This paper presents a highly accurate single-image super-resolution (SR) method using a very deep convolutional network. The method improves accuracy by increasing network depth, using 20 weight layers, and cascading small filters to efficiently exploit contextual information over large image regions. The proposed method uses residual learning and extremely high learning rates (10^4 times higher than SRCNN) with adjustable gradient clipping to address convergence issues. It outperforms existing methods in both accuracy and visual quality. The method is also extended to handle multi-scale SR problems within a single network, making it more efficient and practical. The network is trained using a deep architecture that models image details (residuals) rather than directly modeling high-resolution images, leading to faster convergence and better performance. The method is tested on several datasets and shows significant improvements over state-of-the-art methods, particularly in terms of PSNR and SSIM metrics. The proposed method is also efficient, with training time reduced to 4 hours on a GPU, compared to several days for SRCNN. The method is capable of handling multiple scale factors within a single network, reducing the need for multiple models and improving efficiency. The results show that the proposed method outperforms existing methods in terms of both quantitative and qualitative performance.This paper presents a highly accurate single-image super-resolution (SR) method using a very deep convolutional network. The method improves accuracy by increasing network depth, using 20 weight layers, and cascading small filters to efficiently exploit contextual information over large image regions. The proposed method uses residual learning and extremely high learning rates (10^4 times higher than SRCNN) with adjustable gradient clipping to address convergence issues. It outperforms existing methods in both accuracy and visual quality. The method is also extended to handle multi-scale SR problems within a single network, making it more efficient and practical. The network is trained using a deep architecture that models image details (residuals) rather than directly modeling high-resolution images, leading to faster convergence and better performance. The method is tested on several datasets and shows significant improvements over state-of-the-art methods, particularly in terms of PSNR and SSIM metrics. The proposed method is also efficient, with training time reduced to 4 hours on a GPU, compared to several days for SRCNN. The method is capable of handling multiple scale factors within a single network, reducing the need for multiple models and improving efficiency. The results show that the proposed method outperforms existing methods in terms of both quantitative and qualitative performance.