Abstract: Information distillation has emerged as an effective strategy for developing lightweight image super-resolution (SR) reconstruction networks because it enables efficient feature reuse and reduces computational redundancy. However, most existing information distillation–based SR methods enhance efficiency primarily by simplifying feature extraction and distillation processes, which often leads to insufficient exploitation of discriminative information and the inadvertent introduction of noise during feature distillation. These limitations significantly constrain the reconstruction performance of lightweight models, particularly when dealing with complex textures and fine structural details. To address these issues, a novel lightweight image super-resolution framework, termed information distillation and shared attention network (IDSA-Net), is proposed. The proposed network aims to achieve a more favorable balance between reconstruction accuracy and computational efficiency by enhancing effective feature representation while suppressing redundant and noisy information. Built upon the principle of information distillation, IDSA-Net incorporates several carefully designed attention-based modules to refine feature processing throughout the network. First, to reduce the model complexity and alleviate redundant computations commonly observed in multibranch attention architectures, a shared attention mechanism is introduced, and an attention sharing distillation block (ASDB) module is designed. Unlike conventional attention modules that compute attention weights independently for different feature paths, the ASDB enables multiple distilled feature streams to share the same attention matrices. This design effectively avoids repetitive attention computations, significantly reduces computational overhead, and improves distillation efficiency while maintaining high reconstruction quality. Second, to further enhance the feature representation and improve the utilization of informative features, a hierarchical distillation and feature purification strategy is proposed. Specifically, a large kernel channel information purification block (LCIPB) is constructed to selectively refine channel-wise features during the distillation process. By employing large-kernel convolutions, LCIPB expands the receptive field and captures long-range contextual dependencies, which are crucial for recovering global structures and high-frequency details in image super-resolution. Meanwhile, channel-level feature screening and enhancement are performed to effectively suppress redundant or noisy responses introduced during distillation. This hierarchical purification mechanism ensures that only the most relevant and discriminative features are propagated to subsequent stages without increasing model complexity. Third, to fully exploit complementary information across both channel and spatial dimensions and further strengthen feature representation, an instance normalized hybrid attention module (INAM) is designed. This module integrates channel and spatial attention within a unified framework. In addition, instance normalization is incorporated to stabilize feature distributions and mitigate feature-scale variations across different images, which is particularly beneficial for lightweight networks with limited capacity. The integration of hybrid attention and instance normalization enhances the network’s ability to model complex textures and fine structural details while effectively suppressing noise amplification during the distillation process. Extensive experiments were conducted on four widely used benchmark datasets, namely Set5, Set14, BSD100, and Urban100. Quantitative and qualitative comparisons, together with detailed ablation studies, were performed using 13 representative state-of-the-art image super-resolution methods. Experimental results demonstrate that IDSA-Net consistently outperforms competing approaches in terms of peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) across different datasets and scaling factors. In particular, under the ×4 upscaling setting, the proposed model achieves PSNR values of 32.47 dB on Set5, 28.84 dB on Set14, 27.72 dB on BSD100, and 26.57 dB on Urban100 while maintaining an extremely compact model size of only 752k parameters. In conclusion, IDSA-Net effectively addresses the limitations of existing lightweight information distillation–based super-resolution methods by integrating shared attention, hierarchical feature purification, and hybrid attention mechanisms. By achieving an excellent trade-off between model complexity and reconstruction accuracy, the proposed framework provides an efficient and practical solution for lightweight image super-resolution, particularly in resource-constrained and real-time application scenarios.