Due to the uniqueness and semantic richness of icons, their classification and generation remain challenging. The presence of uncategorized icons further degrades the quality of icon generation, leading to structural inconsistencies and difficulties in style control. To address these issues, we propose MAR-StyleGAN, a style-guided multi-attention residual generative adversarial network for icon generation. Unlike StyleGAN2/3, our method incorporates a multi-attention residual classification stage before generation, enabling more accurate semantic feature learning. Specifically, we integrate attention mechanisms into residual networks, train multiple pre-trained classifiers through transfer learning, and fuse their results to alleviate misclassification in small-sample scenarios. The classified icons are then used to guide StyleGAN training, leading to icons with clearer semantics and more stable style control. Experimental results on the LLD dataset demonstrate that MAR-StyleGAN effectively captures diverse icon features and produces icons of higher quality and distinct styles, showing advantages in both generation stability and diversity.