In light of the limited sample data of insulator defects in transmission lines and the issue of generated images being prone to blurring and structural distortion in detail expression, an unconditional insulator image generation method based on the denoising diffusion probability model (DDPM) is proposed. Building upon the traditional DDPM framework and addressing the limitations of the UNet structure in high-precision image feature extraction, as well as the geometric distortion caused by information dilution in skip connections, this method introduces a backbone network combining residual blocks and a linear attention mechanism, while optimizing the skip connection path through multi-scale attention. This significantly enhances the model’s capacity and quality in capturing complex textures and fine-grained defects, achieving dynamic feature alignment and maintaining structural consistency. Furthermore, this paper constructs a dual-encoder backbone network (PVTv2) based on the fusion of convolutional neural networks (CNNs) and transformers, which effectively strengthens the joint modeling capability of local features and global semantics, improving the model’s robustness under complex background interference. Simultaneously, a parallel double cross-attention mechanism (DCA) is introduced to enhance the semantic interaction between long-range dependencies of multi-scale features and local small targets, thereby enriching contextual information throughout the generation process. Experimental results show that the IS and FID values of the improved model on the transmission line insulator defect dataset significantly outperform those of the traditional DDPM model, thereby validating the effectiveness and superiority of the proposed method.