Accurate separation of retinal arteries and veins is essential for computing vessel-specific morphological biomarkers. Most deep-learning approaches are trained on older public fundus image datasets, so applying them to modern high-resolution images requires downsampling to match vessel scale, distorting vessel geometry and obscuring fine details. This study aims to develop a convolutional neural network capable of artery–vein segmentation directly from fundus images of variable native resolution without global downsampling.