Pixel Ignores, Superpixel Sees: Adverse Weather Image Restoration via Semantic-Center SSM

Chi Man Vong; Dayu Li; Jin Wu; Jufeng Yang; Leizhi Shu; Shihao Zhou

REVIEW 4 major objections 5 minor 68 references

Pixel Ignores, Superpixel Sees: Adverse Weather Image Restoration via Semantic-Center SSM

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that adverse weather image restoration improves when a state space model scans the image region-by-region according to superpixel semantics rather than along fixed pixel paths, and reports an average PSNR of 34.15 dB acros

desk verdict Genuinely new superpixel-guided scan idea, strong benchmarks, but the central 'state confinement' claim is contradicted by the paper's own equations; table numbers also don't line up. read the letter →

arxiv 2608.01760 v1 pith:32AFFS27 submitted 2026-08-03 cs.CV

Dayu Li , Shihao Zhou , Leizhi Shu , Jin Wu , Chi Man Vong , Jufeng Yang This is my paper · ORCID

classification cs.CV

keywords adverseweatherimagerestorationall-in-onestatespacemodelsuperpixel-guidedselectivescanregion-levelgatingrainremovalsnowhaze

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SSR is a unified model for restoring images degraded by snow, rain, haze, and raindrops. Its key claim is that the scanning order of a state space model should follow the image's semantic content, not a fixed geometric path. The paper replaces raster, window, and Hilbert scans with a superpixel-guided selective scan: pixels are first grouped into perceptually coherent regions, and the state propagation is confined within those regions. A companion region-level gating mechanism recalibrates channels using each region's mean and variance, allowing the model to suppress degradation outliers without mixing features across object boundaries. On the Snow100K-S, Snow100K-L, Outdoor, and RainDrop test sets, SSR reports the best average PSNR (34.15 dB) and SSIM (0.9442) among the compared methods, using only 7.05M parameters.

What carries the argument

The Superpixel-guided Selective Scan Mechanism (S3M), a content-adaptive replacement for fixed scan trajectories: a superpixel generator assigns each pixel a region label, a permutation operator makes tokens of the same superpixel contiguous, and the SSM recurrence runs within those contiguous blocks. The companion Region-level Gating Mechanism (RGM) uses each region's mean and variance to gate channels before the state update, performing intra-region calibration along the channel dimension.

What would settle it

Compare SSR against the same model with superpixel labels replaced by random regions of the same size on the Outdoor benchmark: if random grouping matches the superpixel gain, semantic coherence is not carrying the improvement. Alternatively, measure the PSNR gap between SSR and a raster-scan SSM baseline on images with progressively denser snow and heavier rain; the gap should shrink or reverse if over-segmentation breaks the proposed mechanism.

Watch

Extended reading notes

Core claim

SSR is a state space model for all-in-one adverse weather restoration. Its central claim is that the order in which an SSM scans the image should be decided by image content, not by a fixed geometric rule. The paper replaces raster, window, and Hilbert scans with a superpixel-guided selective scan: a superpixel generator partitions the feature map into perceptually coherent regions, a permutation operator groups the tokens of each region contiguously, and the SSM propagates its hidden state only within each region. A region-level gating mechanism then computes each region's mean and variance across channels and uses a small MLP to produce per-region channel gates. On the Snow100K-S, Snow100K

Load-bearing premise

The superpixel generator always groups degraded pixels into perceptually coherent regions that match true scene semantics; under dense snow or torrential rain it can over-segment, splitting the image into fragments that the scan then treats as localized noise-fitting paths.

Editorial extensions

If this is right

A single SSR model can handle snow, rain, haze, and raindrop degradations, reporting the best average PSNR and SSIM among compared methods on the four standard test sets.
Confining state propagation to superpixel regions prevents features from different objects from being mixed, which the paper argues is why small structures and text stay sharper after restoration.
Region-level gating adds about 1.09 dB PSNR over no gating on the Outdoor benchmark while adding only 0.04G FLOPs, so per-region calibration is a cheap and effective component.
At 7.05M parameters and 54.27G FLOPs, SSR reaches its reported scores with fewer parameters than the transformer and diffusion baselines it is compared against, suggesting lightweight deployment is feasible.
The superpixel-guided scan and region gating are modular components that can be inserted into any hierarchical SSM encoder-decoder block; the ablations show consistent gains over raster, fixed-window, and Hilbert scans across four datasets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Semantic-guided scanning may transfer to other SSM-based dense prediction tasks where degradation is spatially non-uniform, such as deblurring or low-light enhancement; the paper does not test those tasks.
The fixed superpixel count K=16 could be made adaptive to image content or degradation density, which might address the over-segmentation failure mode the paper identifies for dense snow and torrential rain.
The paper does not isolate the Pearson correlation loss from the superpixel mechanism; a reader could ablate that loss alone to determine how much of the structural gain comes from scanning versus the extra loss term.
If the superpixel generator is the bottleneck, a learned or degradation-robust grouping module would likely yield larger gains than further changes to the SSM architecture itself.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

X is right to flag the S3M mechanism. The paper's central claim is that the superpixel-guided scan 'strictly confines' state propagation within perceptually coherent regions. But Eqs. (4)–(7) don't do that. P_l just reorders tokens; then SSM runs one global recurrence over the reordered sequence. The hidden state h_k in Eq. (7) carries across the boundary between superpixel r and r+1 with no reset or masking. So the scan is a permutation of the pixels, not a set of independent region recurrences. The ablation in Table 5 may simply be showing that an input-dependent permutation and the RGM help, not that the state is semantically isolated. That's a load-bearing inconsistency, not a rhetorical flourish.

What's genuinely new is using superpixel boundaries to define the SSM scan order. The paper is also honest about the practical cost and the failure mode when superpixel segmentation is corrupted by dense snow or rain (Sec. 4.4). The experimental effort is real: six benchmarks, comparisons with recent SSM and transformer baselines, real-world no-reference evals, FLOPs/parameter tables. The efficiency numbers are attractive if they hold.

The soft spots are addressable but real. First, the state-confinement claim has to be fixed, either by actually resetting the hidden state at superpixel boundaries or by rephrasing what the mechanism does. Second, the ablation tables don't reconcile: Table 4's K=16 row gives Outdoor 31.91 dB, Table 5's superpixel row gives 32.13 dB, and the final model in Table 1 gives 33.22 dB, with no stated protocol change. That makes it hard to attribute the gains. Third, K is selected on the test benchmarks, and there are no error bars anywhere. All of that is fixable in a revision.

So the paper deserves a serious referee. The core idea is worth engaging with, but the authors need to either implement the confinement at the equation level or drop the 'strictly confines' language, and they need to reconcile the table numbers. I'd send it to review with the expectation of major revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes SSR, a unified adverse-weather image restoration model built on a superpixel-guided selective scan mechanism (S3M) and a region-level gating mechanism (RGM). The core idea is to replace content-agnostic SSM scanning trajectories (raster, window, Hilbert) with superpixel-based permutations so that state propagation stays within perceptually coherent regions, while RGM calibrates features per superpixel along the channel dimension. The authors report state-of-the-art or competitive PSNR/SSIM on Snow100K-S, Snow100K-L, Outdoor-Rain, RainDrop, and real-world benchmarks, with 7.05M parameters and 54.27G FLOPs, and release code.

Significance. If the claimed mechanism were correctly specified and the benchmark numbers internally consistent, this would be a meaningful contribution: it demonstrates a content-adaptive scanning strategy for SSM-based image restoration with strong quantitative results and a much smaller parameter budget than recent Transformer and Mamba baselines. The paper also includes extensive ablation comparisons across scanning strategies and gating variants, plus real-world no-reference evaluations with an efficiency discussion. However, the central mechanism as written is not actually region-confined, and several ablation numbers disagree without explanation. These issues must be resolved before the empirical claims can be attributed to the proposed components.

major comments (4)

[Sec. 3.2, Eqs. (4)-(7)] The SSM is applied to the full reordered sequence F_sp_seq, and the recurrence h_k = \bar{A}h_{k-1} + \bar{B}\tilde{x}_k carries the hidden state across the boundary between superpixel r and r+1. There is no reset, mask, or per-region initial state in the equations. Consequently, the claim that state propagation is 'strictly confined within perceptually coherent regions' is not supported by the stated formulation. The gain of the superpixel ordering over fixed scans in Table 5 could be due to the input-dependent permutation or to RGM rather than to region-confined state evolution. Please specify a per-region scan (e.g., h_0^{(r)}=0 for each region) or a boundary mask, or revise the conceptual claim accordingly.
[Tables 4, 5, and 1] There is an unexplained numerical inconsistency among the ablations. Table 4 reports K=16 (described as the default) with Outdoor PSNR 31.91 dB, while Table 5 reports the 'Superpixel (Ours)' no-gating row on Outdoor as 32.13 dB, and Table 1 reports the final model as 33.22 dB. The text states that K=16 is the default for all experiments, and Table 5's superpixel row is the no-gating baseline used for the gating ablation, but no protocol difference is given. These numbers must be reconciled or the text must clearly state which configuration each table uses.
[Sec. 3.2 and Table 4] The superpixel generator [55] is cited but not described: the input to clustering (RGB image, feature map, or intermediate features), the resolution at which labels are extracted, the number of iterations, whether the clustering is differentiable, and how K=16 is applied across the three encoder/decoder scales are all unspecified. Moreover, K is selected by maximizing the same benchmark averages reported in Table 1, so the final comparison is partially a test-set-selected result. Please provide implementation details and, if possible, a validation-based selection or a sensitivity analysis that does not use the final test averages.
[Tables 4 and 5, FLOPs reporting] Table 4 reports FLOPs of 4.30G for K=16 and Table 5 reports 4.28G vs. 4.17G for the raster scan, while Table 2 reports full-model FLOPs of 54.27G. The units are inconsistent unless Tables 4-5 report per-block or per-stage FLOPs, but this is not stated. The efficiency comparison in Section 4.2 depends on the correct interpretation, so please define the FLOPs protocol in the ablations.

minor comments (5)

[Title/Abstract] 'guilded' should be 'guided' in the title and abstract.
[Sec. 3.2, paragraph after Eq. (4)] There is a stray word 'kemudian' in the sentence before the inverse permutation; this appears to be a leftover from a non-English draft.
[Table 3] The table header 'RainDSParamNIQE' is run together; please separate the column names for readability.
[References and author names] Several proper nouns are inconsistent or mistyped, e.g., 'CPLP romptIR' should be 'CPLP-PromptIR', 'TP AMI' should be 'TPAMI', and the Histoformer venue is listed inconsistently between Table 1 (ECCV'24) and reference [30] (JOE).
[Sec. 4.2] The reported real-world no-reference differences (e.g., Q-Align scores in Table 3) are very small; please state whether these are averaged over multiple runs and whether any significance/variance information is available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; SSR's claims are empirical benchmark comparisons with no derivation-level circularity.

full rationale

The paper's central claim is an empirical SOTA comparison on external benchmarks, not a first-principles derivation. The S3M mechanism is specified as a permutation operator plus a standard selective-scan recurrence (Eqs. 4-7); the asserted confinement of state propagation to superpixel regions is a design claim that may be internally inconsistent (Eq. 7 updates h across region boundaries without an explicit reset or mask), but this is a correctness/specification concern, not circularity. The superpixel generator is cited from external work [55], and no load-bearing argument reduces to a self-citation. The authors cite their own prior papers (e.g., refs 6, 18, 33, 61-63) only as task-specific related work, not as evidence for SSR's novelty or results. The choice K=16 in Table 4 is made by comparing PSNR/SSIM on the same benchmarks later used for final reporting, which is test-set hyperparameter tuning and a methodological concern, but it is not a fitted parameter renamed as a prediction and does not make the final benchmark comparison circular. The limitation in Sec. 4.4 explicitly acknowledges dependence on superpixel clustering accuracy; that is an honest stated assumption, not a circular step. Overall, no specific equation or argument is equivalent to its own input by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests mainly on the superpixel generator reliably segmenting degraded images and on the benchmark protocol being fair. No new physical or mathematical entities are introduced. The key tunable is K=16, chosen from the test benchmarks themselves.

free parameters (1)

Number of superpixels K = 16
Selected in Sec. 4.3 (Table 4) as the value with highest average PSNR on the same four test benchmarks used for final evaluation; no separate validation set is reported.

assumptions (3)

domain assumption Superpixel labels from [55] reflect semantic boundaries on degraded input features.
Invoked in Sec. 3.2 as the basis for the permutation operator P_l. The paper's own Sec. 4.4 concedes that dense snow or rain can corrupt clustering, so the assumption is load-bearing and only partially tested.
standard math The selective scan SSM from Mamba-style layers maintains valid state propagation on reordered token sequences.
Equations (4) and (7) rely on the SSM parameterization from prior work [9,25]; the paper does not prove properties of the permuted sequence formulation.
domain assumption The benchmark protocol and reported competitor numbers in Tables 1-3 are accurate and fairly comparable.
The paper follows the standard protocol of prior all-in-one methods, but competitor numbers are compiled without running code, and the internal table inconsistency weakens confidence in the compilation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Pixel Ignores, Superpixel Sees: Adverse Weather Image Restoration via Semantic-Center SSM." pith.science (2026). https://pith.science/paper/32AFFS27

@misc{pith2026260801760,
  author       = {Pith},
  title        = {Pith review of: Pixel Ignores, Superpixel Sees: Adverse Weather Image Restoration via Semantic-Center SSM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32AFFS27}},
  note         = {Machine review of arXiv:2608.01760}
}

abstract

Adverse weather image restoration aims to recover clear visibility from degraded images in complex weather conditions. Existing works attempt to address this problem by modeling relationships between pixels, however, this paradigm defies the spatially non-uniformity fact of degradations and learns non-discriminative features from semantic-conflict regions. In this paper, we propose SSR, a \textbf{S}emantic-center guilded \textbf{S}tate space model for image \textbf{R}estoration. The key idea of SSR is to shift the conventional scanning strategy of pixel-serial to semantic-guilded one. Specifically, we introduce a Superpixel-guided Selective Scan Mechanism ($\text{S}^3$M), which first partitions the image into perceptually coherent regions via superpixel clustering and then performs relations modeling within the semantic-related regions. Moreover, a Region-level Gating Mechanism (RGM) is developed to perform intra-region calibration by modulating degradation outliers within each semantic superpixel unit along the channel dimension. Extensive experiments on \textbf{6} well-established benchmarks demonstrate that SSR performs favorably against state-of-the-art models with competitive computational cost.

Figures

Figures reproduced from arXiv: 2608.01760 by the authors.

**Figure 1.** Comparison of scanning mechanisms. (a-c) Existing predefined scan strategies fall into a content-agnostic paradigm, and may learn non-discriminative features from semantic-conflict regions. (d) Our Superpixel-guided Selective Scan Mechanism (S3M) Scan constrains the trajectory within perceptually coherent regions, effectively preventing information leakage and ensuring structure-aware state evolution. – We introd… view at source ↗

**Figure 2.** Overall architecture of the proposed Semantic-center Guided State Space Model (SSR). Our model adopts a hierarchical U-shaped encoder-decoder structure to capture multi-scale features for unified weather restoration. The backbone consists of multiple S3M Blocks, which integrate the Superpixel-guided Selective Scan Mechanism (S3M) and a Region-level Gating Mechanism (RGM) to perform intra-region calibration by mod… view at source ↗

**Figure 3.** Structure of the S3M Block. (a) The S3M block enables structure-aware feature transitions by integrating the Superpixel-guided Selective Scan Mechanism (S3M) and a Feed-Forward Network (FFN). (b) The S3M incorporates two synergistic innovations: Superpixel-guided Scanning, which constrains state propagation within perceptually coherent regions to prevent learning non-discriminative features from semantic-conflict ar… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Illustration of the Region-level Gating Mechanism (RGM). [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Visual comparisons of SSR against state-of-the-art unified methods. Top Row: Raindrop removal [28]. Middle Row: Joint deraining and dehazing [19, 20]. Bottom Row: Image desnowing [24]. schedule [26] over the remaining iterations. The number of blocks at each stage is s…

**Figure 6.** Figure 6: Visualization of the image partitioning with different values of K. As shown, K = 8 leads to overly coarse and mixed semantic features, while K = 32 results in fragmented regions. K = 16 provides the most continuous and coherent semantic representation. Effect of the S…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages

[55]

In: ICCV (2023)

Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023)

work page 2023
[1]

In: ICONIP (2024)

Bai, J., Yin, Y., He, Q., Li, Y., Zhang, X.: Retinexmamba: Retinex-based mamba for low-light image enhancement. In: ICONIP (2024)

work page 2024
[2]

In: CVPR (2022)

Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi- contrastive regularization: Toward a unified model. In: CVPR (2022)

2022
[3]

In: ECCV (2024)

Conde, M.V., Geigle, G., Timofte, R.: Instructir: High-quality image restoration following human instructions. In: ECCV (2024)

work page 2024
[4]

TPAMI (2023)

Cui, Y., Ren, W., Cao, X., Knoll, A.: Image restoration via frequency selection. TPAMI (2023)

work page 2023
[5]

TPAMI (2024)

Cui, Y., Ren, W., Cao, X., Knoll, A.: Revitalizing convolutional network for image restoration. TPAMI (2024)

work page 2024
[6]

In: CVPR (2025)

Duosheng, C., Shihao, Z., Jinshan, P., Jinglei, S., Lishen, Q., Jufeng, Y.: A polarization-aided transformer for image deblurring via motion vector decomposi- tion. In: CVPR (2025)

work page 2025
[7]

In: CVPR (2017)

Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)

work page 2017

Show all 68 references

[8]

In: ICCV (2023)

Gasperini,S.,Morbitzer,N.,Jung,H.J.,Navab,N.,Tombari,F.:Robustmonocular depth estimation under challenging conditions. In: ICCV (2023)

2023
[9]

In: CoLM (2024)

Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. In: CoLM (2024)

2024
[10]

In: CVPR (2025)

Guo, H., Guo, Y., Zha, Y., Zhang, Y., Li, W., Dai, T., Xia, S.T., Li, Y.: Mambairv2: Attentive state space restoration. In: CVPR (2025)

2025
[11]

In: ECCV (2024)

Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. In: ECCV (2024)

2024
[12]

TPAMI (2010)

He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010)

2010
[13]

In: ECCV (2024)

Huang, T., Pei, X., You, S., Wang, F., Qian, C., Xu, C.: Localmamba: Visual state space model with windowed selective scan. In: ECCV (2024)

2024
[14]

In: CVPR (2017)

Jiang, T.X., Huang, T.Z., Zhao, X.L., Deng, L.J., Wang, Y.: A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In: CVPR (2017)

2017
[15]

TIP (2011)

Kang, L.W., Lin, C.W., Fu, Y.H.: Automatic single-image-based rain streaks re- moval via image decomposition. TIP (2011)

2011
[16]

In: CVPR (2025)

Kong, L., Dong, J., Tang, J., Yang, M.H., Pan, J.: Efficient visual state space model for image deblurring. In: CVPR (2025)

2025
[17]

In: CVPR (2022)

Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)

2022
[18]

In: ICML (2025)

Li, D., Liu, Y., Fu, X., Xu, S., Zha, Z.J.: Fouriermamba: Fourier learning integra- tion with state space models for image deraining. In: ICML (2025)

2025
[19]

In: CVPR (2019)

Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR (2019)

2019
[20]

In: CVPR (2020)

Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR (2020)

2020
[21]

In: ECCV (2024)

Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multi- dimensional data. In: ECCV (2024)

2024
[22]

In: ECCV (2018) Adverse Weather Image Restoration via Semantic-Center SSM 17

Li,X.,Wu,J.,Lin,Z.,Liu,H.,Zha,H.B.:Recurrentsqueeze-and-excitationcontext aggregation net for single image deraining. In: ECCV (2018) Adverse Weather Image Restoration via Semantic-Center SSM 17

2018
[23]

In: ICCV (2021)

Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)

2021
[24]

TIP (2018)

Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. TIP (2018)

2018
[25]

NeurIPS (2024)

Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Jiao, J., Liu, Y.: Vmamba: Visual state space model. NeurIPS (2024)

2024
[26]

In: ICLR (2017)

Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. In: ICLR (2017)

2017
[27]

In: ICLR (2019)

Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)

2019
[28]

TPAMI (2023)

Özdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI (2023)

2023
[29]

NeurIPS (2019)

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. NeurIPS (2019)

2019
[30]

JOE (2024)

Peng, Y.T., Chen, Y.R., Chen, G.R., Liao, C.J.: Histoformer: Histogram-based transformer for efficient underwater image enhancement. JOE (2024)

2024
[31]

NeurIPS (2023)

Potlapalli, V., Zamir, S.W., Khan, S.H., Khan, F.S.: Promptir: Prompting for all- in-one image restoration. NeurIPS (2023)

2023
[32]

In: CVPR (2018)

Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)

2018
[33]

In: CVPR (2026)

Qu, L., Zhou, S., Liang, J., Zeng, H., Zhang, L., Yang, J.: It takes two: A duet of periodicity and directionality for burst flicker removal. In: CVPR (2026)

2026
[34]

Physica D (1992)

Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D (1992)

1992
[35]

In: ICCV (2021)

Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: ICCV (2021)

2021
[36]

TCSVT (2025)

Shi, Y., Xia, B., Jin, X., Wang, X., Zhao, T., Xia, X., Xiao, X., Yang, W.: Vmam- bair: Visual state space model for image restoration. TCSVT (2025)

2025
[37]

In: ECCV (2024)

Sun, S., Ren, W., Gao, X., Wang, R., Cao, X.: Restoring images in adverse weather conditions via histogram transformer. In: ECCV (2024)

2024
[38]

In: CVPR (2022)

Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)

2022
[39]

NeurIPS (2025)

Wang, H., Hu, Q., Guo, X.: Modem: A morton-order degradation estimation mech- anism for adverse weather image recovery. NeurIPS (2025)

2025
[40]

In: ICIP (2021)

Wang, H., Wang, W., Liu, J.: Temporal memory attention for video semantic segmentation. In: ICIP (2021)

2021
[41]

IJCV (2024)

Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W., Li, H.: Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. IJCV (2024)

2024
[42]

NeurIPS (2024)

Weng, J., Yan, Z., Tai, Y., Qian, J., Yang, J., Li, J.: Mamballie: Implicit retinex- aware low light enhancement with global-then-local state space. NeurIPS (2024)

2024
[43]

TPAMI (2025)

Wu, G., Jiang, J., Jiang, K., Liu, X., Nie, L.: Beyond degradation redundancy: Contrastive prompt learning for all-in-one image restoration. TPAMI (2025)

2025
[44]

In: CVPR (2021)

Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR (2021)

2021
[45]

In: ACM MM (2024)

Wu, H., Yang, Y., Xu, H., Wang, W., Zhou, J., Zhu, L.: Rainmamba: Enhanced locality learning with state space models for video deraining. In: ACM MM (2024)

2024
[46]

ICML (2024) 18 D

Wu, H., Zhang, Z., Zhang, W., Chen, C., Liao, L., Li, C., Gao, Y., Wang, A., Zhang, E., Sun, W., et al.: Q-align: Teaching lmms for visual scoring via discrete text-defined levels. ICML (2024) 18 D. Li et al

2024
[47]

In: ACCV (2024)

Yamashita, S., Ikehara, M.: Image deraining with frequency-enhanced state space model. In: ACCV (2024)

2024
[48]

TIP (2024)

Yang, Y., Guo, C.L., Guo, X.: Depth-aware unpaired video dehazing. TIP (2024)

2024
[49]

In: CVPR (2019)

Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: CVPR (2019)

2019
[50]

In: ICCV (2023)

Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: ICCV (2023)

2023
[51]

In: CVPR (2021)

Yue, Z., Xie, J., Zhao, Q., Meng, D.: Semi-supervised video deraining with dynam- ical rain generator. In: CVPR (2021)

2021
[52]

In: CVPR (2025)

Zamfir, E., Wu, Z., Mehta, N., Tan, Y., Paudel, D.P., Zhang, Y., Timofte, R.: Complexity experts are task-discriminative learners for any image restoration. In: CVPR (2025)

2025
[53]

In: CVPR (2022)

Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)

2022
[54]

In: CVPR (2021)

Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)

2021
[56]

TIP (2021)

Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. TIP (2021)

2021
[57]

TIP (2021)

Zhang, K., Li, R., Yu, Y., Luo, W., Li, C.: Deep dense multi-scale network for snow removal using semantic and depth priors. TIP (2021)

2021
[58]

In: ECCV (2024)

Zhao, R., Yan, H., Wang, S.: Revisiting domain-adaptive object detection in ad- verse weather by the generation and composition of high-quality pseudo-labels. In: ECCV (2024)

2024
[59]

arXiv preprint arXiv:2402.04139 (2024)

Zheng, Z., Wu, C.: U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139 (2024)

2024 arXiv
[60]

arXiv preprint arXiv:2405.10030 (2024)

Zhou, H., Wu, X., Chen, H., Chen, X., He, X.: Rsdehamba: Lightweight vision mamba for remote sensing satellite image dehazing. arXiv preprint arXiv:2405.10030 (2024)

2024 arXiv
[61]

In: ICCV (2025)

Zhou, S., Li, D., Pan, J., Zhou, J., Shi, J., Yang, J.: Devil is in the uniformity: Ex- ploring diverse learners within transformer for image restoration. In: ICCV (2025)

2025
[62]

TIP (2026)

Zhou, S., Pan, J., Chen, D., Dong, Y., Yang, J.: Rethinking the importance of high-frequency components in transformers for image restoration. TIP (2026)

2026
[63]

TPAMI (2025)

Zhou, S., Pan, J., Yang, J.: Learning an adaptive sparse transformer for efficient image restoration. TPAMI (2025)

2025
[64]

In: ICML (2024)

Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Effi- cient visual representation learning with bidirectional state space model. In: ICML (2024)

2024
[65]

TIP (2024)

Zhu, R., Tu, Z., Liu, J., Bovik, A.C., Fan, Y.B.: Mwformer: Multi-weather image restoration using degradation-aware transformers. TIP (2024)

2024
[66]

In: CVPR (2023)

Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.w.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR (2023)

2023
[67]

In: ACM MM (2024)

Zou, W., Gao, H., Yang, W., Liu, T.: Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement. In: ACM MM (2024)

2024
[68]

In: ACM MM (2024)

Zou,Z., Yu, H.,Huang, J., Zhao, F.: Freqmamba: Viewingmamba froma frequency perspective for image deraining. In: ACM MM (2024)

2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [55]

In: ICCV (2023)

Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with superpixel token interaction. In: ICCV (2023)

work page 2023

[2] [1]

In: ICONIP (2024)

Bai, J., Yin, Y., He, Q., Li, Y., Zhang, X.: Retinexmamba: Retinex-based mamba for low-light image enhancement. In: ICONIP (2024)

work page 2024

[3] [2]

In: CVPR (2022)

Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi- contrastive regularization: Toward a unified model. In: CVPR (2022)

2022

[4] [3]

In: ECCV (2024)

Conde, M.V., Geigle, G., Timofte, R.: Instructir: High-quality image restoration following human instructions. In: ECCV (2024)

work page 2024

[5] [4]

TPAMI (2023)

Cui, Y., Ren, W., Cao, X., Knoll, A.: Image restoration via frequency selection. TPAMI (2023)

work page 2023

[6] [5]

TPAMI (2024)

Cui, Y., Ren, W., Cao, X., Knoll, A.: Revitalizing convolutional network for image restoration. TPAMI (2024)

work page 2024

[7] [6]

In: CVPR (2025)

Duosheng, C., Shihao, Z., Jinshan, P., Jinglei, S., Lishen, Q., Jufeng, Y.: A polarization-aided transformer for image deblurring via motion vector decomposi- tion. In: CVPR (2025)

work page 2025

[8] [7]

In: CVPR (2017)

Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)

work page 2017