CASR: A Robust Cyclic Framework for Arbitrary Large-Scale Super-Resolution with Distribution Alignment and Self-Similarity Awareness
This addresses a fundamental limitation in image super-resolution for applications requiring high-quality scaling beyond trained scales, though it appears incremental as it builds on existing cyclic and distribution alignment ideas.
The paper tackles the problem of arbitrary-scale super-resolution (ASISR) suffering from noise, blur, and artifacts when scaling beyond training ranges by proposing CASR, a cyclic framework that reformulates ultra-magnification as in-distribution scale transitions, achieving stable inference with a single model and reducing distribution drift.
Arbitrary-Scale SR (ASISR) remains fundamentally limited by cross-scale distribution shift: once the inference scale leaves the training range, noise, blur, and artifacts accumulate sharply. We revisit this challenge from a cross-scale distribution transition perspective and propose CASR, a simple yet highly efficient cyclic SR framework that reformulates ultra-magnification as a sequence of in-distribution scale transitions. This design ensures stable inference at arbitrary scales while requiring only a single model. CASR tackles two major bottlenecks: distribution drift across iterations and patch-wise diffusion inconsistencies. The proposed SDAM module aligns structural distributions via superpixel aggregation, preventing error accumulation, while SARM module restores high-frequency textures by enforcing autocorrelation and embedding LR self-similarity priors. Despite using only a single model, our approach significantly reduces distribution drift, preserves long-range texture consistency, and achieves superior generalization even at extreme magnification.