TranMamba: a lightweight hybrid transformer-Mamba network for single image super-resolution

Transformers excel in modeling long-range dependencies for computer vision, but the quadratic complexity of self-attention complicates lightweight model design. The Mamba model, with linear complexity, offers similar capabilities but underperforms compared to Transformers. Inspired by these insights, we propose TranMamba, a lightweight hybrid Transformer-Mamba network that enhances both performance and efficiency in Single Image Super-Resolution (SISR). Specifically, we reduce the computational cost associated with self-attention by alternating between Transformer and Mamba modules. To balance the extraction of both local and global information, we designed Transformer Aggregation Block (TAB) and Mamba Aggregation Block (MAB) to strengthen feature representation. Additionally, we developed a Reparameterized Spatial-Gate Feed-Forward Network (RepSGFN) to further improve the model's feature extraction capabilities. Extensive experiments demonstrate that TranMamba achieves SOTA performance among models of comparable size.