This paper presents a modified super-resolution framework based on WDSR-GAN, enhanced through orthogonal regularization to improve visual quality and training stability. Single Image Super-Resolution (SISR) remains a challenging task, especially in generating realistic textures while maintaining structural fidelity. While GAN-based methods have made significant progress in this field, they often encounter issues such as overfitting, filter redundancy, and inconsistent convergence. Our approach builds on the WDSR architecture and introduces orthogonal constraints within the generator to encourage feature diversity and reduce redundancy across convolutional filters. This modification not only helps prevent mode collapse but also contributes to more stable adversarial training. The model was trained using the DIV2K dataset and evaluated on widely-used benchmarks including Set5, Set14, Urban100, and BSD100. Experimental results demonstrate that our method achieves better perceptual outcomes, reflected in higher SSIM values and more natural image textures, while maintaining competitive PSNR performance. These findings support the effectiveness of orthogonal regularization in advancing GAN-based Single Image Super-Resolution.