<p>Single image super-resolution (SISR) has achieved remarkable progress in reconstructing image textures and structural details. However, existing CNN–Transformer architectures for super-resolution often suffer from insufficient local texture and global context modeling due to deep stacking or fixed receptive fields, limiting their ability to restore complex textures across multiple scales. To address these limitations, we propose LISC, a hybrid layered attention network designed for high-quality texture reconstruction. Unlike conventional methods that passively integrate multi-scale features through stacked or branched designs, LISC leverages the Efficient Hybrid Attention Block (EHAB) to structurally decouple initial features into two branches : the Channel Convolution Attention Block (CCAB) for local texture modeling and the Self-Correlation Attention Block (SCAB) for global structural modeling. The CCAB and SCAB cooperate with each other to capture global semantics and local details. Furthermore, rather than relying on fixed-window self-attention that only captures single-scale dependencies, SCAB introduces a five-level layered attention mechanism with cross-window interaction, progressively expanding the receptive field to enhance long-range feature aggregation and global semantic representation. Experiments on five benchmark datasets demonstrate that the proposed network achieves high-quality image reconstruction and efficient performance. Specifically, under the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation>4 scaling setting, our method attains an average PSNR of 32.52dB, SSIM of 0.8992, and requires only 25.4G FLOPs, while maintaining consistent performance across Set5, Set14, BSD100, Urban100, and Manga109. These results confirm that LISC effectively balances high-fidelity texture reconstruction with lightweight computational efficiency.</p>