Face Aging Using Various Cycle GAN Models: A Comparative Analysis of Efficient Net Mobile Net and Standard Architectures

Abstract

Human face aging is a fundamental problem in computer vision and biometric analysis, with significant applications spanning forensic identification, entertainment and clinical assessment. Standard ResNet-based generators often exhibit mode collapse and insufficient preservation of identity-discriminative features. Moreover, training stability across a large number of epochs presents difficulties, particularly when balancing adversarial loss, cycle-consistency loss, and identity loss simultaneously. This research proposes and evaluates three distinct CycleGAN architectures for face aging such as CycleGAN-EfficientNetB1, CycleGAN MobileNetV2, and CycleGAN. All three models are trained on a dataset of cropped young and old facial images using the Adam optimizer with learning rate 0.0001 and identical loss weightings. Each model employs a shared adversarial training strategy combining MSE-based GAN loss, L1-based cycle-consistency loss with weight lambda=10, and L1-based identity loss with weight lambda=5. The EfficientNetB1 model uses frozen early encoder layers (first 3 feature blocks) producing 24-channel feature maps, while the MobileNetV2 model leverages the first 7 inverted bottleneck blocks producing 32-channel feature maps. The standard model uses a conventional conv-down-sample-resblock-upsample architecture with 9 residual blocks. Training over 200, 152, and 270 epochs respectively, the models achieved final generator losses of 3.69, 1.71, and 1.56. The experimental findings confirm that architectural choice of backbone significantly influences face aging quality in CycleGAN frameworks. The normal CycleGAN achieves the lowest final loss across all metrics, while MobileNet offers computational efficiency with competitive quality. This study establishes a comparative benchmark for CycleGAN-based face aging research.