Lightweight human pose estimation has long been a research hotspot and challenge in the field. Most existing methods primarily rely on introducing the high-resolution design pattern from HRNet and subsequently performing lightweight modifications. However, the multi-resolution branches in this paradigm result in a bottleneck in terms of throughput. To address this issue, this study proposes StarPose, a single-branch, upsampling-free macro architecture based on HRPVT. By optimizing all micro block designs from a lightweight perspective and reconstructing the network using the advanced star operation design insight, the proposed method can handle high-dimensional features while computing in a low-dimensional space, akin to the mechanism of kernel functions, thereby achieving more effective semantic feature representation. The proposed method achieves 2 × faster inference speed than Lite-HRNet under nearly the same model complexity on the MS COCO benchmark, while maintaining superior accuracy. This significant improvement unlocks greater potential for deployment on resource-constrained edge devices. The code and models can be accessed publicly at https://github.com/george-xu-code/StarPose .