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

Wenhao Guo; Zhaoran Zhao; Peng Lu; Sheng Li; Qian Qiao; DeRui Li

cs.CVFeb 25, 2026

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

Authors: Wenhao Guo, Zhaoran Zhao, Peng Lu, Sheng Li, Qian Qiao, DeRui Li

Organizations: Beijing University of Posts and Telecommunications · Peking University

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 SSAM module aligns structural distributions via superpixel aggregation, preventing error accumulation, while SARM module restores high-frequency textures by enforcing correlation-guided consistency and preserving self-similarity structure through correlation alignment. 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.

Figures & tables

Figure 1: Comparison of cyclic cascade stability across different ASISR. The SIFID measures distribution shifts between reconstructed images and the training data during cascading. Our method achieves notably higher distribution stability than others.

Figure 2: This illustrates the texture inconsistency between patches caused by patch-based super-resolution, where identical repeated objects are reconstructed with different texture patterns.

Figure 3: Illustration of the proposed CASR . The purple module denotes the SSAM, the green block corresponds to the SARM, and the gray U-Net represents the SR backbone.

Panel of Figure 4 — Figure 4: Illustration of the distribution alignment process, where the input image is decomposed into a superpixel representation and a depth map. This decomposition effectively removes artifacts and noise, enabling robust SR.

Figure 5: Illustration of the local self-similarity computation, where structurally similar regions are assigned higher correlation.

Method	DIV8K
	$\times$ 8				$\times$ 12				$\times$ 18
	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
LINF [ 32 ]	0.442	26.01	10.11	8.93	0.528	19.42	11.43	9.87	0.578	17.24	13.37	10.89
BFSR [ 27 ]	0.399	24.30	8.29	7.80	0.500	18.43	10.70	9.18	0.561	17.06	14.24	11.06
IDM [ 9 ]	0.486	24.11	7.23	6.46	0.604	23.42	7.98	6.87	0.656	23.75	7.82	6.87
Kim [ 17 ]	0.491	23.54	8.24	8.84	0.621	21.69	8.80	7.38	0.685	20.06	8.34	7.72

Table 1: Comparison with ASISR methods on the DIV8K synthetic dataset, with the best results in bold . The

\times 4\times 3\times 1.5

column evaluates all methods under progressive upsampling.

Panel of Figure 6 — Figure 6: Qualitative comparison with different methods on the DIV8K dataset. For large-scale super-resolution, our method reconstructs more realistic statue textures and finer fur details on the cat’s ears.

Method	RealSR
	$\times$ 8			$\times$ 12			$\times$ 18			$\times$ 24			$\times$ 30
	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
LINF [ 32 ]	19.58	11.50	9.96	17.22	12.90	10.83	16.74	14.59	11.71	17.26	15.70	12.29	18.29	16.26	12.56
BFSR [ 27 ]	18.27	9.91	8.96	16.56	12.21	10.19	16.28	15.51	11.79	16.88	16.71	12.31	18.04	16.85	12.27
IDM [ 9 ]	28.69	7.50	7.15	26.56	8.14	7.67	33.79	8.00	7.52	31.68	8.17	7.47	28.22	8.35	7.44
Kim [ 17 ]	26.39	8.09	6.94	25.16	7.83	7.85	25.32	8.51	8.42	30.59	8.58	8.46	27.43	8.56	8.43

Table 2: Comparison with ASISR methods on real-world datasets, with the best results in bold . Our approach archives consistently superior performance over others, showcasing strong generalization in real-world image.

Panel of Figure 7 — Figure 7: Qualitative comparison with different methods on the RealSR dataset. Our method produces clearer and more natural results.

Method	CelebA-HQ
	$\times$ 4			$\times$ 6
	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
IDM [ 9 ]	72.79	9.70	6.97	66.77	5.60	5.44
Kim [ 17 ]	43.57	7.83	7.48	41.06	8.06	7.87
CASR	71.12	9.69	7.19	70.86	4.77	4.87
Method	$\times$ 8			$\times$ 12

Table 3: Comparison with diffusion-based ASISR methods on the CelebA-HQ dataset. Our method achieves superior performance at large upsampling scales.

Panel of Figure 8 — Figure 8: Super-resolution results at $\times 12$ on CelebA-HQ. IDM and Kim fail to recover fine facial details, while our method produces cleaner and sharper reconstructions.

SuperPixel	Depth	SARM	$\times$ 18 ( $\times 4\times 3\times 1.5$ )
SuperPixel	Depth	SARM	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
✗	✗	✗	0.585	31.73	7.10	5.91
✓	✗	✗	0.471	42.23	6.61	5.96
✓	✓	✗	0.467	45.18	6.15	5.37
✓	✓	✓	0.450	51.44	6.01	5.24

Table 4: Ablation study of major components in CASR. Each module contributes to the overall performance improvement.

Panel of Figure 9 — Figure 9: Impact of different components. Incorporating the superpixel module (Model1) effectively suppresses accumulated blur and artifacts during cascading, while depth conditioning (Model2) further enhances edge sharpness. The full model produces natural results with consistent textures across patches.

Configuration	$\times$ 18 ( $\times 4\times 3\times 1.5$ )
Configuration	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
w/o global semantic	0.469	49.10	6.30	5.49
w/ global semantic	0.450	51.44	6.01	5.24

Table 5: Ablation on global semantic context within SARM.

$L_{corr}$	$L_{depth}$	$\times$ 18 ( $\times 4\times 3\times 1.5$ )
$L_{corr}$	$L_{depth}$	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
✗	✗	0.462	49.33	6.71	5.91
✓	✗	0.459	50.23	6.24	5.40
✓	✓	0.450	51.44	6.01	5.24

Table 6: Ablation study on loss functions, validating the effectiveness of each loss component.

SuperPixel Size	DIV8K
SuperPixel Size	LPIPS $\downarrow$	MUSIQ $\uparrow$	NIQE $\downarrow$	PI $\downarrow$
3 $\times$ 3	0.513	35.11	7.67	6.81
4 $\times$ 4	0.450	51.44	6.01	5.24
5 $\times$ 5	0.481	53.19	5.33	4.42
8 $\times$ 8	0.516	64.29	6.29	4.89

Table 7: Ablation study on superpixel size.

Panel of Figure 10 — Figure 10: While superpixels effectively suppress degradation artifacts, excessively large superpixel sizes remove fine details and may even alter image content.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

CASISR: Circular Arbitrary-Scale Image Super-Resolution
May 5, 2026Honggui Li, Zhengyang Zhang, Dingtai Li +11Image Super-Resolution
School of Information and Artificial Intelligence, Yangzhou University, Yangzhou 225127, China · Shanghai Qigong Research Institute, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China · LISITE Research Laboratory, Institut Supérieur d'Électronique de Paris, Paris 75006, France +4
DriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting
Oct 3, 2026Wei Zhu, Kai Zhang, Yu Zheng +4Image Super-Resolution Diffusion Models for Image Restoration
Nanjing University of Science and Technology · Huawei · Nanjing University +1
teasr: training-efficient any-step diffusion transformer for real-world image super-resolution
Jun 15, 2026Xiang Gao, Chenxin Zhu, Yushun Fang +2Diffusion Transformer Diffusion Model Distillation
Shanghai Jiao Tong University, China

May 5, 2026J/K move · Enter open · S save

CASISR: Circular Arbitrary-Scale Image Super-Resolution

Honggui Li, Zhengyang Zhang, Dingtai Li, Sinan Chen+10

School of Information and Artificial Intelligence, Yangzhou University, Yangzhou 225127, China · Shanghai Qigong Research Institute, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China · LISITE Research Laboratory, Institut Supérieur d'Électronique de Paris, Paris 75006, France+4

Image Super-Resolution

Open paper AskPDF arXiv

The generalization performance (GP) of deep learning-based arbitrary-scale image super-resolution (ASISR) methods is subject to limited training datasets and unlimited testing datasets. It is vitally significant to enhance the GP of the pretrained ASISR models by making full use of the testing samples. The ASISR models usually employ an open-loop architecture from low-resolution (LR) images to super-resolution (SR) images. The degradation model from SR samples to LR samples is known bicubic down-sampling for the classical ASISR, is supposed down-sampling with additive random noise for the blind ASISR, and is learnable for the real-world ASISR. Combining the ASISR and degradation models, it is potentially possible to adopt a closed-loop architecture based on the automatic control theory for strengthening the GP of the ASISR methods. Therefore, this paper proposes a closed-loop architecture, circular ASISR (CASISR), to lift the capability of image reconstruction. A mathematical nonlinear loop equation is established to describe the CASISR, the reasonability of the CASISR is proven by conditional probability theory, and the stability of the CASISR is proven by Taylor series approximation. The first-order and second-order absolute difference images are defined to compare the image reconstruction performance of the ASISR and the CASISR methods. Comprehensive simulation experiments show that the proposed CASISR approach outperforms the eight state-of-the-art ASISR approaches in the quality of image reconstruction. Especially, the proposed CASISR is extraordinarily suitable for fractional SR scale factors and is extremely effective for text and stripe images with drastically changed edges.

Abstract

Image Super-Resolution

Explore similar work

CASISR: Circular Arbitrary-Scale Image Super-Resolution

DriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting

teasr: training-efficient any-step diffusion transformer for real-world image super-resolution