Multi-Domain Lifespan Face Age Synthesis Using StarGAN v2 with Attention-Based Identity Preservation and Real-ESRGAN Enhancement

Abstract

Lifespan face age synthesis aims to transform a source face into plausible appearances at multiple target ages while retaining identity-defining structure. The problem is difficult because facial aging is neither a uniform texture operation nor a simple change of a single latent variable. Bone structure, facial proportions, skin texture, hairline, pigmentation and soft-tissue distribution change at different rates, while identity cues must remain stable. This paper presents AgeStarNet, a multi-domain framework that combines the scalability of StarGAN v2 with an Attention-Based Identity Preservation Module (AIPM) and a Real-ESRGAN enhancement stage. The proposed design treats aging as a controlled translation problem: a style pathway specifies the target age domain, while an ArcFace-derived identity representation provides an explicit anchor for identity-sensitive features. The AIPM uses eight-head cross-attention so that intermediate generator features can query source identity information before image synthesis is completed. A multi-component objective combines adversarial, style-reconstruction, diversity, cycle-consistency, identity and age-domain terms. A separate Real-ESRGAN stage converts the 64×64 synthesis into a 256×256 output and is fine-tuned on age-synthesized pairs. The experimental reports results on MORPH-II, UTKFace, CACD and FG-NET with FID, identity cosine similarity, age-estimation accuracy and SSIM as complementary measures. The reported UTKFace values are 12.4 FID, 0.913 identity cosine similarity, and 88.7% age-estimation accuracy. The core methodological idea is a bottom-up and open research-generation paradigm where evidence is accumulated from dataset construction, representation learning, synthesis, identity-preservation, enhancement, evaluation and reproducibility rather than considering the final image as the only research artefact.