Unsupervised image super-resolution recurrent network based on diffusion model

Unsupervised image super-resolution offers distinct advantages for real-world applications by eliminating the need for paired high- and low-resolution images. This paper proposes a novel architecture specifically designed for unsupervised learning, consisting of a cycle branch and a diffusion branch. The cycle branch integrates an upsampling and a downsampling network to generate pseudo-paired images from unpaired high- and low-resolution inputs. In parallel, the diffusion branch incorporates two independent diffusion models that refine these pseudo pairs, jointly modeling the processes of image reconstruction and degradation. This collaborative design enhances the authenticity of the pseudo pairs and enriches the detail in the reconstructed images. A key challenge in unsupervised learning is the lack of explicit label supervision, which often leads to inaccurate color restoration. To address this, we introduce a color consistency loss that regulates the cycle branch and promotes color fidelity. Through joint end-to-end training, the two branches complement each other to achieve high-quality reconstruction. Experimental results demonstrate that the proposed method effectively handles real-world low-resolution images, providing a robust and practical solution for image super-resolution.