Synthetic Faces, Real Gains: Improving Age and Gender Classification through Generative Data

Facial image-based age and gender classification is a foundational problem in computer vision, yet performance is often limited by data scarcity and privacy constraints. This paper proposes IDiff-Face-Aged, a novel age transformation framework that utilizes multimodal embeddings to generate realistic and diverse facial images. Two synthetic datasets, Fading (utilizing null text-inversion prompting) and IDiff-Face-Aged <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>, were tested across various architectures, including MultiEfficientNet, MultiLightViT, and MultiMobileNet. Three age estimation methods were used to address both continuous and categorical intervals. Experimental results indicate that synthetic data contributes to improved model accuracy, especially by enhancing representation in under-sampled age groups. On the UTKFace dataset, an accuracy gain of over $3 \%$ was observed. Moreover, models trained with a mix of synthetic and real data show stronger generalization capabilities, particularly when datasets design and alignment are carefully managed. These findings illustrate the potential of synthetic data to supplement real-world datasets, while also pointing to ongoing challenges related to data realism and artifact avoidance.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Method under Patent Pending Process