FaceAge: Scalable Facial Age Estimation via Foundation Model Distillation and Ordinal Regression

We present FaceAge, a face-only age and gender estimation model achieving MAE = 3.58 years on LAGENDA 84k — outperforming body-aware MiVOLO v2 (3.86) using only a single face crop and public training data. Built on DINOv3-ViT-L/16 with CORAL ordinal regression and four-phase progressive fine-tuning, FaceAge ranks #1 among non-LLM methods on the eight-dataset FaceAge benchmark (avg MAE 4.90), winning 4 of 8 datasets. Our results show that a strong foundation model with ordinal supervision extracts sufficient aging information from facial appearance alone, eliminating the need for body detection. Weights and ONNX inference code are publicly available.