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

Source: arXiv - 2602.22159v1

Overview

The paper introduces CASR, a cyclic super‑resolution (SR) framework that can upscale images to any factor with a single model. By treating extreme up‑scaling as a series of small, in‑distribution steps, CASR dramatically reduces the noise, blur, and artifacts that normally explode when the target scale falls outside the training range.

Key Contributions

• Cyclic SR formulation – Re‑expresses arbitrary‑scale up‑sampling as a chain of modest, in‑distribution magnifications, enabling stable inference with one network.
• Structural Distribution Alignment Module (SDAM) – Uses super‑pixel aggregation to align feature distributions across iterations, preventing drift and error accumulation.
• Self‑Similarity Aware Restoration Module (SARM) – Enforces autocorrelation constraints and injects low‑resolution (LR) self‑similarity priors to recover high‑frequency textures.
• Single‑model, multi‑scale solution – No need for separate models or scale‑specific finetuning; the same weights handle 2×, 4×, 16×, or even 64× up‑scaling.
• State‑of‑the‑art performance on standard benchmarks, especially at extreme magnifications where prior methods collapse.

Methodology

1. Cyclic Upscaling Loop

   • Instead of a one‑shot jump from LR to the desired HR size, the image is repeatedly passed through the SR network, each time enlarging it by a modest factor (e.g., 1.5×–2×).
   • This keeps every intermediate output within the distribution the model was trained on, avoiding the “out‑of‑distribution” shock that causes artifacts.

2. Structural Distribution Alignment Module (SDAM)

   • The feature map from the current iteration is segmented into super‑pixels (coherent regions).
   • Statistics (mean, variance) of each super‑pixel are aligned to those of the previous iteration, effectively “re‑centering” the distribution and stopping drift.

3. Self‑Similarity Aware Restoration Module (SARM)

   • Computes an autocorrelation map of the LR input to capture repeating patterns (textures, edges).
   • During each up‑sampling step, SARM injects these self‑similarity cues back into the feature space, encouraging the network to reproduce realistic high‑frequency details rather than hallucinating noise.

4. Training

   • The network is trained on a conventional range of scales (e.g., 1×–4×).
   • Losses combine pixel‑wise L1/L2, perceptual (VGG) loss, and a novel distribution‑alignment loss that penalizes divergence between successive super‑pixel statistics.

Results & Findings

• Extreme scales (×16, ×32, ×64): CASR retains visual fidelity, while competing methods exhibit severe blur and ringing.
• Distribution drift measured by KL‑divergence between successive iterations drops by ~45 % thanks to SDAM.
• Texture consistency (measured via autocorrelation similarity) improves by ~20 % with SARM, confirming that self‑similarity priors are effectively leveraged.

Qualitative examples show clean edge reconstruction and plausible fine‑grained patterns (e.g., fabric weave, foliage) even at 64× magnification.

Practical Implications

• Single‑model deployment – Developers can ship one lightweight SR service that handles any client‑requested zoom level, simplifying CI/CD pipelines and reducing memory footprints.
• Real‑time streaming & VR – The cyclic approach can be throttled adaptively: fewer iterations for low‑latency scenarios, more for high‑quality offline rendering.
• Legacy image restoration – Archivists can upscale historical photos to very high resolutions without training a bespoke model for each target scale.
• Edge devices – Because each iteration works on a modest up‑scale factor, the per‑step compute stays bounded, making it feasible to run on mobile GPUs or NPUs with progressive refinement.

Limitations & Future Work

• Inference latency grows linearly with the number of cycles; extremely high magnifications still require many passes, which may be prohibitive for ultra‑low‑latency use‑cases.
• The current SDAM relies on super‑pixel segmentation, which adds a preprocessing overhead and may struggle with highly textured or noisy inputs.
• Authors note that extending the framework to video SR (temporal consistency) and exploring learned adaptive cycle lengths are promising directions for follow‑up research.

Authors

• Wenhao Guo
• Zhaoran Zhao
• Peng Lu
• Sheng Li
• Qian Qiao
• RuiDe Li

Paper Information

• arXiv ID: 2602.22159v1
• Categories: cs.CV
• Published: February 25, 2026
• PDF: Download PDF