Face recognition using generative Artificial Intelligence has emerged as a powerful technique for learning robust identity representations under varying illumination, pose, and expression. By leveraging data synthesis and latent space modeling, generative AI enhances generalization and improves recognition accuracy. However, conventional generative-based systems often face limitations such as overfitting, noise sensitivity, and insufficient optimization of deep features, leading to performance degradation in real-world scenarios. To overcome these challenges, this study presents a Face Recognition Framework using Generative AI, integrated with the Maximum Entropy Regularized Decision Transformer and Crayfish Optimization Algorithm (MERDT-CryfOA). Initially, facial data are collected from the Labeled Faces in the Wild (LFW) dataset. It undergoes pre-processing using Adjusted Min–Max with Decimal Scaling and Statistical Column Normalization (AMnMx-DS-SCN) to enhance feature uniformity and eliminate intensity bias. The processed data are then fed into an Improved ResNet-34 Algorithm (KANS-ResNet-34) model for feature extraction, capturing fine-grained and nonlinear facial attributes such as wrinkles, skin tone gradients, and contour transitions. These deep features are subsequently classified by the Entropy Regularized Decision Transformer (MERDT), which employs entropy-regularized policy learning to ensure adaptive, robust decision-making. Finally, the Crayfish Optimization Algorithm (CryfOA) is employed to optimize the classifier’s loss function, balancing exploration and exploitation to achieve faster convergence and reduced misclassification. Experimental results show that the proposed MERDT-CryfOA model achieves a recall of 99.4%, precision of 99.3%, FAR of 1.17%, and FRR of 1.09%, outperforming existing deep learning methods. These results confirm its superior accuracy and robustness under varying conditions, making it an effective framework for next-generation face recognition systems.