Computed tomography angiography (CTA) scans provide doctors with an accurate visualization of the vascular system that helps them make diagnostic decisions. To help doctors make a quick diagnosis, the resolution of CTA images needs to be improved. With the advancement of deep learning technology, many neural network models have been proposed to improve image quality. Most current image super-resolution (SR) methods still have problems, such as texture loss and boundary-blurring, and there are few methods for CTA images. In this paper, we propose the Separation Feature Residual Diffusion Model (Dual-ResShift) for generating high-resolution (HR) CTA images with only 15 sampling steps. In the model, we propose a feature-separated SFUNet and a new feature extraction algorithm, GBVS-Enhanced based on Graph-based visual saliency (GBVS), to enhance the extraction of high-frequency features. In addition, we design a loss function named \(\mathcal L_{Edge}\) to guide the diffusion model to optimize the image. Our method was verified on the clinical CTA dataset and AVT dataset. The PSNR reached 26.4165, SSIM reached 0.7440, and LPIPS reached 0.0484 on the clinical CTA dataset. The PSNR reached 28.6791, SSIM reached 0.7805, and LPIPS reached 0.0365 on the AVT dataset. Experiments show that Dual-ResShift can outperform existing methods. Our code and model are put on https://github.com/prefectmoon/Dual-ResShift-code .