Mohamed Jaouad Lahgazi and
Jean-Philippe Tarel
Abstract
Privacy regulations such as GDPR mandate face anonymization in visual data, yet existing methods optimize for benchmark performance while ignoring deployment challenges. State-of-the-art approaches achieve strong results on curated datasets of pre-cropped, aligned faces but lack infrastructure for real-world images containing multiple people at varying scales, poses, and occlusions. More critically, these methods process all detected faces indiscriminately, wasting computation on low-quality detections and risking artifacts on barely-visible faces. We address these deployment gaps with a production-ready pipeline combining diffusion-based inpainting with intelligent face filtering and robust preprocessing. Our visibility scoring system determines which faces warrant anonymization using geometric and recognition cues, preventing wasted computation and artifacts. The end-to-end architecture handles extreme scale variations, multiple overlapping people, and diverse conditions through adaptive mask generation with multi-fallback strategies. While achieving competitive performance on standard benchmarks, our approach demonstrates superior generalization to real-world scenarios including group photos, full-body images, and challenging outdoor scenes. The source codes are publicly available at : github.com/MJLahgazi/vis-aware-diffusion-anonymization.
Reference
@InProceedings{ jpt-icpr26,
author = {Lahgazi, M.-J. and Tarel, J.-P.},
title = {Visibility-Aware Diffusion-Based Face Anonymization for Real-World Deployment},
booktitle = {International Conference on Pattern Recognition (ICPR'26), Part VII},
date = {August 17-22},
address = {Lyon, France},
pages = {264-278},
year = {2026},
url = {http://perso.lcpc.fr/tarel.jean-philippe/publis/icpr26.html}
}