Infrared small-target detection, a complex and critical task in computer vision, faces numerous challenges—including tiny target sizes, low contrast, severe background noise, and limited data availability. These factors significantly impair detection accuracy and real-time performance. Existing deep learning–based algorithms, which predominantly adopt segmentation paradigms via deep encoder–decoder architectures for generating segmentation masks, often exhibit limited precision in complex scenarios due to inadequate feature representation and learning capabilities. Inspired by the notable success of diffusion models in artificial intelligence, this paper introduces a novel approach by reframing infrared small-target detection as a generative task and proposes a conditional denoising network, termed diff-ISTD. By leveraging the strengths of progressive denoising and image reconstruction, diff-ISTD captures the deep statistical properties of infrared images, enabling more precise identification of weak and ambiguous small-target features. The proposed network consists of conditional branching modules for extracting prior knowledge from infrared inputs and denoising branches for refining noisy segmentation masks. In addition, a parallel dual-dimensional self-attention (PDSA) block is introduced to integrate spatial and channel information, significantly enhancing the model's sensitivity to global structures and local details. This design effectively addresses the challenges of target blurring caused by resolution limitations and environmental variability. Comprehensive experiments demonstrate that, under rigorous detection conditions, diff-ISTD outperforms current state-of-the-art segmentation methods in terms of performance and detection efficiency, offering a promising direction for advancing infrared small-target detection technologies.