The rapid evolution of deep learning has revolutionized image denoising, addressing noise-related challenges in diverse real-world applications. However, the reliance on noisy-clean image pairs for supervised methods remains a significant barrier, driving interest in self-supervised techniques that operate without clean data. This survey comprehensively reviews self-supervised image denoising methods, categorizing them into general methods, blind spot network (BSN)-based methods, and transformer-based methods. Each category is analyzed in detail, focusing on principles, algorithmic workflows, strengths, and limitations. Comparative analyses across widely adopted datasets, including BSD68, DND, SIDD, and FMDD, reveal that transformer-based methods achieve state-of-the-art PSNR (e.g., 31.2 on FMDD) and SSIM scores (e.g., 0.96 on SIDD) for complex noise patterns, outperforming BSN-based methods that excel in handling spatially independent noise. General methods remain effective for synthetic Gaussian noise but are limited in adapting to diverse real-world scenarios. The results underscore the trade-offs between computational efficiency and denoising performance across methods. Additionally, challenges such as optimizing resource-intensive transformer models, balancing local–global feature extraction, and handling diverse noise patterns are identified. Promising directions for future research include hybrid approaches combining BSN and transformer principles, adaptive noise modeling, efficient transformer architectures, and multi-task learning that integrates denoising with downstream tasks. By combining in-depth evaluations and actionable insights, this survey provides a valuable resource for advancing self-supervised image denoising.