Vanish

Automatically detect and blur background faces in images and videos using state-of-the-art AI models.

Vanish combines Florence-2 for intelligent face detection with SAM2 for precise segmentation, enabling privacy-preserving content creation with a single command.

Features

• Smart Speaker Detection — Automatically identifies and preserves main speakers while blurring background faces
• Precise Segmentation — Uses SAM2 for pixel-perfect face masking (no ugly rectangles)
• Image & Video Support — Process single images or entire videos
• Adjustable Pixelation — Control blur intensity with customizable pixel size
• GPU Accelerated — Supports CUDA, MPS (Apple Silicon), and CPU

Quick Start

Installation

# Clone the repository
git clone https://github.com/BradenStitt/Vanish.git
cd Vanish

# Install dependencies
pip install -r requirements.txt

# Install SAM2 (required)
pip install git+https://github.com/facebookresearch/segment-anything-2.git

Usage

Command Line:

# Process an image
python vanish.py input.jpg output.jpg

# Process a video
python vanish.py input.mp4 output.mp4

# Adjust pixelation intensity
python vanish.py input.jpg output.jpg --pixel-size 15

# Blur ALL faces (including main speakers)
python vanish.py input.jpg output.jpg --blur-all

Python API:

from vanish import vanish, vanish_video

# Process image
result = vanish("input.jpg", "output.jpg", pixel_size=10)

# Process video
vanish_video("input.mp4", "output.mp4", pixel_size=10)

# Blur all faces (no speaker exclusion)
result = vanish("input.jpg", "output.jpg", exclude_speakers=False)

Jupyter Notebook:

Open vanish.ipynb for an interactive walkthrough with visualizations.

How It Works

Input Image/Video
        │
        ▼
┌───────────────────┐
│    Florence-2     │  ← Object detection: finds all "human face" instances
│   (Detection)     │  ← Phrase grounding: identifies "main speaker" faces
└───────────────────┘
        │
        ▼
┌───────────────────┐
│  Filter Faces     │  ← Excludes main speakers from blur list
└───────────────────┘
        │
        ▼
┌───────────────────┐
│      SAM2         │  ← Generates precise segmentation masks
│  (Segmentation)   │     from bounding boxes
└───────────────────┘
        │
        ▼
┌───────────────────┐
│   Pixelation      │  ← Applies blur effect only to masked regions
└───────────────────┘
        │
        ▼
   Output Image/Video

Requirements

• Python 3.10+
• PyTorch 2.0+
• CUDA GPU recommended (also works on Apple Silicon MPS or CPU)
• ~4GB VRAM for inference

Use Cases

• Content Creation — Blur bystanders in vlogs and street footage
• Privacy Compliance — Anonymize faces for GDPR/CCPA compliance
• Security Footage — Redact identities while preserving context
• Social Media — Protect privacy of people in background

License

MIT License - see LICENSE for details.

Acknowledgments

• Florence-2 by Microsoft
• SAM2 by Meta AI