Disentangled Representation Learning for Realistic and Diverse Child Face Prediction from Parent Images

Abstract

Predicting a child’s facial appearance from their parents’ photos is a challenging task with potential applications in various fields, including kinship verification, age progression, and forensic investigations. Existing methods often struggle to balance the need for accurate genetic representation with the generation of diverse and realistic child faces. We propose a novel approach that leverages a Generative Adversarial Net- work (GAN) framework with factor-based disentanglement and mapping, trained exclusively on a family-focused dataset. Our model explicitly separates and represents distinct facial factors: genetic (inherited traits), external (changeable attributes), and variety (individual differences). By focusing on genetic factors and employing a dedicated mapping module to learn parent-to- child genetic relationships, we aim to achieve higher accuracy and realism compared to traditional style-based or direct mapping methods. Comprehensive experiments on a large-scale Family Face Database demonstrate that our model outperforms existing state-of-the-art approaches in generating realistic and diverse child face images. The predicted faces not only capture the nuanced resemblance between parents and children but also exhibit a wide range of individual variations, aligning with real- world observations. Additionally, our method addresses ethical concerns by focusing on heritable traits and utilizing family- specific data, promoting privacy and minimizing potential biases. This work opens up new possibilities for child face prediction, offering a more accurate and ethically sound approach for future research and applications.