- 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
- 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.
- 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.
- 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.
- 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?
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−ϵneutral and apply ϵR−dttext without additional scaling or normalization. HazeCDG applies the counter-degradation update ϵR−2gt. Both methods use two predictions to form their guidance signal at each of the same K=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}$$