Face Anonymization Made Simple is a cutting-edge implementation developed for the WACV 2025 conference. This project provides a robust solution for face anonymization, ensuring original facial expressions, head positions, eye directions, and background elements are preserved while effectively masking identities. The result is a seamless blend of anonymized faces within the original photographs, ideal for various real-world applications.

Key Features

• Preserve Identity Elements: The technique maintains crucial details of the original image while anonymizing faces, ensuring that the background and facial expressions remain intact.

• Versatile Usability: Whether you have a single aligned face or multiple unaligned faces, our method adapts effectively to your needs.

• Easy Integration: Built upon the popular Diffusers framework, our project enables straightforward integration into your existing workflows.

Usage Examples

Anonymizing Images with a Single Aligned Face

An anonymized version of an image can be created using the following code snippet. The image should contain a single face aligned similarly to those in the FFHQ or CelebA-HQ datasets:

# get an input image for anonymization
original_image = load_image("my_dataset/14795.png")

# generate an image that anonymizes faces
anon_image = pipe(
    source_image=original_image,
    conditioning_image=original_image,
    num_inference_steps=200,
    guidance_scale=4.0,
    generator=generator,
    anonymization_degree=1.25,
).images[0]
anon_image.save("anon.png")

Anonymizing Images with Unaligned Faces

To create an anonymized version for images with one or more unaligned faces, use this approach:

import face_alignment
from utils.anonymize_faces_in_image import anonymize_faces_in_image

# get an input image for anonymization
original_image = load_image("my_dataset/friends.jpg")

# SFD (likely best results, but slower)
fa = face_alignment.FaceAlignment(
    face_alignment.LandmarksType.TWO_D, face_detector="sfd"
)

# generate an image that anonymizes faces
anon_image = anonymize_faces_in_image(
    image=original_image,
    face_alignment=fa,
    pipe=pipe,
    generator=generator,
    face_image_size=512,
    num_inference_steps=25,
    guidance_scale=4.0,
    anonymization_degree=1.25,
)
anon_image.save("anon.png")

Swapping Faces Between Two Images

The tool can also swap faces between two images with ease:

# get source and conditioning (driving) images for face swap
source_image = load_image("my_dataset/00482.png")
conditioning_image = load_image("my_dataset/14795.png")

# generate an image that swaps faces
swap_image = pipe(
    source_image=source_image,
    conditioning_image=conditioning_image,
    num_inference_steps=200,
    guidance_scale=4.0,
    generator=generator,
    anonymization_degree=0.0,
).images[0]
swap_image.save("swap.png")

Additional Resources

For further exploration, we provide a detailed demo.ipynb, which walks you through these functionalities step-by-step.

Acknowledgements

This project is built upon the foundations of the Diffusers framework. The face extractor functionality has been adapted from DeepFaceLab.

Explore the capabilities of face anonymization and experience a versatile approach to privacy and data protection in images.