acceptodds
Under review as a conference paper at ICLR 2027

Reusing Learned Degradation Models via Counter-Degradation Guidance for Diffusion Dehazing

Abstract

Recent diffusion-based dehazing methods increasingly rely on learned haze generators to produce realistic training data for real-world restoration. However, once the restoration model has been trained, the degradation model is typically discarded, despite retaining explicit knowledge of how haze affects the diffusion prediction. We revisit this unused knowledge and propose Haze Counter-Degradation Guidance (HazeCDG), a training-free framework that reuses a pretrained haze generator to improve an already trained diffusion restorer at inference time. HazeCDG compares the pretrained haze generator with its frozen backbone at the same diffusion state, so that their prediction difference reveals how the learned haze process locally changes the diffusion prediction. It then constructs a counter-degradation reference by reversing this learned degradation change and guides the restoration prediction toward it while preserving the relative prediction offset between the restoration and degradation models. Across real-world haze benchmarks, HazeCDG substantially reduces residual haze while preserving the perceptual strengths of the original restorer, bringing DiffDehaze into the performance range of recent dedicated dehazing methods. Importantly, these gains are obtained while keeping both the degradation and restoration models frozen; HazeCDG requires no retraining, test-time parameter adaptation, or manually tuned guidance scale. Our results demonstrate that degradation models learned for data synthesis can remain valuable after training as complementary sources of inference-time knowledge for diffusion restoration. Our code is available at https://anonymous.4open.science/r/HAZECDG-E2C1/README.md.

Primary area:
generative models
Keywords:
Image Dehazing, Diffusion Models, Image Restoration, Degradation Modeling, Generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DvhSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether a haze generator used to synthesize training data can continue to assist image restoration after training. Under the same diffusion state and conditioning, HazeCDG compares the predictions of a frozen haze generator with its degradation branch enabled and disabled, using their difference to characterize the local prediction change induced by that branch. HazeCDG applies this change to the restoration prediction by constructing a reference that reverses the change while preserving its magnitude and by preserving the restorer's prediction offset relative to the degradation model. Sparse probing and reuse across sampling steps control the computational overhead. The experiments cover real-world hazy scenes and transfer the method from its associated multistep restorer to an independently trained single-step restorer.

Critical strengths

The motivation for reusing a degradation model already available from training is clear, and the method has practical value because it requires neither retraining nor test-time parameter updates. Comparisons between neutralization and reverse guidance, analyses of probing frequency, and transfer to a single-step restorer also provide an initial assessment of its effects and applicability trade-offs.

Critical weaknesses

The justification for reverse guidance remains insufficient. The method uses the local prediction change induced by the degradation branch, but the existing neutralization-versus-reversal comparisons do not adequately establish whether the gains primarily come from learned knowledge of haze degradation or from a combination of a generic negative direction, guidance strength, and sampling schedule. No-reference evaluation and visual examples on real haze also make it difficult to directly verify local detail and color fidelity, particularly given the perceptual-quality trade-offs already observed in the single-step transfer setting.

Questions for authors
  1. With the model-call budget and sampling settings matched, the learned degradation difference should be compared with generic negative guidance to further distinguish the roles of the difference direction and the strength of the reverse guidance.
  2. Are the perceptual-quality trade-offs in single-step transfer accompanied by detail completion, structural changes, or color distortion? The authors could provide an independent content-fidelity assessment.
  3. The authors need to clarify the relationship between the real hazy images used to train the haze generator and the URHI evaluation images, and provide the corresponding split or deduplication criteria to establish the independence of the evaluation images.
  4. The experimental analysis on real-world dehazing datasets uses no-reference image-quality metrics such as FADE because paired real reference images are unavailable. However, these metrics are susceptible to visual artifacts or adversarial manipulation. Could the authors provide additional results from double-blind user studies?
1 comment
AuthorsOpenReview ↗

Response to Reviewer DvhS

Q1: Learned degradation guidance versus text-conditioned guidance

We thank the reviewer for suggesting a comparison with generic negative guidance under matched model-call budgets and sampling settings. To further examine the role of the learned haze differential, we consider an alternative guidance strategy based on text-conditioned predictions from the frozen diffusion backbone, without using the learned degradation branch. Specifically, we compute dttext=ϵhaze−ϵneutrald_t^{\mathrm{text}} = \epsilon_{\mathrm{haze}} - \epsilon_{\mathrm{neutral}} and apply ϵR−dttext\epsilon_R - d_t^{\mathrm{text}} without additional scaling or normalization. HazeCDG applies the counter-degradation update ϵR−2gt\epsilon_R - 2g_t. Both methods use two predictions to form their guidance signal at each of the same K=5K=5 sensing anchors over 50 sampling steps.

As shown in the table below, text-conditioned guidance improves over DiffDehaze on all five metrics. HazeCDG achieves stronger haze suppression and better Q-Align, CLIPIQA, and BRISQUE scores, while text-conditioned guidance achieves the best MUSIQ score. This comparison provides additional evidence for the usefulness of the learned haze differential in our counter-degradation formulation. Because the correction magnitudes are not matched, it does not isolate the effect of direction alone. Appendix H of the revised paper reports the quantitative results in Table 10 and the qualitative comparison in Figure 10.

\text{Method} & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline \text{DiffDehaze} & 1.135 & 3.592 & 0.416 & 64.609 & 24.290 \\\\ \text{Text-Conditioned Guidance} & 0.990 & 3.672 & 0.434 & \mathbf{65.467} & 23.793 \\\\ \text{HazeCDG (Ours)} & \mathbf{0.597} & \mathbf{3.691} & \mathbf{0.505} & 64.589 & \mathbf{20.758} \end{array}$$ ## Q2: Content fidelity in single-step transfer to DOD We thank the reviewer for this insightful question. We acknowledge that the single-step DOD results exhibit trade-offs among perceptual-quality metrics. To further examine their implications for content fidelity, we provide additional visual comparisons between DOD and DOD with HazeCDG, including magnified regions highlighting fine details, structural boundaries, and color appearance (Appendix D, Figure 6, p. 19 of the revised paper). These comparisons complement the quantitative results and allow closer inspection of potential structural or color distortions. We discuss the observed trade-offs in Section 4.2 of the paper. ## Q3: Possible overlap between HazeGen training data and the URHI evaluation subset We thank the reviewer for the careful reading of our paper and for highlighting this important point. Since the training data for the pretrained HazeGen model include URHI images and the exact list of training images is unavailable, we cannot rule out potential overlap with our 150-image evaluation subset. We therefore clarify that URHI serves as a supplementary in-domain evaluation rather than an independent generalization benchmark. Our main conclusions are primarily supported by the extensive evaluation on RTTS, complemented by Fattal, where HazeCDG improves haze suppression and several perceptual-quality measures without retraining or parameter updates. ## Q4: Double-blind evaluation of dehazing quality on RTTS We appreciate the reviewer’s suggestion. To complement no-reference metrics such as FADE, we conduct a double-blind rating study on 30 randomly selected RTTS images. Each of the 15 participants evaluates the outputs of all seven methods for every image. The outputs are presented in randomized order without method names. During scoring, neither the participants nor the study administrator knows which method corresponds to each output. Participants assign each output a single overall score from 1 to 5 (5 = best), considering haze removal, naturalness, and preservation of scene details. As shown in the table below, DiffDehaze + HazeCDG receives the highest mean rating among the evaluated methods. These results provide complementary human-perceptual evidence for assessing whether the improvements observed in FADE and other no-reference metrics correspond to visually preferable restoration results. $$\begin{array}{cl|cc} & \text{Method} & \text{Mean rating} & \text{Variance} \\\\ \hline (a) & \text{HNDiff} & 1.473 & 0.658 \\\\ (b) & \text{BiLaLoRA} & 1.982 & 0.960 \\\\ (c) & \text{PDDA} & 2.244 & 1.260 \\\\ (d) & \text{HazeFlow} & 1.867 & 0.947 \\\\ (e) & \text{DiffDehaze} & 3.976 & 1.024 \\\\ (f) & \text{DiffDehaze + DCTTA} & 3.551 & 1.167 \\\\ (g) & \text{DiffDehaze + HazeCDG (Ours)} & \mathbf{4.360} & 0.524 \end{array}$$
Reviewer_7PmZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Haze Counter-Degradation Guidance (HazeCDG), a training-free inference-time method for diffusion-based image dehazing. The key idea is to reuse a pretrained haze generator that was originally employed only for synthesizing training data. At each selected reverse-diffusion step, HazeCDG compares the prediction of the haze generator with and without its learned degradation branch. Their difference is treated as a local haze differential. The method reverses this differential to construct a counter-degradation direction and applies it to the frozen restoration model’s prediction, while preserving the relative offset between the restoration and degradation models. To reduce computation, the haze differential is evaluated at a small number of anchor steps and reused between them. Experiments on RTTS, URHI, and Fattal show that HazeCDG substantially reduces residual haze in DiffDehaze and improves several no-reference perceptual metrics without retraining or test-time parameter updates.

Critical strengths
  1. The paper presents a simple and practically interesting perspective: a learned degradation model may contain complementary information useful at inference time.

  2. The proposed method requires no retraining, test-time optimization, or parameter updates. Its insertion into the diffusion sampling process is straightforward.

Critical weaknesses

The update in Equation~(4) follows from an assumed equal-magnitude reflection around the frozen backbone, but the paper does not establish that this reflection approximates the clean-image score direction. Experiments with ϵR−λgt\epsilon_R-\lambda g_t are needed to demonstrate that λ=2\lambda=2 is principled rather than simply a favorable guidance strength.

Improvements in FADE or perceptual scores cannot establish faithful reconstruction and may reward aggressive contrast enhancement or over-dehazing. Reference-based evaluations using metrics such as PSNR and SSIM would substantially strengthen the conclusions. I understand that datasets such as RTTS do not provide ground-truth images; therefore, additional experiments on paired datasets, such as NH-HAZE, O-HAZE, and RW2AH, are needed.

Questions for authors
  1. How does performance vary with the general update ϵR−λgt\epsilon_R-\lambda g_t? Is λ=2\lambda=2 consistently optimal?

  2. Can the authors provide reference-based metrics (PSNR and SSIM) to evaluate dehazing performance?

  3. Can HazeCDG be evaluated on additional independently trained diffusion restorers?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7PmZ (#1)

Q1: Effect of guidance strength and the choice of α=2\alpha = 2

We appreciate the opportunity to clarify the choice of guidance strength. In our notation, the coefficient λ\lambda in the reviewer’s question is denoted by α\alpha. In HazeCDG, the factor of two is not an empirically tuned guidance strength but follows from our counter-degradation formulation. As derived in Eqs. (2)–(4) of the main text, we construct a counter-degradation reference by reversing the learned degradation displacement with equal magnitude, while preserving the relative prediction offset between the restoration and degradation models. This construction directly yields ϵR−2gt\epsilon_R - 2g_t, whereas α=1\alpha = 1 merely neutralizes the degradation displacement.

To further examine the effect of guidance strength, we evaluated the generalized update ϵR−αgt\epsilon_R - \alpha g_t on RTTS under identical sampling conditions (K=5K = 5), as shown in the table below. These results confirm that the learned haze differential provides effective guidance across different strengths, while revealing different responses among perceptual quality metrics as haze suppression becomes stronger. Although α=2\alpha = 2 is not universally optimal across all metrics, its value follows directly from the symmetric counter-degradation construction rather than empirical parameter tuning. Appendix F of the revised manuscript reports this ablation in Table 8 and shows the visual comparisons across guidance strengths in Figure 8.

\alpha & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline 0\\ \text{(Baseline)} & 1.135 & 3.592 & 0.416 & 64.609 & 24.290 \\\\ 0.5 & 0.995 & 3.639 & 0.438 & 64.996 & 23.350 \\\\ 1.0 & 0.855 & 3.672 & 0.460 & \mathbf{65.124} & 22.377 \\\\ 1.5 & 0.721 & 3.690 & 0.484 & 64.977 & 21.462 \\\\ 2.0\\ \text{(Ours)} & 0.597 & \mathbf{3.691} & 0.505 & 64.589 & 20.758 \\\\ 2.5 & \mathbf{0.491} & 3.673 & \mathbf{0.523} & 64.003 & \mathbf{20.418} \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer 7PmZ (#2)

Q2: Reference-based evaluation on paired real-haze benchmarks

We appreciate the reviewer’s suggestion to report full-reference metrics. We evaluated HazeCDG on two paired real-haze benchmarks, O-HAZE [1] and NH-HAZE [2], which contain 45 and 55 hazy–clean image pairs, respectively. We first evaluated the original HazeCDG without additional training, and then examined whether refining the degradation prior with limited paired supervision could improve fidelity. Specifically, we used 20 pairs from each dataset to fine-tune only HazeGen for 20 epochs, keeping the DiffDehaze restoration model and the HazeCDG guidance formulation unchanged. We refer to this setting as HazeCDG (Adapted). All three methods were evaluated on the same held-out images: 25 O-HAZE pairs and 35 NH-HAZE pairs.

\text{Dataset} & \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline \text{O-HAZE (25)} & \text{DiffDehaze} & 16.55 & \mathbf{0.484} & 0.390 & 1.72 & 3.49 & 0.438 & 24.21 & \mathbf{56.11} \\\\ & \text{HazeCDG (Zero-shot)} & 15.87 & 0.470 & 0.397 & 1.25 & 3.78 & \mathbf{0.472} & 24.51 & 56.64 \\\\ & \text{HazeCDG (Adapted)} & \mathbf{16.73} & 0.480 & \mathbf{0.388} & \mathbf{1.20} & \mathbf{3.87} & 0.469 & \mathbf{25.26} & 56.82 \\\\ \hline \text{NH-HAZE (35)} & \text{DiffDehaze} & \mathbf{14.27} & \mathbf{0.299} & 0.491 & 1.22 & 3.38 & 0.424 & 52.04 & 36.22 \\\\ & \text{HazeCDG (Zero-shot)} & 13.98 & 0.297 & 0.478 & \mathbf{0.85} & 3.61 & 0.498 & \mathbf{53.30} & \mathbf{33.69} \\\\ & \text{HazeCDG (Adapted)} & 14.15 & 0.297 & \mathbf{0.470} & 0.86 & \mathbf{3.65} & \mathbf{0.499} & 53.25 & 34.37 \end{array}$$ Compared with DiffDehaze, zero-shot HazeCDG reduces FADE on both benchmarks but yields lower PSNR and SSIM on both. Thus, the gains in haze suppression do not consistently translate into closer agreement with the paired ground truth. With limited adaptation of HazeGen, PSNR and LPIPS improve over zero-shot HazeCDG on both benchmarks. Since both variants use the same frozen DiffDehaze restorer and the same guidance formulation, these results suggest that refining the learned degradation prior for the target domain can improve reconstruction fidelity through the guidance it provides at inference. Our primary objective is to suppress residual haze in real-world images, for which paired clean references are generally unavailable. In its original zero-shot setting, HazeCDG reuses the degradation knowledge learned by HazeGen to guide a frozen diffusion restorer without additional training on image pairs from either benchmark. The paired evaluation therefore addresses a complementary question: whether stronger haze suppression also brings the restored image closer to its clean reference. The adapted setting separately examines whether reconstruction fidelity can be improved by refining only HazeGen when limited paired supervision is available. Appendix E of the revised manuscript provides the full experimental protocol, quantitative results in Table 7, and qualitative comparisons with the paired ground truth in Figure 7. [1] C. O. Ancuti, C. Ancuti, R. Timofte, and C. De Vleeschouwer, "O-haze: a dehazing benchmark with real hazy and haze-free outdoor images." New Trends in Image Restoration and Enhancement workshop, in IEEE Computer Vision and Pattern Recognition. 2018. [2] C. O. Ancuti, C. Ancuti, and R. Timofte, "NH-HAZE: An image dehazing benchmark with non-homogeneous hazy and haze-free images." 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2020.
AuthorsOpenReview ↗

Response to Reviewer 7PmZ (#3)

Q3: Transferability of HazeCDG to an independently trained diffusion restorer

We appreciate the reviewer's suggestion to test HazeCDG on another independently trained diffusion restorer. We replace DiffDehaze with pretrained CoDehaze [3] while retaining the same frozen HazeGen for degradation sensing. Both CoDehaze runs use the same 50-step sampling schedule. In the guided run, we apply HazeCDG with K=5K=5 sensing anchors, without changing the guidance formulation, retraining either network, or tuning additional parameters.

As shown in the table below, HazeCDG substantially improves haze suppression and several perceptual-quality measures on RTTS. These results show that the learned haze differential can improve haze suppression in a restorer trained independently of DiffDehaze. Appendix G of the revised paper reports the quantitative results in Table 9 and shows the corresponding qualitative comparison in Figure 9.

\text{Method} & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline \text{CoDehaze} & 1.139 & 3.667 & 0.519 & \mathbf{68.754} & 22.033 \\\\ \text{CoDehaze + HazeCDG} & \mathbf{0.472} & \mathbf{3.795} & \mathbf{0.602} & 68.377 & \mathbf{18.719} \end{array}$$ [3] Y.-C. Cheng, C.-H. Chang, C.-S. Fuh, and J.-C. Wu. "CoDehaze: Color-Driven Diffusion with Structured Haze Guidance for Image Dehazing." 37th British Machine Vision Conference (BMVC), 2026.
Reviewer_HhYGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HazeCDG, a training-free inference-time method that reuses the haze generator (HazeGen) from the LHD framework to improve its frozen diffusion restorer (DiffDehaze). At selected reverse steps, the method queries the haze generator with and without its IRControlNet branch. The difference between the two predictions, g_t, is treated as a “haze differential”, and the restoration prediction is corrected as \epsilon_R - 2g_t. The authors argue that the factor of two follows from a counter-degradation construction rather than a tuned guidance scale. Sparse sensing (K = 5 anchors) keeps the overhead at 1.17×. The method is evaluated on RTTS, URHI and Fattal with no-reference metrics only, and includes a feasibility study on the one-step DOD restorer.

Critical strengths

Simple, training-free and cheap. With both models frozen and only 17% extra runtime, the method consistently and substantially reduces residual haze while improving Q-Align and CLIPIQA. It is far cheaper than test-time adaptation.

The idea of reusing a degradation model after training, instead of discarding it, is a clear and interesting perspective. The neutralization vs. counter-degradation ablation supports g_t as a meaningful signal.

Critical weaknesses

The claim that there is “no guidance scale” is not convincing. ϵR−2gt\epsilon_R - 2g_t is essentially negative guidance with a fixed weight w=2w = 2, and the “preserve the relative offset” argument is a design choice rather than a derivation.

The evaluation cannot verify fidelity. All results use no-reference metrics, and FADE rewards aggressive contrast and haze removal. There is no paired evaluation (e.g., O-HAZE, or synthetic data with ground truth) to rule out over-enhancement, color shifts or hallucinated content. MUSIQ and BRISQUE do not improve consistently, and on Fattal, HazeFlow still achieves a much lower FADE.

The scope is narrow. The method depends on an unusual setup in which the degradation and restoration models share one latent-diffusion backbone (LHD). The DOD transfer shows only mixed results, and only dehazing is studied.

Questions for authors

Can you report full-reference metrics (PSNR/SSIM/LPIPS) on a paired real or synthetic haze benchmark to verify that the gains do not come from over-enhancement?

Why does performance slightly degrade as K increases, given that dense sensing should give the most accurate gtg_t?

How sensitive is the method to the haze prompt used for degradation sensing?

6 comments
AuthorsOpenReview ↗

Response to Reviewer HhYG (#1)

Q1: Full-reference evaluation on paired real-haze benchmarks

We thank the reviewer for raising this concern and suggesting a paired evaluation of reconstruction fidelity. To complement our original no-reference evaluation, we conduct additional experiments on two paired real-haze benchmarks, O-HAZE [1] and NH-HAZE [2], containing 45 and 55 hazy-clean image pairs, respectively. We first evaluate the original HazeCDG without any additional training to examine its reconstruction fidelity against ground truth. We then investigate whether the learned degradation information can further improve fidelity when limited paired supervision is available. Specifically, we use 20 image pairs from each dataset to fine-tune only HazeGen for 20 epochs, while keeping the DiffDehaze restoration model and the HazeCDG guidance formulation unchanged. We refer to this additional setting as HazeCDG (Adapted). The remaining 25 O-HAZE and 35 NH-HAZE pairs are reserved for evaluation, where DiffDehaze, the original HazeCDG, and HazeCDG (Adapted) are compared on identical held-out images. We have added the experimental protocol and results to Appendix E of the revised manuscript. The table below reproduces Table 7, including PSNR, SSIM, and LPIPS.

\text{Dataset} & \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline \text{O-HAZE (25)} & \text{DiffDehaze} & 16.55 & \mathbf{0.484} & 0.390 & 1.72 & 3.49 & 0.438 & 24.21 & \mathbf{56.11} \\\\ & \text{HazeCDG (Zero-shot)} & 15.87 & 0.470 & 0.397 & 1.25 & 3.78 & \mathbf{0.472} & 24.51 & 56.64 \\\\ & \text{HazeCDG (Adapted)} & \mathbf{16.73} & 0.480 & \mathbf{0.388} & \mathbf{1.20} & \mathbf{3.87} & 0.469 & \mathbf{25.26} & 56.82 \\\\ \hline \text{NH-HAZE (35)} & \text{DiffDehaze} & \mathbf{14.27} & \mathbf{0.299} & 0.491 & 1.22 & 3.38 & 0.424 & 52.04 & 36.22 \\\\ & \text{HazeCDG (Zero-shot)} & 13.98 & 0.297 & 0.478 & \mathbf{0.85} & 3.61 & 0.498 & \mathbf{53.30} & \mathbf{33.69} \\\\ & \text{HazeCDG (Adapted)} & 14.15 & 0.297 & \mathbf{0.470} & 0.86 & \mathbf{3.65} & \mathbf{0.499} & 53.25 & 34.37 \end{array}$$ Compared with DiffDehaze, zero-shot HazeCDG reduces FADE on both benchmarks, but the lower FADE scores do not consistently correspond to closer agreement with the paired ground truth. All three full-reference metrics worsen on O-HAZE, whereas on NH-HAZE, LPIPS improves while PSNR and SSIM decline. The paired evaluation therefore does not rule out over-enhancement in individual images. Figure 7 of Appendix E presents the restored images alongside their clean references for direct comparison of visibility, color, and scene structure. Our primary objective is to suppress residual haze in real-world images, where paired clean references are generally unavailable. The paired evaluation complements this objective by testing whether improved visibility also corresponds to faithful reconstruction. In the separate adapted setting, refining only HazeGen improves PSNR and LPIPS over zero-shot HazeCDG on both benchmarks, showing that the degradation prior can benefit from limited paired supervision. [1] C. O. Ancuti, C. Ancuti, R. Timofte, and C. De Vleeschouwer, "O-haze: a dehazing benchmark with real hazy and haze-free outdoor images." New Trends in Image Restoration and Enhancement workshop, in IEEE Computer Vision and Pattern Recognition. 2018. [2] C. O. Ancuti, C. Ancuti, and R. Timofte, "NH-HAZE: An image dehazing benchmark with non-homogeneous hazy and haze-free images." 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2020.
AuthorsOpenReview ↗

Response to Reviewer HhYG (#2, #3)

Q2: Effect of increasing the sensing budget KK on restoration performance

We thank the reviewer for this insightful question. In HazeCDG, KK determines how frequently the haze differential is recomputed, rather than how often guidance is applied. Even with K=5K = 5, the most recently sensed differential guides all 50 reverse steps. Although denser sensing provides an updated degradation response at each step, gtg_t represents the local change induced by the learned degradation model, not an optimal direction toward the clean image. Furthermore, each guidance update influences the subsequent restoration trajectory, so more frequent sensing does not necessarily lead to better reconstruction. In contrast, sparse sensing maintains the sensed degradation direction over successive steps, allowing the restoration process to follow a more consistent counter-degradation guidance rather than continually responding to local variations. Thus, the effectiveness of degradation guidance depends not simply on how frequently gtg_t is updated, but on how its information influences the overall restoration trajectory. Furthermore, each guidance update influences the subsequent restoration trajectory, so more frequent sensing does not necessarily improve reconstruction, as the RTTS results in Table 1(a) of the main paper.

Q3: Sensitivity to the haze prompt used for degradation sensing

We appreciate the reviewer’s thoughtful question about how the choice of haze prompt affects HazeCDG. Our results indicate that HazeCDG is robust to the tested haze prompt variations in its overall dehazing effectiveness, although the magnitude of improvement varies with the prompt. To examine this sensitivity, we evaluate shorter and extended versions of the original degradation prompt under identical sampling conditions (K=5K=5, 50 steps).

As shown in the table below, all three prompt configurations improve haze suppression over DiffDehaze. These findings suggest that HazeCDG does not depend critically on a particular wording of the haze prompt, supporting the effectiveness of the learned degradation differential across different haze-related text conditions.

\text{Method} & \text{FADE}\downarrow & \text{Q-Align}\uparrow & \text{CLIPIQA}\uparrow & \text{MUSIQ}\uparrow & \text{BRISQUE}\downarrow \\\\ \hline \text{DiffDehaze} & 1.135 & 3.592 & 0.416 & 64.609 & 24.290 \\\\ \hline \text{Short} & 0.783 & 3.684 & 0.495 & 64.600 & 23.237 \\\\ \text{Original} & \mathbf{0.597} & 3.691 & \mathbf{0.505} & 64.589 & \mathbf{20.758} \\\\ \text{Extended} & 0.643 & \mathbf{3.692} & 0.501 & \mathbf{64.640} & 21.172 \end{array}$$ Degradation prompts: - Short: hazy, foggy - Original: hazy, foggy, misty, obscure, smoggy. - Extended: hazy, foggy, misty, obscure, smoggy, atmospheric haze, dense fog, reduced visibility.

Actually, some previous research has explored the haze difference or related techniques into image dehazing like Depth information assisted collaborative mutual promotion network for single image dehazing. So I do not think the haze differential used in this paper is novel.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.