Pith.

Computer Vision and Pattern Recognition · arXiv:2611.16310 v1 · submitted 12 Nov 2026 · 49 references

Selective Generative Restoration for Video Deraining via Uncertainty-Aware Diffusion

Yi'ang Chen, Han Li, Laibin Chang, Hui Liu, Kui Jiang, Junjun Jiang

The paper says

“Video deraining is challenging due to rain streaks, occlusions, and atmospheric effects. Existing methods mainly follow a deterministic restoration paradigm, assuming that clean scenes can be uniquely recovered from rainy observations. However, rain degradation is spatially heterogeneous.”

more of the abstract

Reliable regions require faithful reconstruction, whereas severely occluded regions contain ambiguous content that require stronger prior-driven inference. Therefore, conventional deterministic restoration often produces over-smoothed results, while global generative restoration may introduce unnecessary hallucinations. We introduce a Selective Generative Restoration paradigm for video deraining, where generative priors are adaptively activated according to restoration uncertainty. Based on this principle, we propose Selective Generative Video Deraining (SGVD), an uncertainty-aware diffusion framework that enables adaptive fidelity preservation and generative recovery. Specifically, SGVD exploits stochastic diversity between dual one-step diffusion predictions to estimate region-level restoration uncertainty, which determines where observation fidelity should be preserved and where generative inference should be activated. Furthermore, an anchor-guided trajectory consistency mechanism is proposed to regularize stochastic diffusion evolution across frames, reducing temporal generation drift and improving video coherence. To achieve practical efficiency, SGVD adopts one-step diffusion inference while maintaining uncertainty-aware generative capability. Extensive experiments demonstrate that SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. SGVD is a well-engineered system that combines uncertainty-guided selective generation with one-step diffusion for video deraining, but the central uncertainty validation is circular and the performance claims are overstated relative to the numbers. This line is wrong

3 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 49 references say.

The paper says, at Abstract; §IV-B, Table I

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The abstract claims that "SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods," and the conclusion claims "strong reconstruction fidelity." On NTURain-Syn, the dataset used for training and the primary test set, Table I reports SGVD at PSNR 26.52, SSIM 0.881, and LPIPS 0.131, versus PSNR 36.34 and 36.71, SSIM 0.9705 and 0.9684, and LPIPS 0.0257 and 0.0281 for RainMamba and DeLiVR.

more of objection 1

The body text carefully restricts the PSNR/SSIM advantage to "the compared diffusion-based methods" and claims only NIQE, CLIP-IQA, MUSIQ, and WE over DeLiVR and RainMamba, which is internally consistent; the abstract and conclusion, however, do not carry that restriction. A roughly 10 dB gap on the trained dataset cannot be described as superior restoration performance or strong reconstruction fidelity. The authors should reformulate the central claim as an explicit fidelity-perception/temporal trade-off, quantify the cost, and ideally support the perceptual claim with a user study, or otherwise demonstrate that the PSNR gap is a metric artifact.

The paper says, at §IV-D, Table III; Eqs. (7)-(9)

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The uncertainty map M is the load-bearing component of the method, but its validation is self-referential. Table III reports Spearman correlation between M and the per-pixel L1 error of SGVD itself on the same three datasets used for the main evaluation, and for NTURain-Syn the same data distribution used for training.

more of objection 2

A high rank correlation shows that the model's error is spatially structured in a way that correlates with the disagreement of two frozen-backbone predictions; it does not establish that M reflects posterior ambiguity of the clean scene, because edges, occlusion boundaries, and rain-streak density also produce spatially structured error without implying multiple plausible clean contents. Moreover, M is computed from the frozen, unadapted backbone (Eq. 7) while the error is that of the adapted model, so the strong agreement is an empirical observation that the paper does not explain. I recommend three additions: (a) a comparison of M against cheap non-generative error proxies (rain density, gradient magnitude, local motion) on a held-out split; (b) a synthetic setup where ambiguity is known by construction, or a dataset with multiple clean references; and (c) an ablation replacing M in Eq. (10) with a uniform mask or a hand-crafted mask, to show that the spatial structure of M, rather than the mere blending operation, is responsible for the gains in Table VIII.

The paper says, at §III-D, Eq. (10); §IV-B

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

In Eq. (10), the fidelity branch keeps z_lq, the latent of the rainy observation itself, in low-uncertainty regions. Since the target of deraining is the clean video, "preserving observation fidelity" in the fusion condition is in direct tension with the task: low-M regions are precisely where the rain signal is retained in the conditioning.

more of objection 3

This design is a plausible contributor to the large PSNR/SSIM/LPIPS deficit of SGVD on NTURain-Syn in Table I. The paper should clarify what the fused condition z_sel encodes after training, analyze whether the final output in low-M regions still contains residual rain, and report a variant in which the LQ branch is replaced by a deterministic deraining output or by the backbone's own prediction. Without such analysis, the claim that USG "preserves faithful observations" refers to fidelity to the degraded observation rather than to the clean scene.

6 minor objections

The paper says, at Table I

Pith says: minor objection 1 of 6

Several entries in the NTURain-Syn rows are typeset without separators (e.g., "36.711526.5174" and "0.85340.97050.9684"), which makes the table very hard to read; the formatting should be fixed.

The paper says, at Eq. (9)

Pith says: minor objection 2 of 6

The symbol N(·) used for the spatial normalization operator collides with the Gaussian notation N(0,I) used in Eqs. (1) and (6); renaming it (e.g., Norm(·)) would remove an avoidable ambiguity.

show 4 more

The paper says, at §IV-D, Table III

Pith says: minor objection 3 of 6

The paper does not state how the Spearman correlation is aggregated (per-frame then averaged, or pooled over pixels and frames) or whether the values are stable across clips; please specify the aggregation and report clip-level variance or confidence intervals.

The paper says, at §IV-C, Table II

Pith says: minor objection 4 of 6

The downstream detection and tracking table omits DeLiVR and RainMamba, the strongest reconstruction baselines in Table I. Adding them would make the downstream claim more convincing, since otherwise it is unclear whether the moderate gains over the rainy input are attributable to SGVD's selective mechanism specifically.

The paper says, at §IV-A and Conclusion

Pith says: minor objection 5 of 6

All quantitative results appear to be single-run. Given the stochastic probes, random crops, EMA consistency target, and the small training set (25 sequences), mean and standard deviation over at least three training seeds for the main table would substantially strengthen the comparisons.

The paper says, at Conclusion

Pith says: minor objection 6 of 6

The conclusion's phrase "strong reconstruction fidelity" is inconsistent with the NTURain-Syn results in Table I and should be revised in light of the reframing requested in the first major comment.

What would settle it

Construct a synthetic video deraining benchmark where each rainy frame is paired with multiple plausible clean ground-truth frames (e.g., by compositing rain onto different clean backgrounds, or by using a generative model to produce diverse clean versions). Compute the true posterior uncertainty as the pixel-wise variance across the multiple clean references. Then compute the Spearman correlation between SGVD's M and this oracle uncertainty. If the correlation drops significantly below the reported 0.91–0.93, the uncertainty map is not a faithful proxy and the selective generation mechanism is not validated.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 16 s · 20,442 tokens read, 708 written
  3. Editor wrote the verdict above 9 s · 21,339 tokens read, 681 written
  4. Referee major revision: 3 major objections, 6 minor 1 min 22 s · 20,792 tokens read, 12,697 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper introduces Selective Generative Video Deraining (SGVD), which uses the disagreement between two one-step diffusion predictions as a proxy for restoration uncertainty, then fuses the generative output with the input according to that map. The core idea—region-adaptive activation of generative priors based on estimated uncertainty—is sensible and the engineering is thorough: they freeze a large pretrained video diffusion model (Wan2.1), train only a small adapter, and add an anchor-guided trajectory consistency loss to reduce temporal flicker. The ablation studies are well-designed and show that both the uncertainty-guided fusion and the temporal consistency loss contribute. The code is released.

That said, there are real soft spots. The biggest is the validation of the uncertainty map. The paper reports Spearman correlations of 0.91–0.93 between the map and the pixel-wise L1 error of SGVD itself, on the same synthetic datasets used for evaluation. This is circular: the map is derived from the same model whose error it predicts, and the correlation is measured on data the model was trained and tuned on. A high correlation here could simply reflect that the model's error is spatially structured in a way the dual-prediction disagreement captures, without establishing that the map reflects true posterior uncertainty about the clean scene. The authors do not acknowledge this circularity. An independent validation—e.g., on a held-out dataset with known ambiguity, or against an oracle that does not depend on SGVD—would be needed to support the central claim.

Second, the performance claims are overstated. On NTURain-Syn, SGVD trails RainMamba and DeLiVR by large margins on SSIM (0.8810 vs 0.9705 and 0.9684) and LPIPS (0.1310 vs 0.0257 and 0.0281). The paper acknowledges this in passing ("Retaining reliable observations can limit texture regeneration") but the abstract and conclusion still claim "superior" performance. The improvements are on no-reference metrics (NIQE, CLIP-IQA, MUSIQ) and on cross-dataset generalization, which is legitimate but should be presented more carefully.

Third, there are no error bars on any quantitative result. Given the stochastic nature of the method, this is a significant omission.

Who is this for? Researchers working on video restoration with diffusion models will find the selective generation paradigm and the ATC loss worth studying. The paper is technically solid and the code is available. It deserves a serious referee, but the authors should be asked to address the circular validation and to present results with error bars and a more balanced discussion of where their method lags.

Recommendation: conditional accept with major revision. The central idea is interesting, but the uncertainty validation needs independent support and the claims need to be calibrated to the actual numbers.

The strongest objection

The central claim that SGVD selectively activates generative priors based on region-level restoration uncertainty rests entirely on the uncertainty map M (Eq. 9). The paper validates M by computing Spearman correlation (ρ ≈ 0.91–0.93) between M and the pixel-wise L1 restoration error of SGVD itself on the same synthetic datasets (NTURain-Syn, RainSynLight25, RainSynComplex25) used for the main evaluation (Table III). This is a circular validation: the uncertainty map is derived from the same model whose error it is supposed to predict, and the correlation is measured on the same data that the model was trained and tuned on. A high correlation could simply reflect that the model's error is spatially structured in a way that the dual-prediction disagreement captures, without establishing that M genuinely reflects posterior uncertainty about the clean scene. An independent validation would require either (a) a held-out dataset with ground-truth uncertainty (e.g., multiple clean references per rainy frame, or a synthetic setup where the ambiguity is known by construction) or (b) a comparison against an oracle uncertainty estimate that does not depend on the SGVD model itself. Without such validation, the entire selective generation pipeline is built on an untested proxy.

The reader's weakest_assumption correctly identifies that the uncertainty map is validated only via post-hoc correlation on the same datasets used for evaluation. My analysis sharpens this: the correlation is not just on the same datasets, but against the L1 error of SGVD itself, creating a circular validation. The paper does not provide any independent ground-truth uncertainty (e.g., multiple clean references per scene). The authors acknowledge in the limitations (b299) that 'generative priors can recover plausible local details that differ from the clean reference,' which implicitly concedes that the L1 error is not a perfect measure of restoration quality, yet they use L1 error as the target for the uncertainty correlation. This is a load-bearing concern because if M does not faithfully reflect true posterior uncertainty, the selective fusion in Eq. 10 may be activating generative priors in the wrong places or with wrong strength, undermining the core contribution. The concrete test I propose would settle this by comparing against an oracle uncertainty derived from multiple clean references, which is feasible in a synthetic setup. The verdict remains CONDITIONAL because the paper is otherwise well-engineered and the concern is testable, not fatal.

The referee's report

If the central mechanism holds, the paper's paradigm is genuinely useful: it converts a large video diffusion prior into an efficient one-step derainer with only 71.67M trainable parameters (0.98%), treats deraining as spatially heterogeneous rather than uniformly generative, and contributes a training-only temporal regularizer. The engineering is unusually thorough: ablations for stochastic probes (Table VIII), loss weights (Table IX), LQ-latent roles (Table VII), text conditioning (Table X), input-length generalization (Table XI), explicit efficiency profiling (Table VI), downstream task evaluation (Table II), and a code release. The two weaknesses that keep this from being an accept are load-bearing: (i) the uncertainty map, which is the paper's main novelty, is validated only by Spearman correlation with SGVD's own restoration error on the same datasets used for training and main evaluation, which does not independently establish that M tracks posterior ambiguity; and (ii) the abstract's claim of "superior restoration performance" is contradicted by Table I on the primary dataset, where the two deraining-specific baselines beat SGVD by roughly 10 dB in PSNR.

Pith's reading of the paper (a machine's summary, not the authors')

The paper introduces Selective Generative Restoration for video deraining, a paradigm that adaptively balances observation fidelity and generative inference according to region-level restoration uncertainty. The authors propose SGVD, a one-step diffusion framework that estimates uncertainty from the disagreement between two stochastic predictions, then uses that map to preserve reliable regions while activating generative recovery in ambiguous ones. An anchor-guided trajectory consistency mechanism regularizes temporal stability across frames. Experiments show SGVD achieves superior reconstruction fidelity, perceptual quality, and temporal consistency compared to state-of-the-art video deraining methods.

What carries the argument. The central mechanism is the Uncertainty-Guided Selective Generation (USG) module, which computes a spatial uncertainty map M = N(Mean_C(|ˆz_a0 − ˆz_b0|)) from two stochastic clean latent predictions, then fuses the generative proposal ˆz0 with the LQ latent via z_sel = M ⊙ ˆz0 + (1 − M) ⊙ z_lq1. This map controls where the frozen diffusion backbone receives a generative condition versus a fidelity-preserving condition.

Core claim. The central claim is that rain degradation is spatially heterogeneous: some regions retain sufficient visual evidence for faithful reconstruction, while others are so occluded that multiple plausible restorations exist. Therefore, a single restoration strategy cannot work well everywhere. The authors discover that the disagreement between two one-step diffusion predictions from the same rainy observation provides a reliable proxy for posterior uncertainty (Spearman ρ ≈ 0.91–0.93). Using this uncertainty map to selectively blend the generative prediction with the original observation allows the model to avoid both over-smoothing (from deterministic methods) and hallucinations (from global generative methods).

The premise it rests on. The L1 difference between two one-step diffusion predictions with different noise seeds faithfully captures the true posterior uncertainty of the restoration problem, rather than just random variation.

If the paper is right (Pith's inferences, not the authors')

If the uncertainty proxy generalizes, any diffusion-based restoration task with spatially varying degradation (deblurring, inpainting, super-resolution) could adopt selective activation of generative priors.

The anchor-guided trajectory consistency mechanism provides a training-only temporal regularizer that imposes no inference overhead, making it applicable to other video restoration pipelines that suffer from stochastic flicker.

One-step diffusion inference with uncertainty-aware conditioning reduces the computational cost of generative video restoration by ~7× compared to 30-step sampling, enabling real-world deployment.

The selective fusion strategy (Eq. 10) offers a principled way to navigate the fidelity–perception trade-off at the pixel level rather than globally.

The uncertainty map derived from dual stochastic predictions could be used as a saliency signal for downstream tasks (e.g., guiding attention in object detection or tracking), not just for fusion weights.

The same selective generation principle could be extended to video inpainting or deblurring, where degradation is also spatially heterogeneous, without retraining the full backbone.

The paper's limitation that reconstruction errors concentrate on complex textures and motion boundaries suggests that the uncertainty estimator may underestimate ambiguity near edges; a multi-scale or frequency-aware uncertainty could improve fidelity in those regions.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Video deraining is challenging due to rain streaks, occlusions, and atmospheric effects. Existing methods mainly follow a deterministic restoration paradigm, assuming that clean scenes can be uniquely recovered from rainy observations. However, rain degradation is spatially heterogeneous. Reliable regions require faithful reconstruction, whereas severely occluded regions contain ambiguous content that require stronger prior-driven inference. Therefore, conventional deterministic restoration often produces over-smoothed results, while global generative restoration may introduce unnecessary hallucinations. We introduce a Selective Generative Restoration paradigm for video deraining, where generative priors are adaptively activated according to restoration uncertainty. Based on this principle, we propose Selective Generative Video Deraining (SGVD), an uncertainty-aware diffusion framework that enables adaptive fidelity preservation and generative recovery. Specifically, SGVD exploits stochastic diversity between dual one-step diffusion predictions to estimate region-level restoration uncertainty, which determines where observation fidelity should be preserved and where generative inference should be activated. Furthermore, an anchor-guided trajectory consistency mechanism is proposed to regularize stochastic diffusion evolution across frames, reducing temporal generation drift and improving video coherence. To achieve practical efficiency, SGVD adopts one-step diffusion inference while maintaining uncertainty-aware generative capability. Extensive experiments demonstrate that SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-13

The authors' main claim, as the classifier picked it out 0.99

“Extensive experiments demonstrate that SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.13 · human participants 0.02 · controlled comparison 0.56 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.94 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Video deraining is challenging due to rain streaks, occlusions, and atmospheric effects. background 1.00 0.05 0.02 asserted 1.00
Existing methods mainly follow a deterministic restoration paradigm, assuming that clean scenes can be uniquely recovered from rainy observations. background 0.92 0.04 0.02 asserted 1.00
However, rain degradation is spatially heterogeneous. background 0.88 0.12 0.02 asserted 1.00
Reliable regions require faithful reconstruction, whereas severely occluded regions contain ambiguous content that require stronger prior-driven inference. background 0.90 0.09 0.06 asserted 1.00
Therefore, conventional deterministic restoration often produces over-smoothed results, while global generative restoration may introduce unnecessary hallucinations. gap 0.63 0.09 0.07 asserted 0.75
We introduce a Selective Generative Restoration paradigm for video deraining, where generative priors are adaptively activated according to restoration uncertainty. contribution 0.99 0.87 0.23 asserted 1.00
Based on this principle, we propose Selective Generative Video Deraining (SGVD), an uncertainty-aware diffusion framework that enables adaptive fidelity preservation and generative recovery. contribution 0.91 0.85 0.15 asserted 1.00
Specifically, SGVD exploits stochastic diversity between dual one-step diffusion predictions to estimate region-level restoration uncertainty, which determines where observation fidelity should be preserved and where generative inference should be activated. method 0.99 0.92 0.08 asserted 1.00
Furthermore, an anchor-guided trajectory consistency mechanism is proposed to regularize stochastic diffusion evolution across frames, reducing temporal generation drift and improving video coherence. method 0.87 0.91 0.04 asserted 1.00
To achieve practical efficiency, SGVD adopts one-step diffusion inference while maintaining uncertainty-aware generative capability. method 0.91 0.90 0.05 asserted 1.00
Extensive experiments demonstrate that SGVD achieves superior restoration performance, perceptual quality, and temporal consistency compared with state-of-the-art video deraining methods. result 1.00 0.95 0.86 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 8,124 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (7)
Figure 1 of the paper
Figure 1. Motivation of Selective Generative Restoration. Determin
Figure 2 of the paper
Figure 2. Overall architecture of the proposed SGVD for video deraining. The rainy video is first encoded into the latent space by a frozen VAE encoder. The Uncertainty-Guided
Figure 3 of the paper
Figure 3. Visual comparisons on synthetic video deraining datasets, including NTURain [6] and RainSynLight25 [18].
Figure 4 of the paper
Figure 4. Visual comparisons on real-world rainy videos from the NTURain-Real dataset [6].
Figure 6 of the paper
Figure 6. Temporal slice comparison of rainy input, SGVD without ATC, full SGVD, and
Figure 7 of the paper
Figure 7. Restoration examples in challenging scenes. From left to right: rain streak map,

Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.

References (49)
  1. [1] J. Li, R. Xu, X. Liu, J. Ma, B. Li, Q. Zou, J. Ma, and H. Yu, “Domain adaptation based object detection for autonomous driving in foggy and rainy weather,”IEEE Transactions on Intelligent Vehicles, vol. 10, no. 2, pp. 900–911, 2025 on Pith
  2. [2] Y . Pan, R. Sun, W. Li, and T. Zhang, “Exploring weather-aware aggregation and adaptation for semantic segmentation under adverse conditions,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 13 952–13 962 on Pith
  3. [3] J. Liu, “RainyTrack: Enhancing object tracking in adverse weather con- ditions with siamese networks,” inProceedings of the 7th International Conference on Image and Graphics Processing. ACM, 2024, pp. 308– 315 on Pith
  4. [4] X. Zhong, S. Tu, X. Ma, K. Jiang, W. Huang, and Z. Wang, “Rainy WCity: A real rainfall dataset with diverse conditions for semantic driving scene understanding,” inInternational Joint Conference on Artificial Intelligence, 2022, pp. 1743–1749 on Pith
  5. [5] M. Li, Q. Xie, Q. Zhao, W. Wei, S. Gu, J. Tao, and D. Meng, “Video rain streak removal by multiscale convolutional sparse coding,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 6644–6653 on Pith
  6. [6] J. Chen, C.-H. Tan, J. Hou, L.-P. Chau, and H. Li, “Robust video content alignment and compensation for rain removal in a CNN framework,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 6286–6295 on Pith
  7. [7] W. Yang, J. Liu, and J. Feng, “Frame-consistent recurrent video derain- ing with dual-level flow,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1661–1670 on Pith
  8. [8] W. Yang, R. T. Tan, J. Feng, S. Wang, B. Cheng, and J. Liu, “Recurrent multi-frame deraining: Combining physics guidance and adversarial learning,”IEEE Transactions on Pattern Analysis and Machine Intel- ligence, vol. 44, no. 11, pp. 8569–8586, 2022 on Pith
  9. [9] H. Wu, Y . Yang, H. Chen, J. Ren, and L. Zhu, “Mask-guided progressive network for joint raindrop and rain streak removal in videos,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 7216–7225 on Pith
  10. [10] S. Sun, J. Lu, X. Chen, J. Wang, D. Lu, G. Zhang, G. Gao, and Z. Zheng, “DeLiVR: Differential spatiotemporal Lie bias for efficient video derain- ing,” inInternational Conference on Learning Representations, 2026 on Pith
  11. [11] H. Wu, Y . Yang, H. Xu, W. Wang, J. Zhou, and L. Zhu, “RainMamba: Enhanced locality learning with state space models for video deraining,” inProceedings of the 32nd ACM International Conference on Multime- dia, 2024, pp. 7881–7890 on Pith
  12. [12] S. Sun, W. Ren, J. Zhou, S. Wang, J. Gan, and X. Cao, “Semi- supervised state-space model with dynamic stacking filter for real-world video deraining,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 114–26 124 on Pith
  13. [13] X. Liao, T. Zheng, J. Zhong, P. Zhang, and C. Ren, “Asymmetric mask scheme for self-supervised real image denoising,” inEuropean Conference on Computer Vision, 2024, pp. 199–215 on Pith
  14. [14] W. Liu, P. Zhang, H. Qin, X. Huang, J. Yang, and M. Ng, “Fast global image smoothing via quasi weighted least squares,”International Journal of Computer Vision, vol. 132, no. 12, pp. 6039–6068, 2024 on Pith
  15. [15] W. Liu, P. Zhang, Y . Lei, X. Huang, J. Yang, and M. Ng, “A generalized framework for edge-preserving and structure-preserving image smooth- ing,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6631–6648, 2022 on Pith
  16. [16] J.-H. Kim, J.-Y . Sim, and C.-S. Kim, “Video deraining and desnowing using temporal correlation and low-rank matrix completion,”IEEE Transactions on Image Processing, vol. 24, no. 9, pp. 2658–2670, 2015 on Pith
  17. [17] W. Ren, J. Tian, Z. Han, A. Chan, and Y . Tang, “Video desnowing and deraining based on matrix decomposition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 4210–4219 on Pith
  18. [18] J. Liu, W. Yang, S. Yang, and Z. Guo, “D3R-Net: Dynamic routing residue recurrent network for video rain removal,”IEEE Transactions on Image Processing, vol. 28, no. 2, pp. 699–712, 2019 on Pith
  19. [19] Q. Wen, Y . Wu, and Q. Chen, “Video waterdrop removal via spatio- temporal fusion in driving scenes,” in2023 IEEE International Confer- ence on Robotics and Automation, 2023, pp. 10 003–10 009 on Pith
  20. [20] T. Varanka, J. L. Gonzalez, H. Kim, P. Garrido, and X. Yao, “Zero- shot video deraining with video diffusion models,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 677–687 on Pith
show 29 more
  1. [21] P. Yang, H. Ci, B. Lin, Y . Song, and M. Z. Shou, “UENR-600K: A large-scale physically grounded dataset for nighttime video deraining,” arXiv preprint arXiv:2604.04402, 2026 on PithPith's reviewarXiv
  2. [22] Y . Zhu, L. Lv, P. Zhang, X. Liu, T. Tang, F. Tian, W. Sun, and H. Lu, “Interactive spatial-frequency fusion Mamba for multi-modal image fusion,”IEEE Transactions on Image Processing, vol. 35, pp. 2380– 2392, 2026 on Pith
  3. [23] H. Sun, L. Lv, P. Zhang, T. Tang, F. Tian, W. Sun, and H. Lu, “Spatial-frequency enhanced Mamba for multi-modal image fusion,” IEEE Transactions on Image Processing, vol. 34, pp. 7684–7696, 2025 on Pith
  4. [24] C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4713– 4726, 2023 on Pith
  5. [25] J. Wang, Z. Yue, S. Zhou, K. C. K. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,”International Journal of Computer Vision, vol. 132, no. 12, pp. 5929–5949, 2024 on Pith
  6. [26] J. Cai, Y . Lei, K. Wang, T. Zhong, F. Zhou, and J. Song, “Condition- enhanced guided diffusion for image super-resolution,”IEEE Transac- tions on Multimedia, vol. 28, pp. 13 079–13 094, 2026 on Pith
  7. [27] Y . Liu, J. Pan, Y . Li, Q. Dong, C. Zhu, Y . Guo, and F. Wang, “Ul- traVSR: Achieving ultra-realistic video super-resolution with efficient one-step diffusion space,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 7785–7794 on Pith
  8. [28] J. Gu, P.-T. Jiang, H. Zhang, M. Zhou, J. Chen, W. Yang, and B. Li, “Improving consistency in diffusion models for image super-resolution,” IEEE Transactions on Multimedia, vol. 28, pp. 12 864–12 875, 2026 on Pith
  9. [29] J. Han, G. Sim, G. Kim, H.-S. Lee, K. Choi, Y . Han, and S. Cho, “DC- VSR: Spatially and temporally consistent video super-resolution with video diffusion prior,” inProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference, 2025, pp. 1–11 on Pith
  10. [30] A. Kim, M. Jang, C. Moon, Y . Shin, J. Jeong, and S.-h. Park, “FiDeSR: High-fidelity and detail-preserving one-step diffusion super-resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 38 270–38 280 on Pith
  11. [31] Y . Zhu, Z. Zhu, M. Hašan, J. Yang, J. Xie, and B. Wang, “Weath- erDiffusion: Weather-guided diffusion model for forward and inverse rendering,”arXiv preprint arXiv:2508.06982, 2025 on PithPith's reviewarXiv
  12. [32] J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021 on Pith
  13. [33] C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 5775–5787. IEEE TRANSACTIONS ON MULTIMEDIA 12 on Pith
  14. [34] Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inInternational Conference on Learning Representations, 2023 on Pith
  15. [35] X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to gen- erate and transfer data with rectified flow,” inInternational Conference on Learning Representations, 2023 on Pith
  16. [36] Team Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P. Li, P. Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Wen... on PithPith's reviewarXiv
  17. [37] Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inInternational Conference on Machine Learning, 2023, pp. 32 211– 32 252 on Pith
  18. [38] S. Luo, Y . Tan, L. Huang, J. Li, and H. Zhao, “Latent consistency models: Synthesizing high-resolution images with few-step inference,” inInternational Conference on Learning Representations, 2024 on Pith
  19. [39] Y . Wang, W. Yang, X. Chen, Y . Wang, L. Guo, L.-P. Chau, Z. Liu, Y . Qiao, A. C. Kot, and B. Wen, “SinSR: Diffusion-based image super- resolution in a single step,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 25 796–25 805 on Pith
  20. [40] R. Wu, L. Sun, Z. Ma, and L. Zhang, “One-step effective diffusion network for real-world image super-resolution,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 92 529–92 553 on Pith
  21. [41] C.-H. Yeh, H.-S. Shiu, C.-Y . Lin, Z. Wang, C.-W. Hsiao, T.-H. Chen, and Y .-L. Liu, “DiffIR2VR-Zero: Zero-shot video restoration with diffusion- based image restoration models,”arXiv preprint arXiv:2407.01519, 2024 on PitharXiv
  22. [42] Y . Liu, Z. Ke, F. Liu, N. Zhao, and R. W. H. Lau, “Diff-Plugin: Re- vitalizing details for diffusion-based low-level tasks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4197–4208 on Pith
  23. [43] S. Chen, T. Ye, K. Zhang, Z. Xing, Y . Lin, and L. Zhu, “Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint,” inEuropean Conference on Computer Vision, 2024, pp. 95–115 on Pith
  24. [44] X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y . Qiao, W. Ouyang, and C. Dong, “DiffBIR: Toward blind image restoration with generative diffusion prior,” inEuropean Conference on Computer Vision, 2024, pp. 430–448 on Pith
  25. [45] Z. Chen, Z. Zou, K. Zhang, X. Su, X. Yuan, Y . Guo, and Y . Zhang, “DOVE: Efficient one-step diffusion model for real-world video super- resolution,” inAdvances in Neural Information Processing Systems, vol. 38, 2025, pp. 94 905–94 924 on Pith
  26. [46] Z. Teed and J. Deng, “RAFT: Recurrent all-pairs field transforms for optical flow,” inEuropean Conference on Computer Vision, 2020, pp. 402–419 on Pith
  27. [47] J. Redmon and A. Farhadi, “YOLOv3: An incremental improvement,” arXiv preprint arXiv:1804.02767, 2018 on PithPith's reviewarXiv
  28. [48] Y . Chen, Y . Cao, H. Hu, and L. Wang, “Memory enhanced global- local aggregation for video object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 10 337–10 346 on Pith
  29. [49] X. Zhou, V . Koltun, and P. Krähenbühl, “Global tracking transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8771–8780 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment