Pith. sign in

REVIEW 5 major objections 8 minor 63 references

The Devil is in the Darkness: Diffusion-Based Nighttime Dehazing Anchored in Brightness Perception

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A brightness-aware diffusion model dehazes night scenes and restores daytime lighting in one pass.

desk verdict Genuinely novel synthesis pipeline, but the paper's core brightness-mapping claim rests on no-reference IQA and unvalidated synthetic constants; deserves peer review with major-revision expectations. read the letter →

arxiv 2506.02395 v1 pith:W4PQ6QQS submitted 2025-06-03 cs.CV

Xiaofeng Cong , Yu-Xin Zhang , Haoran Wei , Yeying Jin , Junming Hou , Jie Gui , Jing Zhang , Dacheng Tao This is my paper
classification cs.CV
keywords nighttimedehazingbrightnessmappingdiffusionmodelimagerestorationdatasynthesisperceptionnon-uniformillumination
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tackles the nighttime dehazing and brightness mapping (NDBM) task: taking a hazy night image and producing a clear image with daytime illumination. It argues that existing datasets and models fail at this task because their brightness mapping is globally uniform and does not encode real-world night-to-day brightness relationships. To fix this, DiffND builds a data synthesis pipeline that uses estimated depth and a sky mask to apply non-uniform brightness adjustment and simulated light distortions, then trains a pre-trained diffusion model with a brightness perception network that injects brightness information into the decoding process. The paper's cross-validation claims that the new data and the new model must be used together to get visually satisfactory real-world results.

What carries the argument

The paper's central mechanism is a two-part pipeline. First, data synthesis converts real daytime images into hazy night images: depth estimation and sky segmentation produce a non-uniform illumination mask $\hat{S}$, sky and distant regions are darkened (Eqs. 2--4), and simulated point/cone light sources with an attenuation map are added through Eqs. 5--6 inside the severe degradation model. Second, lighting reconstruction fine-tunes a pre-trained one-step diffusion model with LoRA, while a brightness perception network $\Phi$ predicts a brightness map $\hat{M}$ from the encoder's first-layer features; this map modulates the first skip connection (Eq. 9) and is trained with the brightness-mean loss $L_M$, together with pixel-wise, adversarial, and diffusion losses.

What would settle it

Take real scenes photographed at night and again in daytime with fixed camera pose and exposure (or a calibrated reference), train DiffND on the synthetic pipeline alone, and compare region-wise brightness of the output with the real daytime image; if the mean absolute brightness error is no better than a global gamma correction, the claimed non-uniform mapping is not doing the work.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the bottleneck in nighttime dehazing with brightness mapping is not network capacity but the combination of training data and lighting prior. It shows that training the proposed diffusion model on existing daytime-label datasets (UNREAL-NH, NHR/NHM/NHC-Daytime) still produces dark or over-bright results, and training existing dehazing networks on the new synthetic dataset damages content; only DiffND, with both the depth/sky-guided non-uniform brightness synthesis and the brightness perception network, yields outputs whose brightness is close to real daytime images on the RWNH test set. The central discovery is that non-uniform, depth-aware synthesis plus a diffusion prior is sufficient to learn night-to-day brightness mapping without aligned real night/day pairs.

Load-bearing premise

The load-bearing premise is that the hand-set non-uniform brightness mapping (Eqs. 2--4) and simulated light distortions (Eqs. 5--6) reproduce the brightness and haze statistics of real night scenes closely enough that a model trained only on synthetic pairs from Places365 generalizes to real RWNH images, a transfer the paper assumes rather than validates with aligned real night/day pairs.

Editorial extensions

If this is right

  • A model trained only on synthetic night/day pairs derived from Places365 generalizes to real-world RWNH images without paired real data.
  • Existing NDBM datasets (UNREAL-NH and the NH-series) are insufficient for the task because their global brightness adjustments do not match real night scenes.
  • The same fine-tuned model can produce dehazed nighttime results by switching the text prompt from 'a clear daytime image' to 'a dark nighttime image' while preserving scene content.
  • Both the depth/sky-guided non-uniform brightness adjustment and the brightness perception network contribute to the reported quality; removing either lowers visual or quantitative performance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the depth/sky-guided synthesis recipe could transfer to other day-to-night translation problems, such as low-light enhancement or nighttime detection, wherever aligned pairs are hard to obtain.
  • We infer that collecting a small set of aligned real night/day images would let a future study test whether the learned brightness mapping is physically consistent, rather than only visually plausible.
  • We infer that the fixed synthesis hyperparameters (for example $\alpha=4$, $\varphi_1=2$, $\varphi_2=1.5$) may need re-tuning for scenes without blue sky, since the training data were selected for blue-sky daytime images.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. This paper proposes DiffND, a framework for the nighttime-dehazing-and-brightness-mapping task (NDBM), i.e., converting hazy nighttime images into clear daytime-equivalent images. The framework has two parts. First, a data synthesis pipeline (Sec. 3.1) starts from real daytime images from Places365 and produces synthetic hazy night images via a depth- and sky-guided non-uniform brightness mapping (Eqs. 2-4) plus simulated point/cone light sources with an attenuation model (Eqs. 5-6). Second, a restoration model (Sec. 3.2) fine-tunes a one-step pre-trained diffusion model with LoRA, conditioned on a fixed 'a clear daytime image' prompt, and adds a brightness perception network Phi whose predicted brightness map modulates the first-level skip connection; Phi is supervised toward the mean brightness of the daytime label (Eqs. 8-10). Training uses 1,546 synthetic image pairs; the model is evaluated on 440 real images from the RWNH set using ten no-reference IQA metrics, cross-validations (training their model on existing datasets and existing models on their dataset), and ablations. The paper claims that DiffND achieves superior haze removal and night-to-day brightness mapping and that both the dataset and the model are necessary for the favorable results.

Significance. The NDBM task is genuinely under-addressed, and the paper's two-part recipe -- depth/sky-guided non-uniform brightness synthesis and a one-step diffusion model with a brightness prior -- is a plausible and timely direction. The paper has real strengths: it evaluates on a real-world test set (RWNH) instead of synthetic benchmarks only; it runs cross-validation in both directions (their model on existing datasets, existing models on their dataset, Tables 2-3); and most of the pipeline is specified concretely enough to implement. The main weaknesses concern the evidence for the central claim: the synthetic-to-real transfer of the hand-set brightness mapping (Eqs. 2-6) is never validated against real night/day statistics; the quantitative case rests entirely on no-reference quality metrics with single runs and no error bars, significance tests, or human study; and the ablation for the headline brightness perception component Phi is reported in an uninterpretable aggregated form. These issues are fixable without changing the method's scope, but they currently prevent me from accepting the paper's performance claims at face value.

major comments (5)
  1. [Sec. 3.1 (Eqs. 2-6); Sec. 4.1-4.2] The central claim -- faithful night-to-day brightness mapping on real images -- rests on an unvalidated synthetic-to-real transfer. The synthesis constants (alpha=4, phi1=2, phi2=1.5, rho, theta=0.98, mu=0.85, xi1=1, xi2=3, xi3=1.8) are set by hand in Sec. 4.1 with no sensitivity analysis, and the abstract's assertion that the pipeline ensures 'brightness consistency between synthetic and real-world scenes' is never verified directly: there are no aligned real night/day pairs, no distributional comparison of synthesized night images against RWNH or other real night statistics, and the only real-world evidence is no-reference IQA (Table 1) plus mean-brightness plots of the final outputs (Fig. 7), which cannot detect systematic mapping errors such as a consistent global offset or a wrong sky/non-sky brightness ratio. I recommend adding (i) a per-region brightness comparison (mean, standard deviation, sky/non-sky ratio) between synthetic night images and real night images, (ii) a sensitivity analysis over the synthesis constants, and (iii) if feasible, a test on real aligned day/night pairs.
  2. [Sec. 4.2 (Table 1); Sec. 4.4] The quantitative evidence for superiority consists of ten no-reference IQA metrics on a single evaluation of 440 real RWNH images: no error bars, no significance tests, and no human evaluation are reported. Several of these metrics (MUSIQ, MANIQA, DBCNN, TOPIQ) are generic naturalness/quality scores uncalibrated for the NDBM task, so high scores do not establish that the output lighting matches the true daytime appearance of the scene. In addition, no simple baseline (e.g., a dehazing model followed by gamma correction, or a standard low-light enhancement method) is included, so the contribution of the generative brightness mapping itself is not isolated. I recommend reporting multi-run variance, adding a perceptual user study on brightness fidelity, and adding such baselines.
  3. [Sec. 4.3 (Tables 2-3)] The cross-validation narrative overstates the tabulated results. In Table 3, Fourmer trained on the proposed dataset exceeds Ours on TOPIQ (0.449 vs 0.443), TReS (64.665 vs 61.050), and DBCNN (0.427 vs 0.414); in Table 2, SD-UNREAL-NH exceeds Ours on TReS (68.299 vs 61.050), MUSIQ (52.881 vs 51.098), and DBCNN (0.426 vs 0.414). The text nevertheless concludes that existing models 'cannot achieve effective dehazing ability in terms of quantitative metrics' and that the proposed data and model 'must be used together.' With single runs and margins this small, the conclusion is not supported by the printed numbers; please report variance and reconcile the text with the tables.
  4. [Sec. 4.4 (Fig. 8)] The ablation of the two methodological centerpieces -- the non-uniform brightness adjustment (BA) and the brightness perception network Phi -- is reported only as a percentage derived from summing metrics across three settings, an aggregation that is not defined and cannot be checked against raw values; the individual metric values for the ablations are not given. The text further concedes that adding Phi only 'maintains' the quantitative evaluation performance, so the necessity of Phi -- a headline contribution ('anchored in brightness perception') -- rests on visual inspection alone. Please provide a standard ablation table with the ten metrics for each setting, and add targeted evidence for Phi (e.g., accuracy of the predicted brightness map against the daytime label's brightness map).
  5. [Sec. 3.2 (Eqs. 8-10)] Phi is described as able to 'identify non-uniform brightness in nighttime hazy images,' but its only direct supervision L_M (Eq. 10) compares the mean of the predicted brightness map with the scalar mean brightness of the daytime label; the spatial structure of the predicted map is not directly supervised. The spatial modulation in Eq. (9) can in principle be shaped indirectly by the pixel-wise loss L_pc (Eq. 11), but as written the claimed non-uniform-brightness perception capability is not a direct consequence of L_M. Please either define M as a per-pixel map (if that is intended), visualize what Phi learns, or add per-pixel supervision.
minor comments (8)
  1. [Sec. 3.1] The prose states that the mask is 'split by the area where depth g(i,j) is larger than the threshold theta,' but Eqs. (2) and (3) use the condition 1-g(i,j) >= theta (i.e., g(i,j) <= 0.02 for theta = 0.98); please align the text with the equations.
  2. [Sec. 3.1] The sentence describing the sky brightness normalization ('the brightness of the sky area is adjusted to make the mean value close to the preset level mu') is not formalized in any equation; with mu = 0.85, the synthesis pipeline cannot be exactly reproduced from the text as written.
  3. [Tables 2 and 3] Several entries are typeset with numerals running together (e.g., '59.6300.4380.276' in Table 2, '52.8810.426' in the same table, and '64.6650.345' in Table 3); please regenerate the tables with proper numeric spacing.
  4. [Figs. 7, 11, 12, 13] The captions and in-figure legends of Figures 7, 11, 12, and 13 contain garbled unicode sequences (e.g., '/uni0000002b/uni00000044/...'), and Figure 2's rendered equations inside the diagram are partially unreadable; please regenerate the figures before resubmission.
  5. [Sec. 4.1] The dataset is a stated contribution but is not released, and no download link, sample, or detailed split (beyond the counts 1,546 training and 440 test images) is provided, which limits reproducibility of the reported experiments.
  6. [Sec. 4.1] There is a typo in 'Places365 datadset' (should be 'dataset'); reference [2] is labeled 'IEEE WAVC' but should be 'IEEE WACV'.
  7. [Sec. 4.2 (Table 1, Fig. 4)] The 'state-of-the-art dehazing model [8]' used to produce the comparison rows is never named in the main text; identifying it as SFSNiD (as in Fig. 4 and Table 3) would make Table 1 easier to interpret.
  8. [Sec. 4.4 (Fig. 13)] The claim that prompt-based brightness changes preserve scene content is supported only by visual inspection; a quantitative content-consistency measure (e.g., LPIPS or SSIM between the daytime- and nighttime-prompt outputs) would make the claim checkable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DiffND is trained on synthetic night/day pairs and evaluated on an external real-world benchmark; the claimed “predictions” are not constructed from the fitted brightness constants.

full rationale

The derivation chain is self-contained. The data synthesis pipeline (Eqs. 1–6) starts from real daytime images y_d^r and constructs synthetic hazy nighttime images x_s^n using hand-set non-uniform brightness mapping (Eqs. 2–4, with α=4, φ_1=2, φ_2=1.5, ρ, and ϑ=0.98) and simulated distortion (Eqs. 5–6, with ξ_1=1, ξ_2=3, ξ_3=1.8). These constants are chosen a priori and are not fitted to the model’s own outputs or to the RWNH test set. The network is trained on these synthetic pairs with a pixel-wise reconstruction loss (Eq. 11) against the original real daytime labels, and the brightness supervision M in Eq. 10 is the mean brightness of the target daytime label, not a quantity derived from the model’s output; hence there is no self-definitional or fitted-input-called-prediction step. Real-world evaluation on RWNH (Fig. 1, Fig. 4, Table 1) is an external benchmark, and the no-reference IQA metrics are external tools, not functions of the synthesis constants. The self-citations to prior work ([8], [15], [30], [58]) are used for baseline implementations, comparison methods, or reusable modules such as Γ_z and pixel-wise Gamma, none of which carries the load of the central claim by itself. The synthetic-to-real transfer assumption is a genuine validation gap—there are no aligned real night/day pairs and no distributional validation of the brightness mapping—but that is a correctness/external-validity concern, not circularity, and the paper does not rename a fitted parameter as a prediction or invoke a self-authored uniqueness theorem. Therefore the paper receives a circularity score of 0.

Assumptions & free parameters 10 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a set of hand-set synthesis parameters (alpha, phi1, phi2, rho, xi1-3, mu, beta, lambda1-2, LoRA ranks) and several domain assumptions about the fidelity of the degradation model, the transferability of a pretrained diffusion prior, and the validity of no-reference IQA metrics. No quantities are fitted to the test set, and the real-world test set RWNH is external to training, so the circularity burden is low. The brightness perception network is an internal module with no independent evidence beyond paper-specific ablations.

free parameters (10)
  • degradation factor alpha = 4
    Gamma power in Eq. 4 that darkens daytime images; set empirically, no sensitivity analysis.
  • brightness difference factor rho = not specified
    Divides the sky brightness in Eq. 3; the paper reports using it but does not give its value.
  • sky illumination factor phi1 = 2
    Multiplies the illumination mask for sky regions in Eq. 2.
  • non-sky illumination factor phi2 = 1.5
    Multiplies the illumination mask for non-sky regions in Eq. 2.
  • sky segmentation threshold rho_vartheta = 0.98
    Threshold on 1-depth to decide sky vs non-sky in Eqs. 2-3; empirically set in Section 4.1.
  • sky brightness level mu = 0.85
    Target mean brightness for sky area; empirically set.
  • point light constant beta = 1
    Constant in Eq. 6 for point light intensity.
  • attenuation parameters xi1, xi2, xi3 = 1, 3, 1.8
    Control the attenuation function in Eq. 5; follow game-engine settings from [30].
  • loss weights lambda1, lambda2 = 0.5, 1
    Weights for adversarial and brightness losses in Eq. 13.
  • LoRA ranks for UNet and encoder/decoder = 8 and 4
    LoRA ranks used for fine-tuning the diffusion model; set in Section 4.1.
assumptions (5)
  • domain assumption The severe degradation model from [27] (Eq. 1) adequately models real nighttime haze when the real distortions are replaced by simulated ones.
    The paper replaces W and T in Eq. 1 with simulated quantities and assumes the linear model still captures real nighttime haze formation.
  • domain assumption Pre-trained SD-Turbo/CycleGAN-Turbo contains a strong prior for clear daytime images that can be adapted via LoRA.
    Used as the backbone in Section 3.2; the prior is assumed to transfer to the NDBM task after fine-tuning.
  • domain assumption Depth Anything V2 and the sky segmentation method produce sufficiently accurate depth and sky masks on Places365 images.
    These masks drive the non-uniform brightness mapping in Eqs. 2-3; errors would propagate into the training data.
  • domain assumption No-reference IQA metrics (QualiCLIP, ARNIQA, TOPIQ, MUSIQ, DBCNN, etc.) are valid for judging enhancement quality.
    All quantitative claims in Tables 1-3 rely on these metrics; no human evaluation is provided.
  • ad hoc to paper The reference brightness label M is the mean brightness of the target daytime image.
    Defined in Section 3.2 to train the brightness perception network; a design choice not derived from theory.
invented entities (1)
  • Brightness perception network Phi
    purpose: Predict a non-uniform brightness map from early encoder features to gate skip connections and guide lighting reconstruction.
    Introduced in Eq. 8; validated only by internal ablations (Figs. 8-9), with no external falsifiable prediction or independent test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Devil is in the Darkness: Diffusion-Based Nighttime Dehazing Anchored in Brightness Perception." pith.science (2026). https://pith.science/paper/W4PQ6QQS

@misc{pith2026250602395,
  author       = {Pith},
  title        = {Pith review of: The Devil is in the Darkness: Diffusion-Based Nighttime Dehazing Anchored in Brightness Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4PQ6QQS}},
  note         = {Machine review of arXiv:2506.02395}
}
read the original abstract

While nighttime image dehazing has been extensively studied, converting nighttime hazy images to daytime-equivalent brightness remains largely unaddressed. Existing methods face two critical limitations: (1) datasets overlook the brightness relationship between day and night, resulting in the brightness mapping being inconsistent with the real world during image synthesis; and (2) models do not explicitly incorporate daytime brightness knowledge, limiting their ability to reconstruct realistic lighting. To address these challenges, we introduce the Diffusion-Based Nighttime Dehazing (DiffND) framework, which excels in both data synthesis and lighting reconstruction. Our approach starts with a data synthesis pipeline that simulates severe distortions while enforcing brightness consistency between synthetic and real-world scenes, providing a strong foundation for learning night-to-day brightness mapping. Next, we propose a restoration model that integrates a pre-trained diffusion model guided by a brightness perception network. This design harnesses the diffusion model's generative ability while adapting it to nighttime dehazing through brightness-aware optimization. Experiments validate our dataset's utility and the model's superior performance in joint haze removal and brightness mapping.

Figures

Figures reproduced from arXiv: 2506.02395 by the authors.

Figure 1
Figure 1. Performance on the real-world nighttime dehazing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The data synthesis, network training, and inference processes. “Distortion” represents the simulated distortion. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of data synthesis procedures. The daytime [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Real-world visual comparisons with existing NDBM-Task solutions. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Real-world visual comparisons of the results obtained by training our model under different datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Real-world visual comparisons of the results obtained by training different models under our dataset. The edge maps obtained [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: Visual evaluation of the brightness adjust (BA) and [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 8
Figure 8. Figure 8: Quantitative evaluation of the brightness adjustment [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 10
Figure 10. Figure 10: Visual evaluation of fine-tuning and the [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 12
Figure 12. Figure 12: Quantitative evaluation of hyperparameters. [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 13
Figure 13. Figure 13: Results obtained with different prompts. By changing [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 56 canonical work pages

  1. [1]

    Quality-aware image-text alignment for real-world image quality assessment.arXiv preprint arXiv:2403.11176, 2024

    Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-aware image-text alignment for real-world image quality assessment.arXiv preprint arXiv:2403.11176, 2024. 6

    arXiv 2024

  2. [2]

    Arniqa: Learning distortion manifold for image quality assessment

    Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. InIEEE WAVC, pages 189– 198, 2024. 6

    work page 2024

  3. [3]

    Night- time image-dehazing: a review and quantitative benchmark- ing.Archives of Computational Methods in Engineering, 28(4):2943–2975, 2021

    Sriparna Banerjee and Sheli Sinha Chaudhuri. Night- time image-dehazing: a review and quantitative benchmark- ing.Archives of Computational Methods in Engineering, 28(4):2943–2975, 2021. 1

    work page 2021

  4. [4]

    IQA-PyTorch: Pytorch toolbox for image quality assessment

    Chaofeng Chen and Jiadi Mo. IQA-PyTorch: Pytorch toolbox for image quality assessment. [Online]. Avail- able:https : / / github . com / chaofengc / IQA - PyTorch, 2022. 6

    work page 2022

  5. [5]

    Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE TIP, 2024

    Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE TIP, 2024. 6

    work page 2024

  6. [6]

    Ni-dehazenet: rep- resentation learning via bilevel optimized architecture search for nighttime dehazing.The Visual Computer, 40(9):6155– 6170, 2024

    Hui Chen, Nannan Li, and Rong Chen. Ni-dehazenet: rep- resentation learning via bilevel optimized architecture search for nighttime dehazing.The Visual Computer, 40(9):6155– 6170, 2024. 1

    work page 2024

  7. [7]

    Style injec- tion in diffusion: A training-free approach for adapting large- scale diffusion models for style transfer

    Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injec- tion in diffusion: A training-free approach for adapting large- scale diffusion models for style transfer. InIEEE CVPR, pages 8795–8805, 2024. 2

    work page 2024

  8. [8]

    A semi-supervised nighttime dehazing baseline with spatial-frequency aware and realistic brightness constraint

    Xiaofeng Cong, Jie Gui, Jing Zhang, Junming Hou, and Hao Shen. A semi-supervised nighttime dehazing baseline with spatial-frequency aware and realistic brightness constraint. InIEEE CVPR, pages 2631–2640, 2024. 1, 2, 3, 4, 5, 6, 7

    work page 2024

Show all 63 references
  1. [9]

    Focal network for image restoration

    Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Focal network for image restoration. InIEEE ICCV, pages 13001–13011, 2023. 1

    2023

  2. [10]

    Omni-kernel net- work for image restoration

    Yuning Cui, Wenqi Ren, and Alois Knoll. Omni-kernel net- work for image restoration. InAAAI, volume 38, pages 1426–1434, 2024. 1, 2

    2024

  3. [11]

    Flare7k: A phenomenological night- time flare removal dataset.NeurIPS, 35:3926–3937, 2022

    Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological night- time flare removal dataset.NeurIPS, 35:3926–3937, 2022. 2

    2022

  4. [12]

    No-reference image quality assessment via transformers, rel- ative ranking, and self-consistency

    S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, rel- ative ranking, and self-consistency. InIEEE WACV, pages 3209–3218, 2022. 6

    2022

  5. [13]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021. 5

    2021 arXiv

  6. [14]

    Raindrop clarity: A dual-focused dataset for day and night raindrop removal

    Yeying Jin, Xin Li, Jiadong Wang, Yan Zhang, and Malu Zhang. Raindrop clarity: A dual-focused dataset for day and night raindrop removal. InECCV, pages 1–17, 2024. 1

    2024

  7. [15]

    Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution

    Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution. InACM MM, 2023. 1, 2, 3, 4, 6, 7

    2023

  8. [16]

    Structure representation network and uncertainty feedback learning for dense non-uniform fog removal

    Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. InACCV, pages 155–172. Springer, 2022. 2

    2022

  9. [17]

    Unsuper- vised night image enhancement: When layer decomposition meets light-effects suppression

    Yeying Jin, Wenhan Yang, and Robby T Tan. Unsuper- vised night image enhancement: When layer decomposition meets light-effects suppression. InECCV, pages 404–421. Springer, 2022. 2

    2022

  10. [18]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. InIEEE CVPR, pages 9492–9502, 2024. 3

    2024

  11. [19]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InIEEE ICCV, pages 5148–5157, 2021. 6

    2021

  12. [20]

    2pcnet: Two-phase consistency training for day-to-night unsupervised domain adaptive object detec- tion

    Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. 2pcnet: Two-phase consistency training for day-to-night unsupervised domain adaptive object detec- tion. InIEEE CVPR, pages 11484–11493, 2023. 1

    2023

  13. [21]

    Multi-path dilated convolution network for haze and glow removal in nighttime images.The Visual Com- puter, pages 1–14, 2022

    Shiba Kuanar, Dwarikanath Mahapatra, Monalisa Bilas, and KR Rao. Multi-path dilated convolution network for haze and glow removal in nighttime images.The Visual Com- puter, pages 1–14, 2022. 1

    2022

  14. [22]

    Ensembling off-the-shelf models for gan training

    Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In IEEE CVPR, pages 10651–10662, 2022. 6

    2022

  15. [23]

    Benchmarking single- image dehazing and beyond.IEEE TIP, 28(1):492–505,

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single- image dehazing and beyond.IEEE TIP, 28(1):492–505,

  16. [24]

    Learning transferable negative prompts for out-of- distribution detection

    Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of- distribution detection. InIEEE CVPR, pages 17584–17594,

  17. [25]

    Nighttime haze removal with glow and multiple light colors

    Yu Li, Robby T Tan, and Michael S Brown. Nighttime haze removal with glow and multiple light colors. InIEEE ICCV, pages 226–234, 2015. 2

    2015

  18. [26]

    9 Hdp-net: Haze density prediction network for nighttime de- hazing

    Yinghong Liao, Zhuo Su, Xiangguo Liang, and Bin Qiu. 9 Hdp-net: Haze density prediction network for nighttime de- hazing. InPacific Rim Conference on Multimedia, pages 469–480, 2018. 1, 3

    2018

  19. [27]

    Nighthaze: Nighttime image dehazing via self-prior learning

    Beibei Lin, Yeying Jin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T Tan. Nighthaze: Nighttime image dehazing via self-prior learning. InAAAI, 2025. 1, 2, 3, 4

    2025

  20. [28]

    Nightrain: Nighttime video deraining via adaptive-rain-removal and adaptive-correction

    Beibei Lin, Yeying Jin, Wending Yan, Wei Ye, Yuan Yuan, Shunli Zhang, and Robby T Tan. Nightrain: Nighttime video deraining via adaptive-rain-removal and adaptive-correction. InAAAI, volume 38, pages 3378–3385, 2024. 2

    2024

  21. [29]

    Single nighttime image dehazing based on image decomposition

    Yun Liu, Anzhi Wang, Hao Zhou, and Pengfei Jia. Single nighttime image dehazing based on image decomposition. Signal Processing, 183:107986, 2021. 1

    2021

  22. [30]

    Nighthazeformer: Single nighttime haze removal using prior query transformer

    Yun Liu, Zhongsheng Yan, Sixiang Chen, Tian Ye, Wenqi Ren, and Erkang Chen. Nighthazeformer: Single nighttime haze removal using prior query transformer. InACM MM,

  23. [31]

    Multi-purpose oriented single nighttime image haze removal based on unified variational retinex model.IEEE TCSVT, 33(4):1643–1657, 2022

    Yun Liu, Zhongsheng Yan, Jinge Tan, and Yuche Li. Multi-purpose oriented single nighttime image haze removal based on unified variational retinex model.IEEE TCSVT, 33(4):1643–1657, 2022. 2

    2022

  24. [32]

    Nighttime image dehazing based on variational decom- position model

    Yun Liu, Zhongsheng Yan, Aimin Wu, Tian Ye, and Yuche Li. Nighttime image dehazing based on variational decom- position model. InIEEE CVPR Workshops, pages 640–649,

  25. [33]

    Yun Liu, Zhongsheng Yan, Tian Ye, Aimin Wu, and Yuche Li. Single nighttime image dehazing based on unified vari- ational decomposition model and multi-scale contrast en- hancement.Engineering Applications of Artificial Intelli- gence, 116:105373, 2022. 1

    2022

  26. [34]

    Zero-shot image-to-image translation

    Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. InACM SIGGRAPH, pages 1–11, 2023. 3, 5

    2023

  27. [35]

    One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024. 2, 3, 5, 6, 8

    2024 arXiv

  28. [36]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image syn- thesis with latent diffusion models. InIEEE CVPR, pages 10684–10695, 2022. 3, 4

    2022

  29. [37]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InECCV, pages 87–103, 2024. 2, 3, 5

    2024

  30. [38]

    Single-image camera response function using prediction consistency and gradual refinement

    Aashish Sharma, Robby T Tan, and Loong-Fah Cheong. Single-image camera response function using prediction consistency and gradual refinement. InACCV, 2020. 4

    2020

  31. [39]

    Freeu: Free lunch in diffusion u-net

    Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. InIEEE CVPR, pages 4733–4743, 2024. 3

    2024

  32. [40]

    Argue: Attribute-guided prompt tuning for vision-language models

    Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. InIEEE CVPR, pages 28578–28587, 2024. 3

    2024

  33. [41]

    Ex- ploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. In AAAI, 2023. 6

    2023

  34. [42]

    Variational single nighttime image haze removal with a gray haze-line prior

    Wenhui Wang, Anna Wang, and Chen Liu. Variational single nighttime image haze removal with a gray haze-line prior. IEEE TIP, 31:1349–1363, 2022. 2

    2022

  35. [43]

    Robust real-world image dehazing via knowledge guided conditional diffusion model finetuning

    Haoran Wei, Qingbo Wu, Chenhao Wu, Shuai Chen, Lei Wang, King Ngi Ngan, Fanman Meng, and Hongliang Li. Robust real-world image dehazing via knowledge guided conditional diffusion model finetuning. InIEEE Workshop on MMSP, pages 1–6. IEEE, 2024. 6

    2024

  36. [44]

    Robust unpaired image dehazing via adversarial deformation con- straint.IEEE TCSVT, 2024

    Haoran Wei, Qingbo Wu, Chenhao Wu, King Ngi Ngan, Hongliang Li, Fanman Meng, and Heqian Qiu. Robust unpaired image dehazing via adversarial deformation con- straint.IEEE TCSVT, 2024. 6

    2024

  37. [45]

    Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023. 6

    2023 arXiv

  38. [46]

    A comprehensive study of multimodal large language mod- els for image quality assessment

    Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language mod- els for image quality assessment. InECCV, pages 143–160. Springer, 2024. 6

    2024

  39. [47]

    Prompt-free diffusion: Taking text out of text-to-image diffusion models

    Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Ir- fan Essa, and Humphrey Shi. Prompt-free diffusion: Taking text out of text-to-image diffusion models. InIEEE CVPR, pages 8682–8692, 2024. 3

    2024

  40. [48]

    Diffu- sion models without attention

    Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffu- sion models without attention. InIEEE CVPR, pages 8239– 8249, 2024. 3

    2024

  41. [49]

    Night- time defogging using high-low frequency decomposition and grayscale-color networks

    Wending Yan, Robby T Tan, and Dengxin Dai. Night- time defogging using high-low frequency decomposition and grayscale-color networks. InECCV, pages 473–488, 2020. 2

    2020

  42. [50]

    A variation-based nighttime image dehazing flow with a phys- ically valid illumination estimator and a luminance-guided coloring model.IEEE Access, 10:50153–50166, 2022

    Chih-Hsiang Yang, Yi-Hsien Lin, and Yi-Chang Lu. A variation-based nighttime image dehazing flow with a phys- ically valid illumination estimator and a luminance-guided coloring model.IEEE Access, 10:50153–50166, 2022. 2

    2022

  43. [51]

    Depth any- thing v2.arXiv:2406.09414, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.arXiv:2406.09414, 2024. 4

    2024 arXiv

  44. [52]

    Superpixel- based single nighttime image haze removal.IEEE TMM, 20(11):3008–3018, 2018

    Minmin Yang, Jianchang Liu, and Zhengguo Li. Superpixel- based single nighttime image haze removal.IEEE TMM, 20(11):3008–3018, 2018. 2

    2018

  45. [53]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. InIEEE CVPR, pages 1191–1200,

  46. [54]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE CVPR, pages 6613–6623, 2024. 3

    2024

  47. [55]

    Nighttime single image dehazing via pixel-wise alpha blending.IEEE Access, 7:114619–114630,

    Teng Yu, Kang Song, Pu Miao, Guowei Yang, Huan Yang, and Chenglizhao Chen. Nighttime single image dehazing via pixel-wise alpha blending.IEEE Access, 7:114619–114630,

  48. [56]

    Fast haze removal for nighttime image using maximum reflectance prior

    Jing Zhang, Yang Cao, Shuai Fang, Yu Kang, and Chang Wen Chen. Fast haze removal for nighttime image using maximum reflectance prior. InIEEE CVPR, pages 7418– 7426, 2017. 2

    2017

  49. [57]

    Nighttime haze removal based on a new imaging model

    Jing Zhang, Yang Cao, and Zengfu Wang. Nighttime haze removal based on a new imaging model. InIEEE ICIP, pages 4557–4561, 2014. 1 10

    2014

  50. [58]

    Nighttime dehazing with a synthetic benchmark

    Jing Zhang, Yang Cao, Zheng-Jun Zha, and Dacheng Tao. Nighttime dehazing with a synthetic benchmark. InACM MM, pages 2355–2363, 2020. 1, 2, 3, 4, 5, 6

    2020

  51. [59]

    Blind image quality assessment using a deep bilinear convolutional neural network.IEEE TCSVT, 30(1):36–47,

    Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network.IEEE TCSVT, 30(1):36–47,

  52. [60]

    Blind image quality assessment via vision- language correspondence: A multitask learning perspective

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision- language correspondence: A multitask learning perspective. InIEEE CVPR, pages 14071–14081, 2023. 6

    2023

  53. [61]

    Places: A 10 million image database for scene recognition.IEEE TPAMI, 40(6):1452–1464, 2017

    Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition.IEEE TPAMI, 40(6):1452–1464, 2017. 6

    2017

  54. [62]

    Fourmer: an efficient global modeling paradigm for image restoration

    Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: an efficient global modeling paradigm for image restoration. InICML, pages 42589–42601, 2023. 3, 6, 7

    2023

  55. [63]

    Edge-aware guidance fusion network for rgb–thermal scene parsing

    Wujie Zhou, Shaohua Dong, Caie Xu, and Yaguan Qian. Edge-aware guidance fusion network for rgb–thermal scene parsing. InAAAI, volume 36, pages 3571–3579, 2022. 6 11

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.