With the continuous scaling of integrated circuit (IC) process nodes, the impact of hotspot issues in lithography layouts on chip performance and reliability has become increasingly significant. Traditional deep learning-based lithography hotspot detection methods struggle to meet the precision and model generalization requirements of advanced IC manufacturing. To address this, we propose a deep learning-based $ k $-nearest vision Transformer (kNViT) model for precise detection of lithography hotspots. The model employs a ContraNorm module (CNM) and a $ k $-nearest neighbor attention module (kNAM) to enhance feature representation and identification accuracy. Additionally, a photolithography layout diffusion model (PLDM) is employed to generate images, significantly improving data diversity. Furthermore, we introduce a circuit-aware loss function (CALF) to optimize the performance of the PLDM in predicting noise. Data augmentation strategies, including rotation and contrast adjustment, are applied to further enhance the model’s generalization capability. Experimental results demonstrate that the kNViT model exhibits high-precision hotspot detection performance across multiple lithography layout datasets. On the ICCAD 2012 dataset, it achieves an average recall rate of 99.7%, accuracy of 98%, precision of 90.9%, and F1 score of 95%. The research indicates that the kNViT model can serve as an auxiliary detection tool, effectively improving detection accuracy and efficiency, with potential for industrial design applications.