Omni-frequency diffusion-based regional feature consistency recovery for realistic image super-resolution

Blind super-resolution (BSR) endeavors to recover a high-quality image from its degraded low-resolution counterpart, which suffers from multiple complex and unknown degradations. Existing diffusion-based methods strive to indiscriminately enhance all intricate aspects, disregarding the consistent interconnectedness between specific details and their surrounding contextual characteristics. Consequently, these approaches typically induces the production of excessively intensified and artificial details. In this paper, we propose a novel Regional feature Consistency Recovery framework based on diffusion model, DiffRCR, which exhibits the capability to hierarchically restore the frequency domain information vanishing in different regions, and generate photorealistic details consistent with the regional characteristics. Specifically, we devise a hierarchical recovery method to explore frequency differences and recover different frequency domains at various levels, enabling the hierarchical recovery of distinct degradation regions. Furthermore, DiffRCR harnesses contextual information surrounding intricate details and regenerates these details ensuring a harmonious consistency with the characteristics of the degraded area. Empirical evidence from comprehensive experiments has substantiated that DiffRCR yields excellent performance in relation to both the fidelity of construction and the perceptual quality when employed for authentic super-resolution tasks. The code is available at https://github.com/huanglab-research/DiffRCR.