CASR: A Robust Cyclic Framework for Arbitrary Large-Scale Super-Resolution with Distribution Alignment and Self-Similarity Awareness
Researchers have developed CASR, a novel cyclic framework designed to overcome limitations in arbitrary large-scale super-resolution (ASISR). This framework addresses the issue of cross-scale distribution shifts that typically lead to artifacts and noise at high magnification levels. By reformulating ultra-magnification as a sequence of in-distribution scale transitions, CASR ensures stable inference with a single model, even at extreme scales. The system incorporates SSAM and SARM modules to align structural distributions and restore high-frequency textures, thereby preserving self-similarity and improving generalization. AI