Deep learning approaches dominate single image super-resolution (SISR) due to their effectiveness. However, these methods often lack fine details, leading to visually unsatisfying results. For SISR, Generative Adversarial Networks (GANs) have shown excellent performance, but training them can be challenging, and improper training can significantly degrade performance. This article introduces an effective method, the Least Squares Multi-Scale Generative Adversarial Network for Super Resolution (MSSRGAN-LS), for high-quality perceptual SISR. It leverages a Least Squares GAN (LSGAN) architecture to generate visually realistic images with rich textural details. The network incorporates a dilated convolution-based feature extraction block to capture multi-scale information, enhancing the SISR process. Spectral normalization is employed to improve the overall GAN performance for SISR tasks. Additionally, a least squares discriminator is utilized to stabilize training and generate higher-quality images. Rigorous experiments on benchmark datasets confirm that MSSRGAN-LS outperforms other approaches in terms of performance, as indicated by qualitative and quantitative metrics.