A Multi-stage Transformer-Diffusion Framework for Blind Image Super-Resolution

Blind super-resolution reconstruction technology can recover high-frequency information from low-resolution images and generate high-fidelity visual content for complex scenarios such as satellite image enhancement, digital restoration of vintage films, and fragment recovery of ancient artifacts. To enhance the algorithm’s capability in capturing global information for structural reconstruction and improving the authenticity of detailed textures, this paper proposes a multi-stage blind super-resolution reconstruction framework based on a diffusion model. A Transformer reconstruction network incorporating dilated window attention and fast Fourier convolution (FFC) is first constructed to strengthen long-range dependency modeling and effectively restore the global image structure. Subsequently, an implicit diffusion model with a conditional control module is introduced to generate high-fidelity texture details through a U-Net encoder and a fidelity balancing mechanism. Experimental results demonstrate that compared to existing solutions, the proposed method achieves superior performance on multiple mainstream image reconstruction datasets in terms of both objective evaluation metrics (e.g., PSNR, SSIM) and subjective visual quality, particularly excelling in complex texture restoration and edge structural coherence.