<p>Handling unknown degradations remains a fundamental challenge in real-world Single Image Super-Resolution (SISR). Most existing methods rely on supervised training with synthetic Low-Resolution (LR) and High-Resolution (HR) pairs, leading to poor generalization under real-world conditions. To address this, we propose the Generalized Real-world Super-Resolution Generative Adversarial Network (GRSGAN), a novel two-stage framework that robustly recovers high-fidelity images from real degraded inputs without requiring paired data. In the first stage, a Zero-Shot Adaptive Degradation Correction Network (ADCN) estimates and reverses complex, unknown degradations such as noise, blur, and compression artifacts through an internal learning process. In the second stage, an Efficient Super-Resolution Generative Network (ESRG-Net) leverages a lightweight GAN architecture with Depthwise Separable Residual Attention Blocks (DSRAB) to reconstruct perceptually rich and structurally accurate HR images. Training of the ESRG-Net is guided by an edge-aware hybrid loss to balance fidelity and visual realism. While the reconstruction stage is designed for efficiency, the integrated zero-shot optimization ensures high adaptability to diverse real-world artifacts. Evaluation on multiple real-world benchmarks demonstrates that GRSGAN significantly outperforms state-of-the-art methods in both distortion-oriented (PSNR/SSIM) and perception-oriented (LPIPS/NIQE/DISTS) metrics. GRSGAN achieved +1.2dB PSNR gain and 18% lower LPIPS compared to the baseline Real-ESRGAN model on the RealSR dataset. The proposed framework demonstrates strong generalization to unseen degradations and establishes a robust baseline for high-fidelity restoration. Finally, this work also provides a clear foundation for future research toward more efficient, broadly adaptable real-world super-resolution systems.</p>