SF-SRGAN: Shift-Free Transformer GAN with High-Frequency Enhancement for Image Super-Resolution

Image super-resolution refers to the process of recovering high-resolution images from low-resolution inputs. However, existing methods face three critical challenges: GAN-based approaches suffer from training instability and artifacts; shifted-window Transformers have issues of computational cost and fixed receptive fields; and high-frequency detail modeling along with multi-source feature fusion remains insufficient. To address these issues, this paper proposes SF-SRGAN, a high-frequency enhanced shift-free Transformer generative adversarial network. The generator introduces a shift-free variable-window Transformer architecture that progressively increases window sizes across layers (4 × 4 → 8 × 8 → 16 × 16), achieving natural receptive field expansion without shift operations or masking overhead. A High-Frequency Positional Encoding module is designed to extract multi-directional high-frequency responses via Laplacian kernels and adaptively inject them through a content-aware gating mechanism. An adaptive dual-path fusion mechanism dynamically integrates global Transformer features with local CNN features. The discriminator adopts a multi-scale architecture combined with R1 gradient penalty and an auxiliary edge prediction loss to stabilize adversarial training and preserve high-frequency structures. Extensive experiments on SR-CelebA and Flickr2K datasets demonstrate that SF-SRGAN achieves superior performance in PSNR, SSIM, and LPIPS, and ablation studies validate the effectiveness of each component.