lossless audio preservation
python3 remove_watermark.py input.mp4 output.mp4

2. Deep Container Inspection (--inspect)

Audit any image for C2PA manifests, AI generation prompts, and EXIF leaks:

python3 remove_watermark.py sample.png --inspect

Example Output:

======================================================================
[INSPECTION REPORT]
File: sample.png (Format: PNG)
Findings: 3
  • [exif] EXIF metadata segment present (size: 142 bytes)
  • [software] Creation tool signature detected: Software
  • [date_time] Timestamp tag present: DateTime
Clean status: UNCLEAN
======================================================================

3. Google SynthID Disruption & Metadata Scrubbing (--synthid-mode)

# Safe mode: Bit-identical pixels with C2PA and prompts stripped
python3 remove_watermark.py input.png cleaned_safe.png --synthid-mode safe

# Nuclear mode: Disrupt DeepMind SynthID latent watermarks
python3 remove_watermark.py input.png cleaned_nuclear.png --synthid-mode nuclear --inspect

4. Advanced Video Options

python3 remove_watermark.py input.mp4 cleaned.mp4 \
    --preset veo_inset \
    --method inpaint \
    --gain 0.60 \
    --size-scale 1.05

CLI Flags Reference: Gemini & Veo CLI Flags Reference:

• --method {inpaint, reconstruct, math}: Inpainting algorithm (default: inpaint).
• --preset {auto, veo_inset, veo_standard, veo_compact, corner, veo_text}: Watermark geometry preset.
• --synthid-mode {none, safe, paranoid, nuclear}: Provenance & invisible watermark cleaning tier.
• --strip-metadata: Enable lossless metadata stripping.
• --inspect: Display detailed JSON/table container audit report.
• --gain FLOAT: Watermark opacity sensitivity (default: 0.60).
• --size-scale FLOAT: Bounding box dilation factor (default: 1.00).

5. NotebookLM Slide & Presentation Deck Cleaning (notebooklm_cleaner.py)

Scrub NotebookLM logos from exported slides, slide captures, and presentation decks:

# Clean a single presentation slide image using boundary-feathered gradient patching
python3 notebooklm_cleaner.py slide.png cleaned_slide.png

# Clean using OpenCV Telea inpainting on complex or textured presentation backgrounds
python3 notebooklm_cleaner.py slide.png cleaned_slide.png --method inpaint

# Batch process an entire directory of presentation slides
python3 notebooklm_cleaner.py ./exported_slides/ ./cleaned_deck/ --method patch

# Fine-tune the bottom-right corner scan region
python3 notebooklm_cleaner.py slide.png cleaned_slide.png --scan-w 0.22 --scan-h 0.08

NotebookLM Flags Reference:

• input: Path to input image file or directory containing slides.
• output: Optional output image path or output directory (defaults to _cleaned suffix).
• --method {patch, inpaint, both}: Removal technique (patch for smooth gradients, inpaint for textures, both to generate comparison outputs).
• --scan-w FLOAT: Bottom-right width scan ratio (default: 0.22).
• --scan-h FLOAT: Bottom-right height scan ratio (default: 0.08).

---

Python API Integration

Integrate Watermark Studio directly into your Python scripts or data pipelines:

from PIL import Image
from remove_watermark import remove_watermark, remove_watermark_from_video
from fingerprint_cleaner import inspect_image_fingerprints, clean_image_fingerprints
from notebooklm_cleaner import clean_notebooklm_image

# 1. Audit image for C2PA and AI provenance
with open("sample.png", "rb") as f:
    report = inspect_image_fingerprints(f.read(), "sample.png")
print(f"Detected {report['finding_count']} metadata items. Has C2PA: {report['has_c2pa']}")

# 2. Remove visible watermark and disrupt SynthID in one call
image = Image.open("sample.png")
cleaned = remove_watermark(
    image,
    gain=0.60,
    size_scale=1.00,
    method="inpaint",
    synthid_mode="nuclear",  # Disrupts SynthID frequency trees
)
cleaned.save("sample_sanitized.png")

# 3. Clean video with lossless audio passthrough
remove_watermark_from_video(
    "input.mp4",
    "output.mp4",
    preset="veo_inset",
    method="inpaint",
)

# 4. Clean NotebookLM presentation slide with mathematical gradient patch
slide = Image.open("notebooklm_slide.png")
clean_slide = clean_notebooklm_image(slide, method="gradient_patch")
clean_slide.save("cleaned_slide.png")

---

Project Structure

developeros-watermark-studio/
├── app.py                         Streamlit Studio web interface with Workspace switcher
├── remove_watermark.py            Core Gemini & Veo removal pipeline + CLI
├── fingerprint_cleaner.py         SynthID disrupter & 12-point C2PA container walker
├── notebooklm_cleaner.py          NotebookLM bottom-right slide & presentation cleaner CLI
├── DESIGN.md                      Semantic Design System (Darkroom Workbench)
├── requirements.txt               Python dependencies
├── packages.txt                   Streamlit cloud debian packages (ffmpeg)
├── assets/
│   ├── bg_48.png                  Small watermark mask template
│   └── bg_96.png                  Large watermark mask template
└── static/
    ├── robots.txt                 Crawler directives (GPTBot, PerplexityBot, ClaudeBot)
    ├── sitemap.xml                Search engine sitemap
    └── notebooklm_embedded.html   In-browser client-side NotebookLM Studio app (pdf.js + jspdf)

---

Privacy & Security

Watermark Studio is engineered with strict privacy principles:

• No Telemetry or Tracking: Zero Google Analytics, cookies, or logging beacons.
• Air-Gapped Processing: No media files are sent to external cloud APIs or third-party servers.
• Ephemeral Storage: Processing executes in RAM and temporary system files which are deleted immediately upon session termination.

---

License & Credits

MIT License. Built with precision by DeveloperOS.