acceptodds
Under review as a conference paper at ICLR 2027

FluidRain: Divergence-Free Rain Flow as an Attention Bias for Loop-in-Loop Video Deraining

Abstract

Existing video deraining methods typically exploit neighbouring frames through either explicit alignment or implicit spatiotemporal aggregation. Explicit alignment relies on accurate motion estimation, which can become unreliable under dense rain, while implicit aggregation avoids alignment but lacks explicit guidance on the directional and temporally coherent structure of rain. This leaves a gap between reliable temporal aggregation and explicit modeling of rain motion. We propose FluidRain, a parameter-efficient video derainer that uses divergence-free rain flow to guide Loop-in-Loop attention across scales and neighbouring frames. Motivated by fluid mechanics, FluidRain models rain motion as an image-space flow, projects the estimated field onto the divergence-free subspace, and converts its local direction and extent into an anisotropic attention bias. The same rain-aware operator is then reused across neighbouring frames and three resolutions in a Loop-in-Loop design. The resulting three-frame model has only 0.80M parameters and remains competitive with substantially larger restoration networks on four benchmarks. Controlled view experiments show that the shared operator benefits most from complementary neighbouring observations rather than repeated computation. To evaluate whether the model remains reliable when rain motion changes across frames, we introduce RainSyn-Gust, which injects controlled changes in rain-streak direction into existing benchmarks. We also use a physics-inspired reference-free diagnostic to analyze removal on real rainy videos.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video deraining, weight-tied recurrence, parameter-efficient architectures, window attention, temporal modelling, low-level vision, video restoration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_voePOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FluidRain, a compact video deraining model that converts an estimated divergence-free rain-flow field into an anisotropic attention bias. A Loop-in-Loop architecture shares parameters across neighbouring frames and multiple scales. The model uses 0.80M parameters and is validated through four benchmarks, controlled architectural ablations, changing rain directions, and real-rain transfer.

Critical strengths

1.The model achieves competitive restoration with only 0.80M parameters, including useful transfer to WeatherBench.

2.Controlled studies on complementary observations, weight sharing, projection, and directional interventions provide substantial evidence for the design.

Critical weaknesses

1.All baselines are retrained using a common 20k-iteration schedule and small crops, without convergence curves or comparisons against adequately trained checkpoints. Different architectures may require different optimization recipes, so the reported rankings primarily establish performance under this particular training budget.

2.The divergence-free derivation assumes restricted camera motion and depth conditions. Although these limitations are acknowledged, direct rain-flow accuracy measurements and tests under assumption violations are missing. Projection removal alone cannot distinguish physically meaningful modeling from generic regularization.

3.DeLiVR already explores geometric motion cues in attention. Comparing complete networks does not reveal how much improvement comes specifically from the proposed bias rather than the backbone and parameter-sharing strategy. A matched-backbone comparison would clarify the methodological advance.

4.Table 1 reports 271 ms for FluidRain, compared with 75 ms for RainMamba and 20 ms for NAFNet. These results warrant a clearer accuracy–latency analysis, together with computational cost and peak memory measurements.

Questions for authors

1.Do the main rankings persist with longer training or official checkpoints evaluated under a verified common protocol? Please provide convergence evidence for the strongest video baselines.

2.Within the same backbone, how does the proposed pathway compare with a generic smoothness-regularized directional field and a DeLiVR-style attention bias?

3.Can you quantify rain-flow direction accuracy before and after projection, and identify whether projection helps or hurts under camera motion that violates the stated assumptions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer voeP

We thank the reviewer for the constructive assessment and for recognizing FluidRain’s competitive restoration with only 0.80M parameters, its transfer to WeatherBench, and the controlled studies supporting our design. Your suggestions help us clarify the role and limitations of the rain-flow prior.

1. W1 & Q1: Training duration and convergence evidence

We examined the NTURain training curves for FluidRain, S2VD, and RainMamba. During the final 2k updates, FluidRain and RainMamba show little change in monitored PSNR, while S2VD shows no further improvement. Their losses also decrease only slightly. For FluidRain, the monitored PSNR gain over the entire final 4k updates is approximately 0.05 dB.

These observations provide evidence that performance stabilizes toward the end of the 20k-update schedule. They address the concern about missing training diagnostics, although they do not establish whether the rankings persist under longer training or official checkpoints. We will include the training-loss, learning-rate, and monitored-PSNR curves, together with the monitoring settings, in the supplementary material.

2. W2 & Q3: Direction accuracy and camera-motion violations

We added direct measurements of direction accuracy before and after projection, together with controlled camera-motion tests.

Direction accuracy. We compare raw and projected directions from the same forward pass after separately averaging flow vectors within each s=1s=1 attention window. Angular errors are measured against structure-tensor targets recomputed per window, using 180∘180^\circ equivalence, identical valid windows, and equal weighting across videos. We use the same model weights as in the pointwise evaluation.

Table R3. Direction error before and after projection (° ↓).

DatasetBeforeAfterChange
NTURain11.04110.248−0.793
RainSynComplex17.94816.752−1.196

Projection improves direction accuracy on both evaluated benchmarks.

Camera-motion tests. We rendered eight scenes under static, forward-motion, and yaw conditions, with 40 evaluated centre frames per condition. Rainy and clean pairs use identical camera trajectories and exposure, with raindrops included only in the rainy version. Checks using the known trajectories confirm nonzero divergence in the forward-motion and yaw conditions.

Table R4. Controlled camera-motion results.

Camera conditionDirection error: before → after (°)ΔPSNR: Full − No-Π (dB)
Static4.704 → 3.896−0.712
Forward motion8.519 → 8.576+0.429
Yaw rotation10.038 → 9.975+0.438

Table R3 shows improved direction accuracy after projection on both benchmarks. Under the camera motions that violate the assumptions in Table R4, projection slightly worsens direction accuracy for forward motion (+0.057°) and slightly improves it for yaw (−0.063°). Nevertheless, Full achieves higher regional PSNR than No-Π in both conditions, by 0.429 and 0.438 dB, respectively.

3. W3 & Q2: Matched-backbone comparison with DeLiVR-style bias

We evaluate a DeLiVR-style SO(2) spatiotemporal-bias adapter within the same Loop-in-Loop backbone.

Table R5. Matched-backbone comparison on RainSynComplex.

Geometric guidancePSNR ↑ (dB)SSIM ↑
DeLiVR-style spatiotemporal-bias adapter28.660.8669
FluidRain30.180.8980

FluidRain improves PSNR by 1.52 dB and SSIM by 0.0311. The adapter uses framewise SO(2) rotations to construct the spatial attention bias and incorporates angular displacement into the existing frame-fusion logits, retaining the other flow-dependent components. FluidRain instead parameterizes local anisotropic attention through the direction and extent of the projected rain-flow field. These results support the advantage of the proposed geometric guidance over the tested DeLiVR-style alternative when the backbone and parameter-sharing strategy are held fixed.

AuthorsOpenReview ↗

Manuscript revisions

Thank you again for your constructive suggestions. We have incorporated the additional analyses discussed in our response into the revised manuscript:

  1. Training diagnostics: Appendix C.5 and Figure 11 now include the training curves and the monitoring setup.
  2. Rain-flow evaluation: Section 4.4 (Tables 5–7) now includes the projection and attention-bias controls, direction-accuracy measurements, and controlled camera-motion tests. Appendix D.7 provides the evaluation protocols.
  3. Matched-backbone comparison: Appendix D.6 (Table 13) now includes the DeLiVR-style spatiotemporal-bias comparison within the same Loop-in-Loop backbone.

We appreciate the time you have devoted to reviewing our work and would be grateful for your feedback on these revisions and any remaining concerns.

Reviewer_4MF5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FluidRain, a 0.80M-parameter, three-frame video deraining network with two components. (i) A rain-flow pathway: a small convolutional head (6.2K parameters) predicts a per-frame "rain-flow" field at quarter resolution, which is projected onto the divergence-free subspace with an FFT-based Helmholtz–Hodge projection. The projection is motivated by a derivation (Prop. 1) that the image-space motion of rain drops in a single depth layer is divergence-free under incompressible horizontal wind and a laterally translating, non-rotating camera. The window-averaged projected flow defines an anisotropic additive attention bias that penalises across-streak displacements and normalises along-streak displacements by a flow-dependent length. (ii) A "Loop-in-Loop" backbone: one 8-block window-attention stack is reused across three scales (outer loop), and within each block the same attention projections read keys/values from the centre frame and its two neighbours (inner loop), fused with weights predicted from a clip descriptor that includes a summary ψ(Δu) of the frame-to-frame rain-flow change. The rain-flow head is supervised only on NTURain (λ_flow = 0.01) and learned purely through the reconstruction loss elsewhere.

Validation consists of comparisons with 11 image/video baselines, all retrained under a unified 20k-iteration protocol on NTURain, RainSynLight, RainSynComplex and RainSynAll100; view/weight-sharing controls (Table 2); rain-flow-pathway ablations and an inference-time flow-rotation intervention (Tables 3–4); a RAFT alignment control (Table 5); a new RainSyn-Gust benchmark with synthetic changes of rain direction; a reference-free diagnostic F_phys for real rain; zero-shot WeatherBench evaluation; KITTI downstream detection/segmentation; and fine-tuning to dehazing/deblurring.

Critical strengths

S1. A well-controlled study of where multi-frame weight sharing pays off. Table 2 holds the parameter count fixed (0.79M) and separates repeated computation (center repeated: small, inconsistent gains), changed spatial view, and genuinely new temporal observations. Reading a second neighbor with tied weights yields 0.93–1.20 dB, whereas untying the per-scale weights costs 2.5× parameters for 0.03–0.59 dB. Combined with the occlusion statistics in Appendix A.3 (≈0.98 / 0.70 of center-frame rain pixels are clean in at least one neighbor for light / heavy rain), this gives a clear, falsifiable explanation of why a tiny shared operator can work for video deraining. This is consequential because it is the most robust and transferable finding of the paper, independent of the physics framing.

S2. Favorable parameter–quality trade-off with a broad evaluation. Within the authors' protocol, a 0.80M model is close to or above 20–30M restoration networks on three of four benchmarks, and the evaluation is unusually broad (alignment control, a direction-change stress test, zero-shot real rain, downstream perception). This would support acceptance if the comparison protocol were sound (see W2), which is why it weighs less in my assessment than S1.

Critical weaknesses

W1. The central claim — that a divergence-free rain-flow prior drives the gains — is not established, and the implementation is largely disconnected from the theory. • On the benchmarks used, rain direction is essentially constant within each sequence (0.03°–0.05° variation, Appendix A.3), and the attention bias only consumes the window-averaged flow u_w. A nearly uniform field is trivially divergence-free, so Π effectively acts as a spectral regulariser / low-pass filter on an unconstrained latent field. On three of four benchmarks λ_flow = 0, so the "rain flow" is never tied to rain geometry, and the accuracy of the estimated flow is never evaluated. • Missing controls: replacing Π with non-physical regularisers (Gaussian low-pass, TV, global mean flow); a fixed vertical or a single learned global-orientation anisotropic bias; a learned anisotropic bias with matched parameters but no flow input.

W2. The comparison protocol does not support the claim of being "competitive with substantially larger restoration networks." • All baselines are retrained for only 20k iterations with batch size 4 and 128² crops — a small fraction of the original schedules of models such as Restormer or DRSformer. Short schedules systematically favour small, fast-converging models. • The retrained baseline numbers are far below published values: S2VD reports 37.37 dB and RainMamba 37.87 dB on NTURain in their original papers, versus 34.17 and 32.65 dB here. S2VD gets 25.36 dB on RainSynLight, and RainMamba is below SwinIR-light on every benchmark — symptoms of under-training or broken configurations. FluidRain's 35.73 dB on NTURain is about 2 dB below published state of the art.

Questions for authors
  1. Baseline protocol. Can you report baselines trained with their official schedules (or published numbers on matching splits) together with FluidRain trained to convergence? Please explain the 3–5 dB gap between your retrained S2VD / RainMamba results on NTURain and the published 37.37 / 37.87 dB, and why S2VD reaches only 25.36 dB on RainSynLight.

  2. Is the gain physical? Please replace Π with non-physical alternatives at the same cost (Gaussian low-pass of û, TV smoothing, a global-mean flow) and add a fixed-direction (vertical) anisotropic bias. If Π clearly beats these, my concern in W1 would be largely resolved.

  3. Flow accuracy. What is the angular error of the predicted orientation against the structure-tensor targets on NTURain, and on the RainSyn benchmarks where the head is unsupervised?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4MF5

We thank the reviewer for the constructive assessment and recognition of our controlled evidence on temporal observations and weight sharing, broad evaluation, and favorable parameter–quality trade-off. These comments help us strengthen the evaluation of the rain-flow prior and clarify the scope of our baseline comparisons.

1. W1 & Q2: Projection versus generic regularization and anisotropic biases

We independently retrained all six alternatives under the protocol in Appendix C.2. FluidRain achieves 35.73 dB on NTURain, outperforming every alternative in Table R1.

Projection replacements. Gaussian filtering, TV smoothing, and global-mean flow achieve 35.28, 35.34, and 35.16 dB, respectively. The proposed projection improves over the strongest replacement by 0.39 dB.

Attention-bias alternatives. Fixed vertical and globally learned orientations achieve 35.11 and 35.24 dB. The approximately parameter-matched learned bias without direct flow input reaches 35.39 dB, 0.34 dB below FluidRain. Thus, the tested smoothing and anisotropic-bias alternatives do not recover the full model's performance.

The mathematical distinction is explicit in Eq. (29):

Πk=I−kk⊤∥k∥2,k≠0,Π0=I.\Pi_{\mathbf{k}}=I-\frac{\mathbf{k}\mathbf{k}^{\top}}{\lVert\mathbf{k}\rVert^{2}},\quad \mathbf{k}\neq0,\qquad \Pi_{\mathbf{0}}=I.

The projection preserves divergence-free components at all frequencies, unlike radial low-pass filtering. The experiments therefore support the benefit of this specific structural prior and its use in attention.

Table R1. Projection and attention-bias controls on NTURain.

Control groupConfigurationPSNR ↑ (dB)
ReferenceFluidRain (full)35.73
Projection replacementGaussian low-pass filtering35.28
TV smoothing35.34
Global-mean flow35.16
Attention biasFixed vertical orientation35.11
Single learned global orientation35.24
Learned anisotropic bias without direct flow input35.39

Projection controls replace Π throughout the flow pathway; bias controls change only the attention bias, retaining the remaining flow components.

2. W2 & Q1: RainMamba training objective and comparison protocol

In Table 1, RainMamba was trained with a shared L1 reconstruction loss. We retrained the adapted model from scratch for 20k updates using RainMamba's original loss components: Charbonnier, perceptual, and dynamic contrastive losses, weighted by 1, 0.3, and 0.1. We retained its architecture and training data, restored five-frame supervision, and scaled the contrastive schedule to 20k updates.

NTURain results (RGB PSNR/SSIM).

MethodTraining objectivePSNR ↑SSIM ↑
RainMambaL132.650.9424
Original loss components34.170.9551
FluidRainOurs35.730.961

RainMamba’s lower result in Table 1 partly reflects the use of L1 loss in our retraining. Restoring its original objective improves PSNR by 1.52 dB and SSIM by 0.0127, while FluidRain retains a 1.56 dB PSNR advantage under the same 20k-update budget. We will include these additional results in the supplementary material.

Training stabilization. Retrospective diagnostics on 16 fixed NTURain test clips show that FluidRain gains only approximately 0.05 dB over the final 4k updates, with negligible improvement over the final 2k. Its reconstruction loss also levels off. The original S2VD, RainMamba (L1), and Restormer runs show final-4k gains below 0.22 dB. These measurements provide evidence of late-stage stabilization under the evaluated schedule.

We will include these additional results and the training-loss, learning-rate, and monitored-PSNR curves in the supplementary material.

3. Q3: Accuracy of the predicted rain-flow orientation

We evaluate projected rain-flow orientations pointwise on the quarter-resolution grid, computing angular errors against structure-tensor targets at individual grid locations and then averaging these errors. Orientations differing by 180° are treated as equivalent. FluidRain achieves lower mean angular errors than the fixed-vertical reference on all four benchmarks, indicating closer alignment with the reference rain-streak orientations.

Table R2. Pointwise rain-flow orientation errors on the quarter-resolution grid (↓).

DatasetFluidRain (projected flow)Fixed vertical
NTURain6.4°13.28°
RainSynLight8.3°17.45°
RainSynComplex10.8°12.90°
RainSynAll10011.7°13.51°

Thanks for the new results reported here.

AuthorsOpenReview ↗

Manuscript revisions

Thank you for your follow-up. We have incorporated the additional results and clarifications into the revised manuscript:

  • Projection and attention-bias controls: Section 4.4 and Table 5, with the projection clarification in Section 3.1.
  • RainMamba objective comparison and training diagnostics: Appendices C.3 and C.5, Table 11, and Figure 11.
  • Pointwise flow-orientation evaluation: Appendix D.7 and Table 14, with the attention-window evaluation reported separately in Table 6.

Thank you again for the time and care you have devoted to reviewing our work. Your suggestions have helped us strengthen the experimental evidence and make the manuscript clearer. If these additions have addressed your main concerns, we would be grateful if you could consider raising your score. We would also welcome the opportunity to clarify any concerns that remain.

Reviewer_4n9nOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper aims to enhance the visibility of rain videos, as poor visibility in rain is a problem for many downstream tasks, such as video surveillance, object detection, and segmentation. Existing methods rely heavily on explicit alignment or implicit spatiotemporal aggregation to build their deraining frameworks. However, accurate explicit alignment is an open problem under adverse weather conditions, while implicit aggregation avoids alignment but lacks explicit guidance on the directional and temporally coherent structure of rain. To address these issues, this paper proposes FluidRain, which uses divergence-free rain flow to guide Loop-in-Loop attention across scales and neighbouring frames. The experimental results show that the proposed FluidRain achieves promising performance on synthetic datasets.

Critical strengths
  1. This paper introduces a fluid-mechanical rain-flow representation, which is a very interesting idea for video deraining. To the best of my knowledge, this is the first paper that uses fluid-mechanical representations to enhance temporal aggregation in degraded videos.

  2. The proposed framework is parameter-efficient. The whole video deraining framework requires only 0.8M parameters while maintaining promising results on synthetic datasets.

  3. The experimental setup is easy to understand and follow, which improves the reproducibility of this paper.

Critical weaknesses
  1. This paper focuses only on rain-streak removal, which is not the key challenge for rain videos. Rain fog/haze is the most important degradation in rain videos. Dense rain always comes with haze effects, which significantly weaken the background signal. This paper aims to enhance the visibility of rain videos. However, the authors do not consider the rain accumulation problem, which is really disappointing. The real rain video shown in the supplementary material demonstrates that there is a severe haze effect under dense rain conditions.

  2. This paper shows only one real-world rain video in the supplementary material, which is not enough. Since there is a significant domain gap between real-world and synthetic rain degradations, a model that works well on synthetic datasets may not generalize to real-world rain videos. Therefore, it is really necessary to show as many real-world deraining results as possible. However, this paper focuses more on the synthetic results.

  3. The training datasets are synthetic and simulate ideal cases of rain streaks. However, in the real world, rain streaks are more complex and may contain noise. How does the proposed method handle such cases? Can the fluid-mechanical rain-flow representation trained on synthetic datasets generalize to real-world rain-streak scenes?

Questions for authors
  1. Following Weakness 1, the authors should clarify that this paper focuses only on rain-streak removal, not on video deraining as a whole. The proposed fluid-mechanical representation seems to be designed for rain streaks only, which is too narrow. I am wondering whether the proposed method can be generalized to haze/fog effects. To be honest, addressing rain streaks only is not impressive enough at present. For example, the visibility of the derained real rain video in the supplementary material is still very low.

  2. Following Weakness 2, the authors should provide as many real-world deraining results as possible. Currently, the real-world deraining results are not impressive. There is not much difference between the proposed method and the other methods.

  3. Following Weakness 3, I am wondering whether the authors have considered these domain-gap problems. Why would the fluid-mechanical rain-flow representation still be robust to real-world rain streaks? Does the current method have any mechanism to overcome this? If not, how can this domain-gap problem be prevented?

  4. Most of the PSNR/SSIM results are based on synthetic data. The authors should also conduct some quantitative evaluations on real data, for example, using no-reference metrics or VLM-based evaluation. However, the most important thing is to provide real-world qualitative results.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4n9n

Thank you for recognizing the potential of our rain-flow formulation and its parameter efficiency. We address your concerns about rain accumulation, real-world evaluation, and synthetic-to-real generalization below.

1. W1 & Q1: Rain accumulation and dehazing

We agree that rain-induced haze is important for visibility restoration. While FluidRain’s physical prior describes visible rain-streak motion, our evaluation also includes rain accumulation: on RainSynAll100, FluidRain achieves 35.64 dB PSNR and 0.948 SSIM (Table 1). Appendix D.5 further evaluates dehazing using the unchanged architecture after 3,000 iterations of task-specific fine-tuning, improving PSNR from 16.28 to 19.88 dB and SSIM from 0.727 to 0.849. Figure 14 illustrates reduced haze and improved background contrast.

These results support the architecture’s adaptability beyond rain-streak removal under the evaluated setting; they do not imply that the rain-flow prior explicitly models atmospheric scattering or that dehazing requires no adaptation. We acknowledge the residual haze in the supplementary video and will clarify this limitation while making the existing dehazing evidence more prominent.

2. W2, Q2 & Q4: Real-world evaluation

Our evaluation extends beyond the single supplementary video. It includes seven real NTURain sequences, WeatherBench, and captured rainy scenes, with qualitative comparisons in Figures 7, 15, and 17. For NTURain without clean references, Table 10 reports FphysF_{\mathrm{phys}} precision/recall of 0.858/0.131, compared with 0.848/0.117 for the backbone without the rain-flow pathway. This reference-free diagnostic characterizes removal behavior, with its limitations discussed in Appendix E.4.

We also evaluate NTURain-trained checkpoints on all 200 paired WeatherBench test images without fine-tuning. Table 11 reports 28.37 dB PSNR, 0.8823 SSIM, and 0.3232 LPIPS on WB200: the best PSNR and LPIPS, and joint-best SSIM among the compared methods. On WB145, which retains distinct neighboring observations, FluidRain achieves 29.27 dB / 0.8616 / 0.3074, ranking second across all three metrics. The rain-flow pathway improves PSNR over the same backbone by 0.51 dB on WB200 and 0.22 dB on WB145.

These reported results support transfer to the evaluated real captures with only 0.80M parameters, although visual improvements vary across scenes. We will consolidate this evidence and expand the supplementary presentation with side-by-side videos showing residual rain, detail preservation, and temporal behavior.

3. W3 & Q3: Generalization mechanisms and limitations

FluidRain estimates a field from each observed frame at quarter resolution and uses its window-averaged direction and extent to guide attention. The Helmholtz–Hodge projection removes its divergent component:

\qquad \nabla^2\phi_\tau=\nabla\cdot\hat u_\tau.$$ The projected field enters as an additive bias alongside content-based attention, followed by learned temporal fusion: $$A_\delta= \operatorname{Softmax}\left( \frac{QK_\delta^\top}{\sqrt d} +B_{\mathrm{rel}} +B_{\mathrm{flow}}\left(\frac{u_{t+\delta}}{s}\right) \right)V_\delta.$$ $$X_s\leftarrow X_s+\sum_{\delta\in\lbrace 0,-1,+1\rbrace}g_\delta A_\delta,$$ where $s$ denotes the spatial scale and $g_\delta$ are learned fusion weights. Restoration therefore combines directional guidance, image content, and complementary neighboring observations without requiring exact rain-streak correspondences. This provides a structural inductive bias, rather than an explicit domain-adaptation mechanism. The projection is not a general-purpose noise filter, and under departures from the ideal camera-motion and depth assumptions, it serves as a structured prior. The real-data results above provide empirical transfer evidence; they do not guarantee robustness to arbitrary noise or elimination of the domain gap. We will clarify this distinction.
AuthorsOpenReview ↗

We Have Revised the Manuscript

Following our response, we have revised the manuscript, with the relevant changes highlighted in blue:

  1. Dehazing clarification. Appendix D.5 now clarifies the fine-tuning procedure and evaluation setup for dehazing.
  2. Real-rain evidence. Section 4.7 and Appendices E.3–E.5 present the evidence discussed in our response, including quantitative results in Tables 16–17 and qualitative comparisons in Figures 7, 16, and 18. We would appreciate it if you could review our response together with these revisions and let us know whether any concerns remain. Thank you for your time and feedback.

Thank you for the response. However, I am not satisfied by the rebuttal.

Regarding W1 & Q1, the authors' evaluation relies heavily on PSNR and SSIM computed on synthetic datasets. However, due to the significant domain gap between synthetic and real-world data, these metrics are not reliable for assessing real-world performance. Additionally, Figure 14 ("Efficiency") does not demonstrate real-world dehazing results. I assume the authors meant Figure 15, which includes a dehazing example. Looking closely at that result, the dehazing performance appears limited. Specifically, there is a noticeable drop in resolution and visible blurring in the output image. Furthermore, the authors do not provide a fair or comprehensive comparison against other relevant dehazing baselines.

Regarding W2, Q2 & Q4: The visual evidence cited by the authors is still insufficient and unconvincing:

  • Figure 7 shows only a single image with a small zoomed-in patch, where residual rain streaks remain clearly visible.

  • Figure 15 presents only one qualitative example for dehazing and deblurring, where the visual improvement is marginal, and no comparisons against relevant dehazing/deblurring baselines are provided.

  • Figure 17 displays only quantitative plots rather than actual visual results on real-world scenes.

Furthermore, the video demo on WeatherBench shows many remaining rain streaks in "Ours," which makes the visual performance unconvincing.

Overall, my final rating is not affected by W1 and Q1, as I acknowledge that this paper focuses on video rain-streak removal. However, the visual results on real-world videos are not sufficiently promising. Therefore, I decide to maintain my initial Weak Reject rating.

Thank you for the follow-up. We respond to each point below.

  1. Figure numbering. We apologize for the confusion. Our first response used the numbers from the original submission. In the revised PDF, real-rain restorations are shown in Figures 7, 16 (NTURain), and 18 (WeatherBench); Figure 15 is the dehazing and deblurring example; and Figure 17 contains the F_phys plots.

  2. Synthetic metrics. Our real-world quantitative evidence does not rely on synthetic data. Table 17 evaluates real rainy captures with paired references (WeatherBench), using NTURain checkpoints without fine-tuning. On WB200, FluidRain achieves the best PSNR (28.37 dB) and LPIPS (0.3232) among all eleven compared methods with only 0.80M parameters.

  3. Limited real-rain examples and residual streaks (Figure 7, WeatherBench demo). Figure 7 is a summary in the main text. Figures 16 and 18 provide further real-rain examples on NTURain and on heavy and light WeatherBench scenes. Residual streaks can remain in very dense real rain; quantitatively, FluidRain still achieves the best overall restoration quality on WB200.

  4. Dehazing and deblurring (Figure 15). These experiments test whether the unchanged architecture can be adapted to other degradations, not whether it outperforms dedicated restoration methods. The dehazing example uses the synthetic haze-only frames of RainSynAll100. With short L1 fine-tuning (3,000 iterations for dehazing, and 2,000 iterations on only 282 frames for deblurring), PSNR increases by 3.60 dB for dehazing and by 0.39 dB for deblurring. The smoother appearance of the dehazed output and the smaller deblurring gain reflect this limited fine-tuning budget, and the deblurring row is shown as an enlarged crop.

Thank you again for your careful review.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.