CASR: A Robust Cyclic Framework for Arbitrary Large-Scale Super-Resolution with Distribution Alignment and Self-Similarity Awareness

2026-02-25

AI Summary

This work addresses the challenge of arbitrary-scale super-resolution, where performance degrades significantly when the inference scale lies outside the training distribution, leading to noise amplification, blurring, and artifact accumulation. To mitigate this cross-scale distribution shift, we propose CASR, a recurrent framework that decomposes large upscaling factors into a sequence of in-distribution scale transitions, enabling stable arbitrary-scale super-resolution with a single model. The core innovations include a Structure Distribution Alignment Module (SDAM) based on superpixel aggregation to suppress error propagation, and a Texture Restoration Module (SARM) that integrates self-attention with low-resolution self-similarity priors to enhance high-frequency details. Extensive experiments demonstrate that our approach effectively alleviates distribution drift, preserves long-range texture consistency, and exhibits strong generalization even under extreme upscaling factors.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Unsupervised & Self-Supervised Learning
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

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.

Problem

• Arbitrary-Scale Super-Resolution
• cross-scale distribution shift
• distribution drift
• texture consistency
• super-resolution

Innovation

• Arbitrary-Scale Super-Resolution
• Distribution Alignment
• Self-Similarity Awareness
• Cyclic Framework
• Structural Distribution Matching