acceptodds
Withdrawn from ICLR 2027

Coupling-Aware Compositional Restoration: Zero-Shot, Training-Free Removal of Composite Adverse Weather

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world adverse weather rarely appears in isolation: low light, haze, rain, and snow frequently co-occur and interact, producing composite degradations that are far harder to remove than any single effect. State-of-the-art methods for composite weather are fully supervised and must be trained on a combinatorially large set of degradation combinations, while existing training-free restorers built on diffusion priors are designed for a single, known degradation and break down once degradations are stacked. We present CACR, a zero-shot, training-free framework for composite adverse-weather restoration. CACR detects which base weathers are present, derives a physically consistent removal order, and applies image-space physical operators in a coupling-aware manner, re-estimating each operator’s parameters from the partially restored image rather than from the raw input. Restoration is performed entirely by zero-shot classical operators: no restoration network is trained, and no paired restoration data is used. Weather detection itself runs zero-shot from physical cues, with an optional lightweight few-shot probe that only sharpens detection, never the restoration. We show that single-degradation zero-shot restorers can fall below the degraded input on stacked weather, that the coupling, not the individual operators, drives our improvement, and that diffusion-inversion editing is bounded by a reconstruction ceiling that motivates restoring in image space. On the CDD-11 composite-weather benchmark, CACR provides the first zero-shot, training-free reference point across all eleven classes, reported in PSNR, SSIM, and LPIPS, with a clear characterization of the remaining gap to supervised state of the art.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Composite adverse weather restoration, Zero-shot image restoration, Training-free image restoration, Coupling-aware restoration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_4jwzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CACR for restoring combinations of low light, haze, rain, and snow without training a restoration network. It detects degradations using physical cues and an optional few-shot classifier on frozen DINOv2 features, then applies illumination correction, dehazing, and particle removal in a fixed order. Its central idea is to account for degradation interactions by re-estimating parameters from intermediate images and selecting operators according to the detected weather. Evaluation includes comparisons with classical and diffusion-based training-free baselines on CDD-11, component ablations, and tests on independent single-weather datasets.

Critical strengths

Not provided.

Critical weaknesses

1.The writing quality and experimental validation of this manuscript are far below the standards for submission to ICLR. Specifically, regarding experimental validation, the visualization section of the experiments not only lacks comparisons with other methods, but the provided examples also fail to demonstrate any clear advantages of the proposed method. In terms of writing quality, the author frequently and indiscriminately used italics and other special fonts in the manuscript, which seriously affected the reading experience. The aforementioned issues not only affect the quality of presentation but also impact readers' understanding and judgment of the experimental results, leading me to believe that the authors lacked sufficient thoroughness in their checks before submitting the manuscript.

2.The design of the core method lacks sufficient justification, and its substantive contributions remain unclear. The authors provide a qualitative explanation of the fixed degeneration process order, but do not fully explain how this order is derived from the mentioned formula. While existing ablation experiments can demonstrate that coupling brings additional benefits, they do not explain the contribution of each mechanism separately, and therefore are insufficient to support the conclusion that "the main benefits come from coupling rather than operators". Therefore, it is currently difficult to determine the substantial advantages of this method compared to a well-designed series of traditional operators.

Questions for authors

Besides the above weaknesses, the following lists some other concerns:

1.Section 3.3 states that the removal order “low → haze → particle” follows from the physics of Eq. (1), but the manuscript does not establish this connection. Please provide the derivation or evidence supporting this specific order.

2.Since composite weather restoration is evaluated only on CDD-11, what evidence supports generalization to composite weather beyond this benchmark?

3.Why does Figure 2 omit visual results from all competing methods?

4.The PSNR in Figure 1 is labeled as 17.41 → 11.38 → 20.15 → 20.07, but the caption claims that it is monotonically increasing. In Table 3, the PSNR of rain and snow also decreased from 23.1 to 21.8 and 23.9 to 20.9 respectively, why is it still referred to as "do-no-harm" in the manuscript? Please reconcile these inconsistencies and clarify which conclusions remain supported.

Reviewer_hYo1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Coupling-Aware Compositional Restoration (CACR), a zero-shot training free composite degradation restoration approach. The authors address composite degradations by leveraging their underlying physical models, and capturing how different degradations interact with each other in a composite setting (known as "coupling"). Specifically, the three couplings the authors identify are (1) addressing airlight of haze after addressing illumination, (2) assessing haze presence after illumination is corrected and after particles are removed (rain/snow), and (3) addressing haze using depth-guidance only when lowlight is absent. Thus, the degradations are removed using a predefined order, i.e. low-light, haze and rain/snow removal. The degradations in the image are detected using a Dino-based degradation classifier. The approach achieves state-of-the-art performance among training-free approaches on mixed degradation sets.

Critical strengths
  1. It is interesting to see that a physics-based approach is able to tackle composite degradations to a certain extent, without the use of any pre-trained model-based priors (like many existing training-free approaches).

  2. The problem being tackled, i.e. addressing composite degradations using a training-free approach, is interesting.

Critical weaknesses
  1. The performance gain from the coupling, which is the core component of the work, is not significant. For instance, on average (Table 2) the coupling only adds 0.35 dB of performance. Given these results, the contribution of the work does not seem novel. The authors propose a training-free approach for composite degradation handling, which I agree is interesting on an idea level. However, the core contribution is the coupling which still largely uses existing physical models and lacks a strong quantitative justification.

  2. The effectiveness of de-particle stage (rain/snow removal) is not clear. For instance in Fig. 2, there is no difference after departicle. In fact in many instances, the PSNR seems to drop. Furthermore, the caption of Fig. 2 mentions a monotonic increase in PSNR which is not true.

  3. The qualitative results are very hard to follow. Specifically, Fig. 2, which contains the main qualitative results of the paper, needs a complete overhaul as it is very difficult to read the PSNR values, view the small image tiles, and read the captions for each row or column.

Questions for authors
  1. Can the authors provide more evidence for where the coupling actually improves restoration performance? The quantitative results in Table 2 are not supporting its contribution.

  2. De-particle does not seem to remove the particles effectively. Is there any specific reason for this?

  3. Can the authors provide some results on real-world composite degradations? Currently the real-world results are only provided for single degradations (Table 4)

Reviewer_YArpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies composite adverse-weather restoration without training a restoration network. CACR selects low-light, haze, rain, and snow operators using physical cues, optionally combined with a small probe on frozen DINOv2 features. It applies illumination correction, dehazing, and masked particle removal in a fixed coarse-to-fine order. The proposed couplings include estimating airlight after brightening, assessing haze after preliminary particle removal and illumination correction, and selecting depth-based or dark-channel transmission according to the detected weather. Experiments at 256×256 compare CACR with individual classical operators and three diffusion-based configurations on CDD-11. The paper reports improvements over the input on all seven composite classes, detector and operator ablations, and additional evaluations on single-weather benchmark images.

Critical strengths
  1. The operating regime is worthwhile. Composite weather is a practical setting in which paired training data can be expensive, and a method that uses existing encoders and physical operators offers a useful reference for what can be achieved without training a restorer. Table 1 reports a composite PSNR improvement from 15.42 to 17.47 dB, with SSIM increasing from 0.685 to 0.702 and LPIPS decreasing from 0.331 to 0.304. Table 3 shows positive mean PSNR changes on each of the seven composite classes. I also appreciate the transparency about the strongest detector using 30 training images per class: the label-free variant is reported separately at approximately 17.1 dB, and the limitations acknowledge the distinction. A training-free reference can be valuable even when its absolute accuracy remains below that of supervised models.

  2. The paper includes useful controls and qualifications. Table 2 holds the operator suite and ground-truth weather labels fixed when comparing coupling settings, which is a better attribution experiment than comparing only with isolated DCP or gamma correction. It also compares detector choices and targeted versus global particle removal. The independent single-weather evaluations provide some evidence that the operators are useful beyond CDD-11, particularly for low light. Finally, the limitations explicitly acknowledge the synthetic, low-resolution composite setting, single-weather scope of the external evaluation, and losses on relatively clean rain and snow inputs. These disclosures make it possible to assess the contribution within a concrete scope.

Critical weaknesses
  1. The controls support an incremental benefit from coupling, but do not support the stronger mechanism claims. With ground-truth detection and identical operators, Table 2 gives 17.66 dB with coupling off and 18.01 dB with coupling on. The former already improves substantially over the 15.42 dB input; the latter adds about 0.35 dB. The larger 0.83 dB gain on low+haze is encouraging, but the evidence does not establish that coupling, rather than the operator cascade, drives the overall improvement. Sequentially invoking the dehazing operator defined in Section 3.4 after gamma correction would already estimate airlight from the brightened input, so the distinction from that ordinary cascade needs clarification. Both compared variants use a fixed order, and oracle labels bypass the second coupling's haze-selection decision. Ordering, airlight estimation, detector preprocessing, and depth gating therefore need to be distinguished.

  2. The claimed behavior of the stages is also overstated. Figure 1 displays 17.41→11.38→20.15→20.07 dB, contradicting its monotonic-improvement caption. On single rain and snow, Table 3 gives 23.1→21.8 and 23.9→20.9 dB, so restricting edits to a mask does not establish do-no-harm. The final limitations correctly acknowledge these losses; that qualification should also govern the earlier interpretation. None of these observations rules out a useful final restoration result. They do mean that the paper has not established the stronger picture of a physically determined sequence whose stages consistently improve fidelity. Noncommutativity alone also does not determine an inverse order without specifying the forward composition being inverted.

  3. The diffusion comparison and evaluation recipe are not sufficiently specified to support the claimed structural failure of existing restorers. Section 4.1 explicitly says that the BIRD row is a prior-only round-trip because no blind weather operator is used. This is a reconstruction diagnostic, rather than BIRD's restoration procedure, which jointly optimizes the image and degradation parameters. The other two methods also deserve a more careful characterization: Invert2Restore supports fully blind restoration, and LD-RPS learns a feature/pixel alignment mapping at inference. Their poor scores cannot simply be explained by assuming that all these methods require a known forward degradation operator. The supplied paper does not give the priors, inversion/sampling budgets, optimization settings, or implementation checks needed to interpret their composite-weather results. I could not locate the round-trip experiment or assumptions behind the repeatedly claimed reconstruction ceiling.

  4. The CACR and dataset descriptions leave related verification gaps. The reproducibility statement supplies gamma, omega, the transmission floor, and the streak threshold, but omits beta, which controls the depth-based transmission. The detection thresholds and the preliminary haze-assessment procedure are also unspecified. Table 4 calls every row a full test split, yet Snow100K lists 500 images while its official test set contains 50,000, and NH-HAZE lists all 55 dataset pairs rather than the five-image challenge test split. Evaluating a fixed pipeline on a larger pool is not itself evidence of training leakage, but the actual protocol needs to be named. There is also a numerical inconsistency: the eleven LD-RPS entries in Table 3 average to 15.80 dB, outside the rounding range compatible with the 16.01 dB in Table 1. Together, these issues make it difficult to distinguish a reproducible baseline finding from implementation and reporting differences.

Questions for authors
  1. What exactly changes between the coupling-off and coupling-on runs, and how do they differ from a standard cascade that estimates each operator's parameters from its current input? Please specify the order, airlight estimation, depth gate, and treatment of the second coupling. What evidence isolates the haze-selection coupling, which ground-truth detection bypasses? Please also clarify the forward composition for which the inverse order is intended.

  2. Which pretrained priors and inversion, sampling, optimization, and alignment settings were used for BIRD, Invert2Restore, and LD-RPS? Do the reimplementations recover a reasonable reference result on a native single-degradation task? Please separate the BIRD round-trip diagnostic from restoration, and provide the analysis underlying the claimed reconstruction ceiling, including what quantity is bounded.

  3. Could you provide the complete CACR evaluation recipe, particularly beta, depth normalization, detector thresholds, filtering kernels, and the preliminary transformations used by DETECT? Please identify the CDD-11 subset and the exact Snow100K and NH-HAZE evaluation pools, and reconcile the LD-RPS aggregate with its per-class results. If the all-composite improvement claim also covers label-free detection, please provide its per-class results alongside the probe-based results.

Reviewer_GRkqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

CACR focus on composite adverse-weather restoration by detecting degradation types and sequentially applying gamma correction, dehazing, and particle filtering. Handcrafted coupling rules adjust detection and operator parameters. Evaluation covers CDD-11 and several single-weather datasets.

Critical strengths

The problem is practically relevant, and the pipeline is interpretable and requires no restoration-network training.

Critical weaknesses
  1. Signigicantly limited novelty. The method largely assembles established operators and pretrained models through handcrafted rules, which is not intuitive in the deep-learning era. Non-commutativity does not justify the prescribed low-light → haze → particle removal order, and alternative orders are not systematically evaluated. The central coupling mechanism contributes only 0.35 dB in Table 2, insufficient to substantiate the claim that coupling drives the improvement.

  2. Very poor performance and insufficient validation. Composite PSNR reaches only 17.47 dB, versus 15.42 dB for the input. More concerningly, rain and snow performance deteriorates by 1.3 and 3.0 dB, directly contradicting the repeated "do-no-harm" claim. The visulized results in Fig.2(c) also demonstrates under-quality restoration performance.

  3. Unconvincing comparisons. BIRD is reduced to a "prior-only round-trip," which does not constitute a faithful evaluation of its restoration capability. Implementation details for the other diffusion baselines are insufficient. Individual DCP and gamma operators are weak controls for a composite restoration pipeline, while systematic comparisons with tuned cascades and alternative orders are absent. External benchmarks compare only against degraded inputs. These experiments cannot support even basic conclusion about the failure of existing zero-shot restorers.

Questions for authors
  1. What justifies the fixed removal order beyond intuition? How does CACR compare with alternative orders and a tuned classical cascade?
  2. Why is BIRD evaluated only through a prior-only round-trip? Can the authors provide faithful baseline implementations and complete configurations?
  3. How can the "do-no-harm" claim be reconciled with the 1.3/3.0 dB losses on rain/snow, and the detection-bottleneck claim with the limited oracle-detection performance?

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

Related papers

Discussion (0)

Sign in to comment.