Image matching is essential in computer vision for applications like pose estimation, object retrieval, 3D Reconstruction, and SLAM, which typically involve Front-End (detection/description), Middle-End (matching/filtering), and Back-End (integration) stages, all constrained by computational budgets. For efficient on-device inference, sparse Front-End methods represent images as keypoints and lines forming wireframes; however, these elements are typically detected and processed independently. We propose a unified neural network that directly produces wireframes with discriminative descriptors for each node (keypoint/junction). This unified approach enhances the efficiency of the Front-End by simultaneously finding wireframe elements and their node descriptors in a single pass, thereby reducing computational cost and simplifying the pipeline compared to traditional multi-stage methods. Our model, trained via multitask distillation from state-of-the-art point and line feature extractors, achieves a competitive trade-off between accuracy and efficiency.