Magnetic resonance imaging (MRI) is an important clinical imaging modality that offers advantages such as the absence of ionizing radiation, high soft-tissue contrast, and multiplanar and multiparametric imaging capabilities. It has been widely used in disease screening, clinical diagnosis, and treatment response assessment. However, MRI acquisition is affected by factors such as scanning conditions and hardware systems, resulting in varying degrees of noise contamination. To address this problem, this paper proposes a self-paced learning (SPL)—driven training-free deep generative method for MR image denoising, termed self-paced deep image prior (SP-DIP). Specifically, built upon the classical deep image prior (DIP) framework, SP-DIP partitions the observed image into patches and dynamically assigns weights to them according to their fitting difficulty during network optimization. Patches with better data consistency and more stable structures are preferentially used to guide parameter updates in the early stages, while patches containing complex textures, fine edges, or noise-sensitive regions are gradually introduced in subsequent iterations, thereby enabling a more robust progressive denoising strategy. Comparative experiments on multiple datasets under various noise levels demonstrate that SP-DIP outperforms traditional iterative filtering methods in terms of denoising accuracy and remains competitive with supervised deep learning methods under high-noise conditions.