AnonNET: A Unified Framework for Expression-Consistent Anonymization in Talking Head Videos

Anil Egin^1,2, Andrea Tangherloni², Antitza Dantcheva^1
1Inria Méditerranée, ²Bocconi University

AnonNET is a multi-stage face anonymization pipeline designed to preserve non-identifying facial attributes (e.g., age, gender, race, and expression) while ensuring robust identity obfuscation in both images and talking-head videos. It combines attribute-conditioned diffusion-based inpainting with landmark-free motion synthesis, making it suitable for real-world and privacy-critical video anonymization applications.

This repository provides the implementation of AnonNET as described in our paper:

📄 Now You See Me, Now You Don’t: A Unified Framework for Expression-Consistent Anonymization in Talking Head Videos
🗣️ Oral Presentation at the IEEE International Conference on Computer Vision (ICCV) 2025, Workshop on Computer Vision for Biometrics, Identity & Behaviour (CV4BIOM), Hawaii, USA.

---

Key Features

• Attribute-aware anonymization
  Guided by facial attribute recognition (age, gender, race, expression)

• Diffusion-based inpainting
  Utilizes Stable Diffusion v1.5 with ControlNet conditioning (e.g., face parsing masks)

• Motion synthesis
  Landmark-free reenactment using LIA or LivePortrait

---

Anonymization Samples

🖼️ Image Anonymization

Here are examples of AnonNET’s image anonymization (Original → Anonymized):

🎥 Video Anonymization

Here are video anonymization examples (Original → Anonymized):

Sample 1: Original → Source → Anonymized

Sample 2: Original → Source → Anonymized

---

🛠 Installation

Requires: Python 3.9 and CUDA >= 12.1 (GPU required for inpainting and motion synthesis)

git clone https://github.com/anilegin/AnonNET.git
cd AnonNET
python3.9 -m venv AnonNET
source AnonNET/bin/activate  # On Windows use: AnonNET\Scripts\activate
pip install -r requirements.txt

---

Extra Dependencies

Manually download or cache models for:

vox.pt – Required for LIA motion synthesis
Download from the Releases page and place under Generation/pretrained_weights.

Other model weights (Stable Diffusion, RetinaFace, etc.) are expected to be downloaded and stored automatically when the script is initalized.

---

📸 Image Anonymization

python image_anonymize.py \
    --image path/to/input.jpg \
    --segment face \
    --prompt "A photorealistic portrait of a middle-aged Asian woman with a neutral expression" \
    --strength 0.9 0.4 0.3 \
    --save_folder results/

Parameters

Argument	Description
--image	Input image path
--segment	Type of mask: face or head
--prompt	(Optional) Otherwise attribute-aware one will be generated
--strength	Mask, lineart, and openpose guidance strengths
--steps	Denosing steps
--seed	Random seed (optional)
--save_folder	Output folder for anonymized image

---

🎬 Video Anonymization

python anonymize.py \
    --driving_path path/to/video.mp4 \
    --motion lp \
    --segment face \
    --save_folder results/

Parameters

Argument	Description
--driving_path	Path to input video
--motion	lp for LivePortrait, lia for Latent Image Animator
--segment	face or head segmentation
--max_len	Max clip length (optional)
--save_folder	Output folder
--no_stitch	No eye-lip retargeting for LivePortrait

---

Contact

For questions or feedback, please contact: [email protected]

📃 License

This repository is released under the MIT License.

The code of InsightFace is released under the MIT License. The models of InsightFace are for non-commercial research purposes only.

If you want to use the AnonNET project for commercial purposes, you should remove and replace InsightFace’s detection models to fully comply with the MIT license.