acceptodds
Withdrawn from ICLR 2027

HiHaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world haze is often spatially non-uniform, while the high resolution of modern images imposes substantial computational and memory costs on diffusion-based dehazing. Although one-step latent diffusion provides an efficient alternative to iterative sampling, two challenges remain. First, latent compression and spatially coarse conditioning can weaken local degradation variations, making it difficult to adapt restoration strength to different haze densities. Second, high-resolution inference commonly relies on independently processed patches, which can introduce visible seams and inconsistent color or dehazing strength across the image. To address these issues, we propose **HiHaze**, a haze-aware one-step diffusion framework for high-resolution real-world image dehazing. At the model level, we introduce **Haze-Aware Feature Modulation (HFM)**, consisting of two complementary designs. **Attentive Skip Compensation (ASC)** selectively retrieves structural information from VAE encoder features according to decoder queries, compensating for compression loss while suppressing the direct propagation of haze degradation. **Spatial Degradation Adapter (SDA)** encodes multiple degradation cues into multi-scale spatial features and injects them into the denoising network, enabling region-adaptive restoration for non-uniform haze. For high-resolution inference, we further propose **Global-Guided Patch Fusion (GPF)**, a training-free strategy that combines globally coherent low-frequency information from an additional downsampled full-image prediction with high-frequency details recovered from local patches, while adaptively preserving reliable textures according to local structural consistency. Extensive experiments on multiple real-world dehazing benchmarks demonstrate that HiHaze achieves strong restoration quality and perceptual fidelity while retaining the efficiency of one-step diffusion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Dehazing, Diffusion, Fusion

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

This paper presents Hihaze, a one-step latent-diffusion system for high-resolution real-world dehazing. It uses attentive skip compensation (ASC) to transfer reliable encoder features, a spatial degradation adapter (SDA) to inject multi-scale haze features into the U-Net, and training-free global-guided patch fusion (GPF) to align patch predictions. The method reports improvements in both quality and speed across four real-world benchmarks.

Critical strengths

One-step denoising significantly accelerates the diffusion model.

The dehazed results are clean and thorough, and the method demonstrates good generalization.

Critical weaknesses

The paper does not provide reference code. Figures 2 and 4 are well designed, but they do not clearly explain how GPF relates to the overall Hihaze framework.

The test sets overlap substantially. Many test images largely duplicate those in O-HAZE, so evaluating on multiple datasets provides limited additional value.

The authors repeatedly reintroduce abbreviations, such as “(SDA),” after they have already been defined. This suggests that the AI-generated text was not carefully checked.

The authors do not report runtime comparisons with non-diffusion-based models. More importantly, the related-work discussion appears to deliberately omit prior high-resolution dehazing studies, such as 4KDehazing and DehazeXL. The paper should at least discuss the limitations of these existing approaches to better justify its novelty. The implementation of GPF also appears somewhat similar to those of 4KDehazing and FreeScale.

Questions for authors

The paper is reasonably complete, but it still requires substantial revision. The AI-generated language is insufficiently polished, and the apparent omission of relevant prior work weakens the claimed novelty. Based on the technical content alone, I would lean toward a weak accept after major revision. However, the lack of reference code makes the work less reproducible than other submissions that provide it. Therefore, my final decision is Weak Reject.

Reviewer_eg5tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hihaze, a one-step diffusion framework for high-resolution real-world image dehazing. Hihaze introduces two main architectural components: an attentive skip compensation (ASC) module for recovering useful encoder information and a spatial degradation adapter (SDA) for modeling spatially varying haze degradation and the corresponding restoration requirements. For high-resolution inference, this paper further presents a training-free global-guided patch fusion (GPF) strategy. GPF uses a global prediction to align the appearance of local patch predictions and reduce the influence of unreliable regions, thereby mitigating boundary artifacts and illumination inconsistencies introduced by patch-wise processing. Experiments on four real-world dehazing benchmarks show that Hihaze achieves strong performance relative to the compared methods.

Critical strengths

S1. Spatially adaptive degradation modeling: The proposed supervision strategy for the SDA prior encoder is intuitive and well motivated. It encourages the encoder to represent spatially varying degradation and restoration requirements rather than treating haze as spatially uniform. The visualizations in Figure 3 provide useful qualitative evidence that the learned priors vary spatially in a manner consistent with regional haze density.

S2. Practical high-resolution inference: GPF provides an intuitive and practical solution to the inconsistencies caused by patch-wise processing of high-resolution images. In particular, using a global prediction to harmonize local predictions is well aligned with the observed problems of boundary artifacts and illumination mismatch. The experiments further suggest that GPF is reasonably general, as it improves performance when applied to multiple restoration backbones.

Critical weaknesses

W1. Insufficient justification and validation of the SDA supervision: Although the intended roles of TdT_d, TrT_r, and TsT_s are intuitive, the paper does not sufficiently explain why these particular formulations were selected. The construction of Φ(⋅)\Phi(\cdot) is also unclear, especially how the Sobel-gradient and contrast statistics are computed and combined. More importantly, the ablation study does not isolate the effect of the proposed supervision targets, for example, by comparing SDA with and without explicit prior supervision or by evaluating the contribution of each target separately. Consequently, it is difficult to determine whether the gains are obtained from the proposed supervision design or primarily from the additional adapter capacity.

W2. Incomplete specification and limited validation of detail compensation in GPF: The computation of the structural agreement map ss is not specified clearly, and the self-guided filtering used to obtain HH and HfH^f lacks an explicit formulation, sufficient implementation details, or a supporting reference. Because these quantities directly determine the final detail compensation, the method is difficult to reproduce precisely. Moreover, the relatively modest improvement from detail compensation in the ablation study weakens the empirical justification for this part of the design.

Questions for authors

Q1: How do the formulations of TdT_d, TrT_r, and TsT_s encode their intended degradation and restoration properties? In particular, how are the Sobel-gradient and contrast statistics computed and combined in Φ(⋅)\Phi(\cdot)? A more detailed ablation, i.e., comparing SDA without target supervision, SDA with all targets, and SDA with each target removed individually, would help establish the contribution of the proposed supervision strategy.

Q2: Could the authors provide an explicit formulation for the structural agreement map ss in GPF? Please also clarify the self-guided filtering procedure used to obtain HH and HfH^f, including its parameters or an appropriate reference. These details are important for reproducing the detail-compensation stage.

Q3: More consistent notation between the figures and the main text would improve readability. For example, the fused prediction is denoted by FF in Figure 4 but appears to correspond to IfI^f in the text. Figure 2 may label the features associated with ASC and SDA more explicitly, for example, ElE_l and DlD_l, and use consistent color coding for the modules throughout the paper.

Q4: Hihaze is trained using LMHaze and RW2AH and then evaluated on I-HAZE, O-HAZE, Dense-Haze, and NH-HAZE, constituting a cross-dataset generalization setting. Some prior methods, however, report results obtained under dataset-specific training protocols. Could the authors clarify the training data and evaluation protocol used for every baseline in the comparison tables? If the reported baseline numbers come from different protocols, this should be stated explicitly. Comparisons under a unified training protocol would make the results easier to interpret. The cross-dataset setting may itself be more challenging, but the distinction should be made clear.

Q5. Could the authors include comparisons with stronger recent non-diffusion dehazing methods, such as ConvIR (TPAMI 2024) and SGDN (AAAI 2025)? Because SGDN introduced the RW2AH dataset, and Hihaze uses RW2AH as part of its training data, reporting results on a held-out RW2AH test split would also provide a useful in-domain evaluation.

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

This paper proposes HiHaze, a one-step diffusion framework for high-resolution real-world image dehazing. It introduces attentive skip compensation (ASC) and a spatial degradation adapter (SDA) to improve detail restoration and handle spatially varying haze. In addition, a training-free global-guided patch fusion (GPF) module is proposed to improve consistency across patches.

Critical strengths
  1. The paper is well-motivated and addresses practical challenges in high-resolution image dehazing.
  2. The proposed method achieves strong quantitative and qualitative results on four real-world dehazing benchmarks.
  3. GPF is training-free and can be applied to other dehazing methods without retraining, demonstrating good transferability.
Critical weaknesses
  1. Limited novelty of ASC and SDA. Both modules appear to be task-specific adaptations of existing techniques rather than fundamentally new designs. Similar ideas, such as selective skip connections and degradation-aware feature modulation, have been explored in prior work. The authors should better clarify how ASC and SDA differ from existing methods and provide stronger comparisons to demonstrate their advantages.
  2. Formatting issues in Figure2. The text inside the green boxes of the SDA and ASC modules is not rendered correctly, making the figure difficult to read. The authors should fix these issues.
  3. Lack of comparison with other high-resolution dehazing methods. Since GPF is a key component for high-resolution dehazing, it would be valuable to compare HiHaze with DehazeXL (CVPR 2025), which also focuses on high-resolution dehazing and global-local information fusion. Such a comparison would help better demonstrate the effectiveness of GPF.
Questions for authors
  1. Since SDA is designed to handle spatially non-uniform haze, why are the ablation studies conducted only on O-HAZE? Additional ablations on NH-HAZE would help demonstrate its effectiveness.
  2. SDA uses three handcrafted supervision targets. What is the individual contribution of each target? An ablation of these targets would help justify the design.

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

Related papers

Discussion (0)

Sign in to comment.