Head Anonymization with Realistic Vision V5 and ControlNet

This repository provides a head-anonymization workflow that uses:

• head-segmentation for head masks
• SG161222/Realistic_Vision_V5.0_noVAE as the base model
• ControlNet for inpainting
• DeepFace for generating consistent demographic attributes (age, gender, race)

Table of Contents

1. Introduction
2. How It Works
3. Installation
4. Usage
   • Segment the Head
   • Run the Anonymization
   • Optional Prompts
5. Output
6. Model & References
7. License

---

Introduction

This pipeline aims to anonymize faces while preserving certain identity-related features (like approximate age, race, and gender). By analyzing faces with DeepFace, we build a Stable Diffusion prompt that matches the subject’s demographic attributes, ensuring a realistic yet anonymized result by staying above 0.6 cosine distance from the original image.

Example:

Prompt:
A photorealistic portrait of a young adult White man, with a neutral expression, looking directly at the camera. Ultra-detailed, 8k resolution, professional photography.

Sample Image

Mask

Output:

---

How It Works

1. Segment the Head
   We use head-segmentation to generate a mask of the head region in the input image.

2. Analyze Demographics
   DeepFace analyzes the original face to predict:

   • Age
   • Gender
   • Race (dominant race)

3. Generate Prompt
   A custom prompt is built using the analyzed demographic data. For example:

   "A photorealistic portrait of a {age_group} {race_descriptor} {gender}, with a neutral expression,
   looking directly at the camera. Ultra-detailed, 8k resolution, professional photography"
   

4. Inpaint with ControlNet
   ControlNet uses the generated mask and the prompt to inpaint the head region in the style of Realistic Vision V5.

5. Save the Results
   The anonymized image is saved, preserving the original image background but anonymizing the head region.

---

Installation

1. Clone the Repository

   git clone https://github.com/anilegin/AnonHead.git
   cd AnonHead

2. Create a virtual environment (Optional)

   python -m venv anonymized_env
   source ./anonymized_env/bin/activate

3. Install Dependencies
   Make sure you’re in a Python virtual environment, then run:

   pip install -r requirements.txt

   Note: You may need to install system-level dependencies for PyTorch, e.g., CUDA, if you plan to run on GPU.

4. Install Pre-trained weights for Realistic Vision v5.0
   Run:

   python download-weights

---

Usage

Segment the Head

Run the segment script to generate a head mask from your input image:

python segment.py --im_path path/to/your/image.png

• Output: The segmented mask is saved to the ./masks folder by default.

Run the Anonymization

Use the predict script to inpaint the face with a new identity:

python predict.py --image path/to/your/image.png --mask path/to/your/mask.png

• Output: The anonymized image is saved under the ./res folder.

Optional Prompts

You can pass custom prompts if you want to override the defaults:

python predict.py \
   --image path/to/your/image.png \
   --mask path/to/your/mask.png \
   --prompt "A photorealistic portrait of a middle-aged Hispanic woman, neutral expression, looking at the camera" \
   --negative_prompt "ugly, blurry, low-res"

If not provided, the script uses DeepFace to detect age, race, gender, and builds an automated prompt accordingly.

---

Output

• Masks: Saved in ./masks/<image_name>_mask.png
• Anonymized Images: Saved in ./res/<image_name>_out.png

Example structure:

project/
├── masks
│   └── input_image_mask.png
├── res
│   └── input_image_out.png

---

Model & References

• Realistic Vision V5 model from:
  SG161222/Realistic_Vision_V5.0_noVAE
• Head Segmentation from:
  wiktorlazarski/head-segmentation
• DeepFace for demographic analysis:
  DeepFace GitHub

---

License

• Data & Model Usage: Comply with each model’s license (for Realistic Vision, DeepFace, etc.).
• Code: MIT License Copyright (c) 2025 ... Permission is hereby granted, free of charge, to any person obtaining a copy ...

Feel free to open an issue or pull request for improvements and suggestions!