With the rapid advancement of facial recognition (FR) technology, personal privacy protection has become a growing concern. Some studies generate adversarial examples by introducing perturbations into facial images to prevent recognition by malicious FR systems. However, existing methods often struggle with transferability and imperceptibility, limiting their real-world applicability. To address this, we propose a novel method, IA-LDiffProtect, which improves adversarial example generation using latent diffusion models, multi-dimensional gradient guidance, and facial parsing masks. Our method creates gradients through multi-dimensional loss functions, ensuring the generated adversarial examples align with target images in feature space, frequency domain, and perceptual layers, thus improving the success rate of transfer attacks across various FR models. Additionally, facial parsing masks enhance the imperceptibility of adversarial examples by reducing perturbations in key facial areas while concentrating disturbances in less significant regions. Extensive experiments on the CelebA-HQ and FFHQ datasets show that IA-LDiffProtect outperforms existing state-of-the-art methods regarding transferability and imperceptibility.