Arterial spin labeling (ASL) magnetic resonance imaging (MRI) is a powerful imaging technology that can measure cerebral blood flow (CBF) quantitatively. However, since only a small portion of blood is labeled compared to the whole tissue volume, conventional ASL suffers from low signal-to-noise ratio (SNR), poor spatial resolution, and long acquisition time. In this paper, we proposed a super-resolution method based on a multi-scale generative adversarial network (GAN) through unsupervised training. The network only needs the low-resolution (LR) ASL image itself for training and the T1-weighted image as the anatomical prior. No training pairs or pre-training are needed. A low-pass filter guided item was added as an additional loss to suppress the noise interference from the LR ASL image. After the network was trained, the super-resolution (SR) image was generated by supplying the upsampled LR ASL image and corresponding T1-weighted image to the generator of the last layer. Performance of the proposed method was evaluated by comparing the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) using normal-resolution (NR) ASL image (5.5 min acquisition) and high-resolution (HR) ASL image (44 min acquisition) as the ground truth. Compared to the nearest, linear, and spline interpolation methods, the proposed method recovers more detailed structure information, reduces the image noise visually, and achieves the highest PSNR and SSIM when using HR ASL image as the ground-truth.