<p>Recent advancements have demonstrated the effectiveness of retentive networks in natural language processing and high-level vision tasks. However, their potential in low-level vision, such as image super-resolution (SR), remains underexplored. In this paper, we introduce RetNetSR, a novel retentive network architecture designed specifically for image super-resolution. Our approach leverages a spatial prior derived from the Manhattan distance to enhance the Self-Attention mechanism, effectively translating the temporal decay concept of RetNet into the spatial domain. At the core of RetNetSR is the Manhattan Self-Attention module, which integrates Self-Attention with multi-layer perceptrons and depthwise convolution. Building on this, we propose the Manhattan Self-Attention Block and the Manhattan Self-Attention Group, the latter further enriched with 3 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation> 3 convolutions and Enhanced Spatial Attention modules for more effective deep feature extraction. Experimental evaluations across multiple runs on standard image SR benchmarks show that RetNetSR consistently achieves competitive or superior results, with statistical analysis (paired t-test, <i>p</i> <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(&lt;\)</EquationSource> </InlineEquation> 0.05) confirming that improvements—such as raising PSNR from 32.76 to 32.86 on Urban100 (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>2 scale) and SSIM from 0.9151 to 0.9156 on Manga109 (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>4 scale) are statistically significant. These findings highlight the novelty of RetNetSR in unifying temporal-inspired retentive modeling with spatial feature enhancement, establishing a new and efficient architectural direction for low-level vision that has not been explored in prior SR literature</p>