Single Image Super-Resolution (SISR) has emerged as one of the key beneficiaries of the persistent evolution in deep learning architectures. Following the breakthrough of transformers in natural language processing, the computer vision community recognized their potential for visual applications. However, a significant limitation of transformer architectures in vision tasks is their tendency to prioritize local image regions while struggling to capture comprehensive global context. To overcome these limitations, the proposed architecture, referred as PRISM incorporates channel attention modules for comprehensive global feature extraction, complemented by windowed self-attention mechanisms that maintain localized spatial details. To enhance the SR quality, we introduce an overlapping cross-attention mechanism that establishes dynamic connections between adjacent window partitions, enabling more comprehensive feature extraction. Further, through consistent testing, training, and evaluation of the proposed model, we experimented with many additional features and methods on major datasets. The PRISM model achieved 9th place in both the Restoration Track and the Perceptual Track of the NTIRE 2025 Image Super-Resolution ( \(\times 4\) ) Challenge, demonstrating its effectiveness for high-quality image SR [4].