Generative Adversarial Networks in Cosmetic and Medical Image Simulation: A Survey on Hyperparameter Tuning with Metaheuristic Algorithms

Generative Adversarial Networks (GANs) have revolutionized the synthesis of photorealistic facial and skin images, enabling unprecedented applications in cosmetic treatment preview and medical simulation. Despite remarkable progress in architectures such as Pix۲Pix, CycleGAN, and StyleGAN, the extreme sensitivity of GAN training to hyperparameters remains a critical bottleneck that often results in mode collapse, vanishing gradients, or poor visual fidelity. Recent research has increasingly turned to metaheuristic optimization algorithms (Particle Swarm Optimization, Ant Colony Optimization, Genetic Algorithms, and their hybrids) to automatically discover robust hyperparameter configurations. This paper presents the first systematic survey dedicated to metaheuristic-driven hyperparameter tuning of GANs for cosmetic and dermatological image synthesis. Through rigorous analysis of state-of-the-art works, we categorize approaches, compare quantitative improvements in Fréchet Inception Distance (FID), perceptual quality, and training stability, and identify persistent challenges including computational cost, fairness across skin phototypes, and ethical implications of synthetic medical imagery.

کلیدواژه ها:

Generative Adversarial Networks ، Metaheuristic Optimization ، Hyperparameter Tuning ، Cosmetic Simulation ، Facial Image Synthesis ، Ant Colony Optimization ، Particle Swarm Optimization ، Fairness in AI