FaceAge+: training-free synthetic augmentation for deep learning-based biological age estimation from facial images

Abstract Accurate biological age estimation from facial images has emerged as a promising non-invasive marker of human aging. It is potentially relevant to health assessment and geriatric research. However, progress in this field is constrained by a severe age imbalance in publicly available datasets, leading to poor representation of children and older adults. This limitation causes unreliable aging predictions in populations where accurate estimates are clinically relevant. In this work, we present FaceAge+, a data augmentation based method, that addresses age imbalance without requiring additional data collection or end-to-end model retraining. The proposed approach uses a novel training-free, age-conditioned face synthesis technique based on geometrically guided W-space traversal of StyleGAN2, enabling the disentangled generation of synthetic faces across target age groups. We incorporate a Vision Language Model-based perceptual face filter and self-supervised feature extraction using the state-of-the-art DINOv3 model. Extensive experiments on two datasets demonstrate that our method outperforms recent approaches, achieving mean absolute errors of 3.27 years on UTKFace and 2.20 years on IMDb-Wiki. This consistently improves predictions for underrepresented age groups, particularly individuals aged 60 and above. Age-stratified evaluation shows reductions of up to 25% in error for elderly cohorts compared to models trained on the original data. Our code is available at https://github.com/qaixerabbas/faceage-plus .