<p>Deep learning has revolutionized Single Image Super-Resolution (SISR), yet existing methods often fail to recover high-frequency details, resulting in blurry textures and unrealistic reconstructions. Moreover, GAN-based approaches, while effective in perceptual quality, commonly suffer from unstable training. To address these challenges, we propose High-Frequency Learning for Stable and Realistic Single Image Super-Resolution (HFLSR). The generator employs a Frequency Regulated Attention Block (FRAB) that explicitly emphasizes high-frequency information through complementary High-Frequency Spatial Attention (HFSA) and Simplified Channel Attention (SCA) modules, ensuring adaptive refinement of critical structural and textural features. To further enhance realism, we incorporate a high-frequency aware loss function, encouraging faithful reconstruction of edges and fine textures in alignment with the ground truth. On the discriminator side, we introduce a Relativistic Average Least Squares GAN (RaLSGAN) framework, which provides stable training dynamics while enforcing relative realism between real and generated images. Extensive experiments on benchmark datasets demonstrate that HFLSR consistently outperforms state-of-the-art methods in both objective measures (PSNR, SSIM, LPIPS, PI) and subjective perceptual quality. Specifically, compared to the baseline SRGAN model, HFLSR achieves a PSNR gain of 1.25dB (6.2%) and PI reduction of 0.44 (11.8%) on the Urban100 dataset.</p>