The paper introduces the Laplacian Pyramid Super-Resolution Network (LapSRN), a deep convolutional network designed for fast and accurate single-image super-resolution (SR). LapSRN progressively reconstructs high-frequency residuals of high-resolution (HR) images using a cascade of convolutional layers and transposed convolutions. Unlike existing methods that use bicubic interpolation or pre-processing steps, LapSRN directly predicts sub-band residuals from coarse-resolution feature maps, reducing computational complexity. The network is trained with deep supervision using a robust Charbonnier loss function, which improves accuracy and handles outliers effectively. LapSRN generates multi-scale predictions in a single feed-forward pass, making it suitable for resource-aware applications. Extensive evaluations on benchmark datasets show that LapSRN outperforms state-of-the-art methods in terms of both speed and reconstruction quality. The paper also discusses the limitations of the method, such as its inability to hallucinate fine details in low-resolution inputs.The paper introduces the Laplacian Pyramid Super-Resolution Network (LapSRN), a deep convolutional network designed for fast and accurate single-image super-resolution (SR). LapSRN progressively reconstructs high-frequency residuals of high-resolution (HR) images using a cascade of convolutional layers and transposed convolutions. Unlike existing methods that use bicubic interpolation or pre-processing steps, LapSRN directly predicts sub-band residuals from coarse-resolution feature maps, reducing computational complexity. The network is trained with deep supervision using a robust Charbonnier loss function, which improves accuracy and handles outliers effectively. LapSRN generates multi-scale predictions in a single feed-forward pass, making it suitable for resource-aware applications. Extensive evaluations on benchmark datasets show that LapSRN outperforms state-of-the-art methods in terms of both speed and reconstruction quality. The paper also discusses the limitations of the method, such as its inability to hallucinate fine details in low-resolution inputs.