Image super-resolution aims to enhance the resolution of low-quality images by generating high-resolution counterparts, significantly benefiting various practical applications, such as high-quality television, gaming, and medical imaging. Our work stems from the success of SwinIR, a state-of-the-art model that leverages Swin Transformers and window-based self-attention mechanisms to effectively model distant dependencies. Firstly, we employ the CutBlur augmentation method to increase both the size and diversity of the training data. This technique cuts and pastes random regions between low- and high-resolution images, forcing the model to tackle both degraded and detailed areas simultaneously. Secondly, we replace the Swin Transformer layer in the original model with a Locally-Enhanced Feed-Forward Network (LEFF) layer. This modification improves the model’s ability to capture local context by incorporating a depth-wise convolutional block within the feed-forward network. Experimental results demonstrate that our proposed approach consistently outperforms several baselines across various benchmarks. Notably, on the Set5 and Set14 datasets, our model achieves PSNR values of 38.37 dB and 34.17 dB, respectively, surpassing the baseline SwinIR. On the BSD100 and Manga109 datasets, our model achieves a PSNR of 39.61 dB, maintaining superior performance. Although the performance on the more challenging Urban100 dataset is slightly lower, it remains comparable to the baseline, leaving room for further improvement.