This paper explores deep learning-based translation of iris images from the visible (VIS) to the near-infrared (NIR) spectrum, aiming to combine the robustness of NIR-based recognition–less sensitive to illumination and reflections–with the accessibility and low cost of VIS imaging, including mobile devices. Unlike existing approaches that focus primarily on visual quality, this study emphasizes the preservation of biometric discriminability. A key innovation is the integration of a dedicated feature-preserving module, LogGaborNet that guides the generation of NIR-like images by enforcing retention of discriminative features inspired by Daugman’s log-Gabor filters. Two models are proposed: a U-Net autoencoder and a Pix2Pix conditional GAN (cGAN), both extended with LogGaborNet to ensure that the synthesized images are not only visually convincing but also biometrically meaningful. Experiments show that the U-Net-based model notably improves segmentation performance, while the cGAN-based translation module excels in recognition accuracy. These results demonstrate the potential of feature-aware VIS-to-NIR translation models to enable reliable iris recognition even in VIS-only acquisition scenarios.