Scan Clusters, Not Pixels: A Cluster-Centric Paradigm for Efficient Ultra-high-definition Image Restoration

Chen Wu; Ling Wang; Zhuoran Zheng; Yuning Cui; Zhixiong Yang; Xiangyu Chen; Yue Zhang; Weidong Jiang

cs.CVFeb 25, 2026

Scan Clusters, Not Pixels: A Cluster-Centric Paradigm for Efficient Ultra-high-definition Image Restoration

Authors: Chen Wu, Ling Wang, Zhuoran Zheng, Yuning Cui, Zhixiong Yang, Xiangyu Chen, Yue Zhang, Weidong Jiang,

Organizations: National University of Defense Technology · HKUST(GZ) · Qilu University of Technology · Technical University of Munich ·

Abstract

Ultra-High-Definition (UHD) image restoration is trapped in a scalability crisis: existing models, bound to pixel-wise operations, demand unsustainable computation. While state space models (SSMs) like Mamba promise linear complexity, their pixel-serial scanning remains a fundamental bottleneck for the millions of pixels in UHD content. We ask: must we process every pixel to understand the image? This paper introduces C $^2$ SSM, a visual state space model that breaks this taboo by shifting from pixel-serial to cluster-serial scanning. Our core discovery is that the rich feature distribution of a UHD image can be distilled into a sparse set of semantic centroids via a neural-parameterized mixture model. C $^2$ SSM leverages this to reformulate global modeling into a novel dual-path process: it scans and reasons over a handful of cluster centers, then diffuses the global context back to all pixels through a principled similarity distribution, all while a lightweight modulator preserves fine details. This cluster-centric paradigm achieves a decisive leap in efficiency, slashing computational costs while establishing new state-of-the-art results across five UHD restoration tasks. More than a solution, C $^2$ SSM charts a new course for efficient large-scale vision: scan clusters, not pixels.

Figures & tables

Figure 1: The scanning strategies in existing Mamba-based methods and our proposed method. (a) Vmamba [ 21 ] employs a Z-shaped scan path that incurs VRAM bottlenecks when processing UHD images due to its full-pixel scanning. (b) EfficientVMamba [ 23 ] reduces scanning costs by omitting sampling steps, this compromises global modeling accuracy. (c) The proposed cluster-centric scanning strategy.

Figure 2: The overview of our proposed C 2 SSM. C 2 SSM employs an asymmetric U-Net architecture whose decoder integrates the Cluster-Centric Scanning Module and Spatial-Channel Feature Modulator to achieve spatial-channel global feature coupling.

Figure 3: Visual quality comparisons on UHD-LOL4K dataset [ 32 ] . The last row shows the color histogram of the image.

Methods	Type	Venue	PSNR	SSIM	Param
Z_DCE++ [ 17 ]	non-UHD	TPAMI’21	15.58	0.934	79.42K
Uformer [ 33 ]		CVPR’22	29.98	0.980	20.63M
Restormer [ 42 ]		CVPR’22	36.90	0.988	26.11M
NSEN [ 40 ]	UHD	MM’23	29.49	0.980	2.67M
UHDFour [ 16 ]		ICLR’23	36.12	0.990	17.54M
LLFormer [ 32 ]		AAAI’23	37.33	0.988	24.52M

Table 1: Comparison of quantitative results on UHD-LOL4K dataset [ 32 ] .

Methods	Type	Venue	PSNR	SSIM	Param
Z_DCE++ [ 17 ]	non-UHD	TPAMI’21	16.41	0.630	10.56K
Uformer [ 33 ]		CVPR’22	19.28	0.849	20.63M
Restormer [ 42 ]		CVPR’22	22.25	0.871	26.11M
UHDFour [ 16 ]	UHD	ICLR’23	26.22	0.900	17.54M
LLFormer [ 32 ]		AAAI’23	22.79	0.853	13.15M
UHDformer [ 29 ]		AAAI’24	27.11	0.927	0.34M

Table 2: Comparison of quantitative results on UHD-LL dataset [ 16 ] .

Methods	Type	Venue	PSNR	SSIM	Param
RCDNet [ 30 ]	non-UHD	CVPR’20	30.83	0.921	3.17M
SPDNet [ 39 ]		ICCV’21	31.81	0.922	3.04M
IDT [ 38 ]		TPAMI’22	32.91	0.948	16.41M
Restormer [ 42 ]		CVPR’22	33.02	0.934	26.12M
DRSformer [ 3 ]		CVPR’23	32.96	0.933	33.65M
UDR-S2Former [ 2 ]		ICCV’23	33.36	0.946	8.53M

Table 3: Comparison of quantitative results on 4K-Rain13k dataset [ 1 ] .

Figure 4: Visual quality comparisons on 4K-Rain13k dataset [ 1 ] . The last row shows the error map of the image.

Methods	RCDNet [ 30 ]	SPDNet [ 39 ]	IDT [ 38 ]	Restormer [ 42 ]	DRSformer [ 3 ]	NeRD-Rain [ 4 ]	MambaIRv2 [ 10 ]	ERR [ 43 ]	Ours
NIQE	9.997	9.917	9.067	8.636	8.49	9.139	9.397	9.136	8.198
PIQE	63.816	64.774	55.049	60.335	60.441	60.070	62.913	59.735	54.90

Table 4: Comparison of quantitative results on 4K-RealRain dataset [ 1 ] .

Methods	Type	Venue	PSNR	SSIM	Param
MIMO-Unet++ [ 5 ]	non-UHD	ICCV’21	25.03	0.752	16.1M
Restormer [ 42 ]		CVPR’22	25.21	0.752	26.1M
Uformer [ 33 ]		CVPR’22	25.27	0.752	20.6M
Stripformer [ 26 ]		ECCV’22	25.05	0.750	19.7M
FFTformer [ 13 ]		CVPR’23	25.41	0.757	16.6M
UHDformer [ 29 ]	UHD	AAAI’24	28.82	0.844	0.34M

Table 5: Comparison of quantitative results on UHD-Blur dataset [ 29 ] .

Figure 5: Visual quality comparisons on UHD-Blur dataset [ 29 ] . The last row shows the color histogram of the image.

Methods	Type	Venue	PSNR	SSIM	Param
Restormer [ 42 ]	non-UHD	CVPR’22	12.72	0.693	26.11M
Uformer [ 33 ]		CVPR’22	19.83	0.737	20.63M
DehazeFormer [ 25 ]		TIP’23	15.37	0.725	2.5M
MB-TaylorFormer [ 24 ]		ICCV’23	20.99	0.919	2.7M
UHD [ 45 ]	UHD	ICCV’21	18.04	0.811	34.5M
UHDformer [ 29 ]		AAAI’24	22.59	0.942	0.34M

Table 6: Comparison of quantitative results on UHD-Haze dataset [ 29 ] .

Methods	Type	Venue	PSNR	SSIM	Param
Uformer [ 33 ]	non-UHD	CVPR’22	23.71	0.871	20.63M
Restormer [ 42 ]		CVPR’22	24.14	0.869	26.12M
SFNet [ 8 ]		ICLR’23	23.63	0.845	7.05M
UHD [ 45 ]	UHD	ICCV’21	29.29	0.949	34.5M
UHDformer [ 29 ]		AAAI’24	36.61	0.988	0.34M
UHDDIP [ 31 ]		TCSVT’25	41.56	0.990	0.81M

Table 7: Comparison of quantitative results on UHD-Snow dataset [ 31 ] .

Variants	PSNR	SSIM	Param
(a) replace CCSM with ResBlock	37.32	0.989	2.38M
(b) replace CCSM with Vanilla Mamba [ 11 ]	37.43	0.990	3.11M
(c) replace CCSM with ASSM [ 10 ]	38.71	0.991	2.96M
(d) replace SCFM with ResBlock	39.25	0.992	2.69M
(e) remove CCSM	35.87	0.987	2.21M
(f) remove SCFM	39.05	0.992	2.53M

Table 8: Ablation study of proposed blocks on UHD-LOL4K dataset [ 32 ] .

Number of Centers	2	4	6	8
UHD-LOL4K [ 32 ]	37.91/0.992	39.61/0.992	39.23/0.992	39.19/0.992
UHD-Blur [ 29 ]	30.42/0.881	31.53/0.890	31.72/0.890	31.65/0.889
UHD-Haze [ 29 ]	23.27/0.940	24.08/0.942	24.03/0.942	23.86/0.942

Table 9: Ablation study of the number of centers.

Method	MambaIR [ 11 ]	Wave-Mamba [ 48 ]	EVSSM [ 14 ]	MambaIRv2 [ 10 ]	Ours
Parm	1.32M	1.25M	17.13M	0.84M	2.71M
FLOPs	4.774G	0.881G	7.893G	4.940G	0.407G
Venue	ECCV’24	MM’24	CVPR’25	CVPR’25	-

Table 10: Comparison of different scanning strategies. FLOPs are measured with an image of the size

64\times 64

pixels.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

CoDe-SSM: Context-Detail Decoupled State Space Model for Efficient UHD Image Restoration
Jul 31, 2026Jiaxu Su, Zhijian Wu, Jun Li +2Visual SSMs Image Restoration
School of Computer and Electronic Information, Nanjing Normal University, China · Medical Artificial Intelligence Laboratory, Westlake University, China
Ultra-High-Definition Restoration Transformers with Correlation Matching Transformation
Aug 18, 2026Cong Wang, Liyan Wang, Jinshan Pan +4Image Restoration All-in-One Image Restoration
School of Computing and Communications, Lancaster University
From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark
Apr 17, 2026Chen Zhao, Yunzhe Xu, Zhizhou Chen +5Frequency-Domain Image Processing Image Restoration
State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China · School of Computer Science and Technology, Dalian University of Technology, Dalian 116023, China · Department of Computer Science and Engineering, Michigan State University, East Lansing, MI 48824, U.S.A.

Jul 31, 2026J/K move · Enter open · S save

CoDe-SSM: Context-Detail Decoupled State Space Model for Efficient UHD Image Restoration

Jiaxu Su, Zhijian Wu, Jun Li, Bo Zhang+1

School of Computer and Electronic Information, Nanjing Normal University, China · Medical Artificial Intelligence Laboratory, Westlake University, China

Visual SSMs Image Restoration Image Super-Resolution

Open paper AskPDF arXiv

Ultra-high-definition (UHD) image restoration must balance the aggregation of spatially recurring degradation cues with the preservation of localized image structures. Compact aggregation can reduce redundant processing but may attenuate edges, textures, and other fine structures. Existing approaches manage UHD restoration cost through downsampling, window partitioning, or cluster-based token reduction; yet many of them do not explicitly retain information that is poorly represented by shared aggregation. In this study, we propose a Context-Detail Decoupled State Space Model (CoDe-SSM) for UHD restoration, which processes aggregated context and clustering residuals in separate pathways. The context modeling pathway, implemented by the Global Cluster Scan Module (GCSM), aggregates features into

K

input-dependent cluster centers and applies selective SSM reasoning over the resulting fixed-order sequence, enabling cross-region context sharing while decoupling computational cost from spatial resolution. The detail recovery pathway, implemented by the Local High-Frequency Module (LHFM), processes the clustering residual with an input-derived high-frequency mask and a sparse mixture of convolutional experts. Extensive experiments on five UHD benchmarks and five degradation types demonstrate that our explicit context-detail decoupling strategy yields substantial gains in restoration quality while maintaining desirable efficiency.

Abstract

Image Restoration

Explore similar work

CoDe-SSM: Context-Detail Decoupled State Space Model for Efficient UHD Image Restoration

Ultra-High-Definition Restoration Transformers with Correlation Matching Transformation

From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark