Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Single-Step Latent Diffusion for Underwater Image Restoration

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SLURPP restores underwater images in one latent diffusion step by jointly predicting the clear scene and water medium, hitting 25.66 dB PSNR on the Osmosis benchmark at over 200 times its speed.

desk verdict Strong synthetic results and a real speedup, but the real-world SOTA claim rests on reference-free metrics after post-hoc image filtering; deserves review with demands for better real-world evidence. read the letter →

arxiv 2507.07878 v2 pith:T4FVCUDH submitted 2025-07-10 cs.CV

Jiayi Wu , Tianfu Wang , Md Abu Bakr Siddique , Md Jahidul Islam , Cornelia Fermuller , Yiannis Aloimonos , Christopher A. Metzler This is my paper
classification cs.CV
keywords underwaterimagerestorationlatentdiffusionmodelssingle-stepinferenceformationmodelsyntheticdatagenerationtransmissionestimationbackscatteringphysics-basedtraining
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 claims that underwater image restoration can be reduced to a single forward pass of a fine-tuned latent diffusion model, rather than the iterative denoising used by prior diffusion approaches. The proposed network, SLURPP, takes an underwater image and jointly predicts the clear scene $J$, the transmission $T$, and the backscattering $B$ under the formation model $I = J \cdot T + B$, with two U-Net branches sharing cross-attention and initialized from different pretrained diffusion priors. Because the target distribution in restoration is narrow, the authors argue the diffusion process can be collapsed to one step, which lets them supervise directly with image-space losses and a self-supervised reconstruction loss. If this is correct, diffusion-based underwater restoration becomes practical for real-time and high-resolution use while improving quality: 25.66 dB PSNR on the Osmosis synthetic benchmark, about 3 dB above the prior diffusion baseline, with inference at 0.75 seconds for a 512x512 image.

What carries the argument

The load-bearing mechanism is a dual-branch single-step latent diffusion fine-tuning scheme held together by a physics-constrained training objective. Two U-Net branches process the frozen latent encoding of the input: the scene branch, initialized from a text-to-image latent diffusion model, predicts the clear-scene latent, while the medium branch, initialized from a monocular depth diffusion model, predicts transmission and backscattering latents. A cross-attention module connects the branches, and a cross-latent decoder with zero-convolution skip connections carries high-frequency input details into the restored image. The governing identity is the dense scene-medium decomposition $I^c = J^c T^c + B^c$, enforced during training by a reconstruction loss (weight 0.4) alongside $L_1$, SSIM, and LPIPS losses on each output. Because inference is single-step, the outputs can be supervised directly in image space, which the paper finds beats a 50-step latent-loss fine-tuning baseline (25.66 versus 23.46 dB PSNR).

What would settle it

Build a calibrated imaging setup in a tank with measurable water optics: place color charts at several known depths, capture images under controlled turbidity, and compare SLURPP's predicted transmission and backscattering against values computed from the known attenuation coefficients and depths. Systematic divergence of predicted medium parameters or restored colors as turbidity increases would indicate that the synthetic-training assumption does not carry over to real water.

Watch

Extended reading notes

Core claim

SLURPP is a single-step dual-branch latent diffusion network that jointly predicts the clear scene $J$ and the dense medium parameters $T$ (transmission) and $B$ (backscattering) from one underwater image, subject to the physically derived decomposition $I = J \cdot T + B$. The scene branch is initialized from a pretrained text-to-image latent diffusion model and the medium branch from a pretrained monocular depth diffusion model; inter-branch cross-attention lets scene content and medium effects serve as mutual cues. Training uses synthetic underwater images generated by physically calibrated sampling of metric depth, attenuation coefficients from measured water types, and background light extracted from real underwater images, with image-space losses on all outputs plus a reconstruction loss enforcing the formation model. The paper reports state-of-the-art restoration on the Osmosis synthetic benchmark (25.66 dB PSNR, 0.95 SSIM), better per-pixel transmission and backscattering estimates than Osmosis, and superior reference-free quality on UIEB and USOD10K, while running more than 200 times faster (0.75 seconds for 512x512 images).

Load-bearing premise

The central assumption is that underwater scenes are well described by the homogeneous-medium formation model used to generate all training data, so if real turbid or spatially varying water deviates from that model, the reported real-world performance may not hold; the paper acknowledges that turbidity is not explicitly modeled.

Editorial extensions

If this is right

  • A single forward pass of a fine-tuned latent diffusion model can replace iterative denoising for restoration tasks whose output distribution is narrow, cutting inference cost by orders of magnitude.
  • Joint medium prediction provides physically interpretable transmission and backscattering maps, so restoration comes with per-pixel estimates of how the water has degraded the image.
  • The physically calibrated synthesis pipeline (metric depth, real measured water optics, real background light) is shown to be worth about 0.9 dB over randomized water parameters and about 3 dB over the prior diffusion baseline.
  • The approach supports restoration at up to 2K resolution, compared with 256x256 for the pixel-space RGBD diffusion baseline.
  • Direct image-space supervision during single-step fine-tuning outperforms latent-space noise supervision, suggesting that reconstruction-quality losses matter more than denoising trajectory fidelity for this task.

Reading between the lines

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

  • The same single-step recipe should transfer to other inverse problems with narrow target distributions, such as dehazing, deblurring, or low-light enhancement, where current few-step diffusion distillations inherit the teacher's ceiling.
  • Because transmission and backscattering are depth-correlated, the predicted medium maps could serve as a dense depth proxy for underwater robots, a use the paper does not develop.
  • A testable extension is fine-tuning on unpaired real underwater images using only the reconstruction loss; if that loss indeed closes the synthetic-to-real gap, labeled synthetic data would matter less.
  • The failure mode most worth probing is turbidity: since the formation model does not include it, performance should degrade in turbid water, and the paper's own video results show flickering in scenes with particles and ambiguous depth.
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

3 major / 5 minor

Summary. This manuscript introduces SLURPP, a single-step latent diffusion network with a scene branch and a medium branch that jointly predict the clear underwater image J and the dense transmission T and backscattering B maps. Training uses a physics-based synthetic data pipeline built on the Jaffe-McGlamery formation model, with metric depth from Depth Pro, attenuation sampled from Jerlov water types, and background light extracted from real underwater images via ULAP. The authors report 25.66 dB PSNR on the Osmosis/NYU synthetic benchmark, about 3 dB over Osmosis, with 0.75 s inference for 512x512 images, and report state-of-the-art UIQM/MUSIQ scores on the USOD10K and UIEB real-world datasets.

Significance. The method is a well-engineered application of latent diffusion priors to underwater restoration. The synthetic evaluation is internally consistent: ablations isolate the contributions of single-step training, the reconstruction loss, the cross-latent decoder, and the data synthesis pipeline, and the medium-prediction comparison in the supplementary shows large gains over Osmosis on synthetic ground truth. If the real-world claim were supported by stronger evidence, the roughly 200x speedup over iterative diffusion baselines would be practically important. The paper also provides a clear disentangling of network design and training-data contributions, which is genuinely informative for the community.

major comments (3)
  1. [Section 4.2, Table 1] The real-world claim of state-of-the-art performance is based on no-reference metrics UIQM and MUSIQ computed after removing approximately 10% of images with 'apparent image artifacts' from the evaluation set. This exclusion is not applied to baseline methods, no standard deviations or significance tests are reported, and no-reference metrics primarily reward contrast and colorfulness rather than physical fidelity of the restored radiance. The statement 'Our method achieves state-of-the-art performance across all metrics on both datasets' is therefore not adequately supported; the authors should report metrics on the full unfiltered datasets, provide score distributions, and supplement with a human perceptual study or a real-world dataset with reference ground truth such as SQUID.
  2. [Section 3.5.1, Eq. (4), Table 5] The reconstruction loss L_UIFM is claimed to 'play a key role in mitigating the domain gap' because it enforces Eq. (2). However, Eq. (2) is an identity for any predicted J if one defines B = I - J*T, so this loss cannot by itself enforce physical correctness of the decomposition; physical correctness is inherited from the synthetic supervision on J, T, and B. The observed 0.86 dB improvement from L_UIFM in Table 5 on the synthetic benchmark is consistent with a regularization effect and does not directly evidence domain-gap mitigation. The authors should either soften the claim or demonstrate the effect on a synthetic test set with domain mismatch, such as simulated non-homogeneous water parameters.
  3. [Section 3.4 and Section 11] The training pipeline renders degradation with a single global background light and spatially homogeneous attenuation coefficients per image, and Section 11 concedes that turbidity is not modeled. The real-world generalization of the method therefore rests on an unverified assumption that Eq. (1) with these simplifications is a faithful proxy for real turbid, spatially varying water. The qualitative real-world results in Figs. 6, 7, and 11, while visually appealing, are not accompanied by any quantitative validation of the predicted T and B against real measurements, so the paper's 'physically grounded' claim for real scenes is not established. Please add a quantitative evaluation on real scenes with known optical properties, or explicitly reframe the real-world contribution as qualitative restoration quality rather than physical parameter recovery.
minor comments (5)
  1. [Section 4.2] The speedup comparison ('over 200×') compares 0.75 s for a 512×512 image with over 200 s for a 256×256 image from Osmosis, conflating resolution with inference cost; please report timings at the same resolution or a per-pixel cost.
  2. [Table 2] The Phaseformer row cites reference [24], but Phaseformer is reference [25] in the bibliography; please correct the citation.
  3. [Introduction and Supplementary Section 7] There are typographical errors: 'diffcult' should be 'difficult' in the Introduction, and 'objectice' should be 'objective' in Supplementary Section 7.
  4. [Table 1] Please report the exact number of images used after filtering and the standard deviation of the metrics; without this, the margin over the second-best method on UIQM is difficult to interpret.
  5. [Section 3.5.2] The cross-latent decoder is described as fine-tuned with 'zero convolution' skip connections, but the details of how the skip connections are integrated into the frozen VAE decoder are not fully specified, which may hinder reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central restoration result is an empirical supervised-learning claim validated on held-out synthetic benchmarks, and it does not reduce to its own inputs.

full rationale

The paper's core claim is an empirical result: a dual-branch single-step latent diffusion model trained on synthetic underwater images and evaluated on held-out synthetic data (Osmosis/NYU) and on reference-free real-world metrics. The derivation chain does not reduce to its inputs: ground-truth J, T, and B are generated from Eq. (1) using independently sourced depth (Depth Pro), Jerlov water-type attenuation coefficients, and ULAP background lights; the network is supervised on these labels, and the held-out Osmosis benchmark is disjoint from the training set. The only identity-like term is the self-supervised reconstruction loss L_UIFM enforcing I = J·T + B, since for any J one can define B = I - J·T; however, the paper does not claim this loss alone enforces physical correctness — physical meaning is inherited from the synthetic labels, and L_UIFM is presented as a regularizer for domain gap. The cited self-works (e.g., [9], [17], [21]) are design precedents and related-work citations, not load-bearing uniqueness theorems. Overall, no circular step merits flagging; the main caveats (synthetic-to-real transfer, reference-free metrics, turbidity not explicitly modeled) are empirical-risk and correctness concerns, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method is a trained neural network, not a mathematical derivation; the central results rest on the fidelity of the synthetic training distribution and on the pretrained diffusion priors. The loss weights and synthesis thresholds are hand-set, and several modeling assumptions (homogeneous water, Jerlov sampling, Depth Pro accuracy) are taken from prior literature without independent verification in this paper.

free parameters (4)
  • Loss weights lambda_T, lambda_B, lambda_L = 0.5, 0.5, 0.4
    Chosen by hand in Eq. (4); the reconstruction loss weight lambda_L=0.4 contributes a 0.86 dB PSNR gain in the ablation, so the central result depends on these settings.
  • Number of background light clusters K = 10
    K-means cluster count for B_inf extraction in Sec. 3.4; affects diversity of training water colors.
  • Attenuation bound and resampling criterion = not specified
    Beta values are bounded and samples are resampled 'when excessive information loss occurs', but the exact thresholds are not reported; this is a hand-tuned data-generation setting.
  • Training image resolution and learning rate = 512x512, 1e-5
    Standard training hyperparameters chosen without reported tuning; stage 1 takes 2 days, stage 2 1 day on one A6000.
assumptions (4)
  • domain assumption Jaffe-McGlamery underwater image formation model, Eq. (1), with homogeneous attenuation and backscatter coefficients and known depth z.
    Used to synthesize all training labels (J, T, B) and to define the reconstruction loss in Eq. (4); real scenes with turbidity or non-uniform water violate it, as the authors acknowledge in Sec. 11.
  • domain assumption Pretrained Stable Diffusion V2 and Marigold contain natural-image and depth priors useful for underwater restoration.
    Scene branch is initialized from SDV2 and medium branch from Marigold (Sec. 3.3); the method's performance depends on these upstream models' priors, which were not trained on underwater data.
  • domain assumption Depth Pro provides accurate metric depth estimates for terrestrial images used in synthesis.
    Metric depth maps from Depth Pro are used to compute transmission/backscatter in Eq. (1); errors in Depth Pro propagate into training labels.
  • domain assumption Jerlov water-type measurements mapped to RGB attenuation coefficients at 600/525/475 nm are representative of real water types.
    Random sampling from 10 Jerlov water types guides beta in Sec. 3.4; the band-to-RGB mapping is a modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Single-Step Latent Diffusion for Underwater Image Restoration." pith.science (2026). https://pith.science/paper/T4FVCUDH

@misc{pith2026250707878,
  author       = {Pith},
  title        = {Pith review of: Single-Step Latent Diffusion for Underwater Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4FVCUDH}},
  note         = {Machine review of arXiv:2507.07878}
}
read the original abstract

Underwater image restoration algorithms seek to restore the color, contrast, and appearance of a scene that is imaged underwater. They are a critical tool in applications ranging from marine ecology and aquaculture to underwater construction and archaeology. While existing pixel-domain diffusion-based image restoration approaches are effective at restoring simple scenes with limited depth variation, they are computationally intensive and often generate unrealistic artifacts when applied to scenes with complex geometry and significant depth variation. In this work we overcome these limitations by combining a novel network architecture (SLURPP) with an accurate synthetic data generation pipeline. SLURPP combines pretrained latent diffusion models -- which encode strong priors on the geometry and depth of scenes -- with an explicit scene decomposition -- which allows one to model and account for the effects of light attenuation and backscattering. To train SLURPP we design a physics-based underwater image synthesis pipeline that applies varied and realistic underwater degradation effects to existing terrestrial image datasets. This approach enables the generation of diverse training data with dense medium/degradation annotations. We evaluate our method extensively on both synthetic and real-world benchmarks and demonstrate state-of-the-art performance. Notably, SLURPP is over 200X faster than existing diffusion-based methods while offering ~ 3 dB improvement in PSNR on synthetic benchmarks. It also offers compelling qualitative improvements on real-world data. Project website https://tianfwang.github.io/slurpp/.

Figures

Figures reproduced from arXiv: 2507.07878 by the authors.

Figure 1
Figure 1. Real-world underwater restoration using our method. We develop a single-step underwater restoration method that leverages pretrained latent diffusion priors. Given an underwater input image (top row), our method jointly predicts the clear image (middle row), and the per-pixel underwater medium parameters, specifically the backscattering (bottom row left) and transmission (bottom row right) parameters. In this figure… view at source ↗
Figure 2
Figure 2. Our method captures the depth-varying change of water medium properties. In this figure we demonstrate the depth-dependent nature of the underwater medium effects and show that our method can correctly capture this in our medium predictions. We can see in the zoomed-in regions of the input underwater image (right first row), that the water medium effects increase as we move from the foreground to the background. Thi… view at source ↗
Figure 3
Figure 3. Pipeline overview of our single-step dual-branch underwater restoration method. Our pipeline takes in an underwater image and aims to predict a clear image without water effects, along with the transmission and backscattering properties of the water medium in a single step. The input image is first encoded into latent space using the frozen VAE from pretrained Stable Diffusion (SD) [4]. This latent image is then fed… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Physically accurate underwater image synthetics pipeline for diverse data generation. Our model is trained on realistic underwater images synthesized from large-scale terrestrial data using precise mod￾eling of depth, attenuation, and background light for physically ac…
Figure 6
Figure 6. Figure 6: Qualitative comparisons of restoration results in USOD10K [52] UIEB [3]. We show extensive comparisons against previous methods [3], [22], [23], [24], [25], [26], [27], [53], [54]. As illustrated in the comparison, previous methods often struggle to achieve physically …
Figure 7
Figure 7. Figure 7: Comparison with Osmosis [27] on the UIEB dataset [3]. In this figure we show the predicted clear image and medium-related parame￾ters for our method and Osmosis. In the medium visualization of both methods, objects in the foreground have lower depth/backscattering, whi…
Figure 9
Figure 9. Figure 9: Cross-latent decoding enhances restoration details. Even though our method uses image loss during training, the limitations of the vanilla SD [4] decoder could still hallucinate high-frequency details. The cross-latent decoding allows us to obtain restorations with bet…
Figure 10
Figure 10. Figure 10: More restoration results of our method on real-world datasets [3], [52]. We showcase more real-world restoration image results of our method that we find visually appealing. TABLE 3 Comparing our single-step method to iterative latent loss fine-tuning methods. We pres…
Figure 11
Figure 11. Figure 11: Visualizing our model’s performance under challenging underwater scenarios. We show method restoring real world examples from [3], [52] with challenging lighting scenarios and strong turbidity. color accuracy of our restored images. We show this effect in [PITH_FULL_…
Figure 12
Figure 12. Figure 12: Our method restores more accurate colors compared to other methods. When evaluating restored outputs on real-world underwater datasets [3], [52], we observe that our method more accurately recovers the true color profiles of objects in both the foreground and backgrou…
Figure 13
Figure 13. Figure 13: Additional real-world comparisons on real-world underwater datasets [3], [52]. Our method (second column from the left) achieves physically consistent results across varying depths, with improved performance in degraded distant regions. It also accurately estimates pe…
Figure 14
Figure 14. Figure 14: More visualizations of synthetic training data. We show extensive visualizations of training data samples. We use a wide range of terrestrial image data sources as the clean image. Combined with our physically-accurate data synthesis pipeline, we generate diverse and …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 52 canonical work pages

  1. [1]

    A Revised Underwater Image Formation Model,

    D. Akkaynak and T. Treibitz, “A Revised Underwater Image Formation Model,” in CVPR, 2018, pp. 6723–6732

    work page 2018

  2. [2]

    AquaFuse: Wa- terbody Fusion for Physics Guided View Synthesis of Underwater Scenes,

    M. Siddique, J. Wu, I. Rekleitis, and M. J. Islam, “AquaFuse: Wa- terbody Fusion for Physics Guided View Synthesis of Underwater Scenes,” IEEE Robotics and Automation Letters (RA-L), vol. 10, no. 5, 2025

    work page 2025

  3. [3]

    An underwater image enhancement benchmark dataset and beyond,

    C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,” IEEE transactions on image processing, vol. 29, pp. 4376–4389, 2019

    work page 2019

  4. [4]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022

    2022

  5. [5]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” NeurIPS, 2020

    work page 2020

  6. [6]

    LAION-5b: An open large- scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. W. Gordon, R. Wight- man, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Worts- man, P . Schramowski, S. R. Kundurthy, K. Crowson, L. Schmidt, R. Kaczmarczyk, and J. Jitsev, “LAION-5b: An open large- scale dataset for training next generation image-text models,” in NeurIPS Systems Datasets and Benchmarks Track, 2022

    work page 2022

  7. [7]

    Marigold: Affordable adaptation of diffusion- based image generators for image analysis,

    B. Ke, K. Qu, T. Wang, N. Metzger, S. Huang, B. Li, A. Obukhov, and K. Schindler, “Marigold: Affordable adaptation of diffusion- based image generators for image analysis,” arXiv preprint arXiv:2505.09358, 2025

    arXiv 2025

  8. [8]

    Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image,

    X. Fu, W. Yin, M. Hu, K. Wang, Y. Ma, P . Tan, S. Shen, D. Lin, and X. Long, “Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image,” in ECCV, 2024

    work page 2024

Show all 63 references
  1. [9]

    Flash-split: 2d reflection removal with flash cues and latent diffusion separation,

    T. Wang, M. Xie, H. Cai, S. Shah, and C. A. Metzler, “Flash-split: 2d reflection removal with flash cues and latent diffusion separation,” arXiv preprint arXiv:2501.00637, 2024

    2024 arXiv

  2. [10]

    Uni-renderer: Unifying rendering and inverse rendering via dual stream diffusion,

    Z. Chen, T. Xu, W. Ge, L. Wu, D. Yan, J. He, L. Wang, L. Zeng, S. Zhang, and Y.-C. Chen, “Uni-renderer: Unifying rendering and inverse rendering via dual stream diffusion,” in CVPR, 2025

    2025

  3. [11]

    Viewactive: Active viewpoint optimization from a single image,

    J. Wu, X. Lin, B. He, C. Fermuller, and Y. Aloimonos, “Viewactive: Active viewpoint optimization from a single image,”arXiv preprint arXiv:2409.09997, 2024

    2024 arXiv

  4. [12]

    Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion,

    T. Xiong, J. Wu, B. He, C. Fermuller, Y. Aloimonos, H. Huang, and C. Metzler, “Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion,” in 8th Annual Conference on Robot Learning, 2024

    2024

  5. [13]

    Sea-Thru: A Method for Removing Water From Underwater Images,

    D. Akkaynak and T. Treibitz, “Sea-Thru: A Method for Removing Water From Underwater Images,” in CVPR, 2019, pp. 1682–1691

    2019

  6. [14]

    UDepth: Fast Monocular Depth Estimation for Visually-guided Underwater Robots,

    B. Yu, J. Wu, and M. J. Islam, “UDepth: Fast Monocular Depth Estimation for Visually-guided Underwater Robots,” in ICRA, 2023

    2023

  7. [15]

    Fine-tuning image-conditional diffusion models is easier than you think,

    G. Martin Garcia, K. Abou Zeid, C. Schmidt, D. de Geus, A. Her- mans, and B. Leibe, “Fine-tuning image-conditional diffusion models is easier than you think,” in WACV, 2025

    2025

  8. [16]

    Fast Underwater Image En- hancement for Improved Visual Perception,

    M. J. Islam, Y. Xia, and J. Sattar, “Fast Underwater Image En- hancement for Improved Visual Perception,” IEEE Robotics and Automation Letters (RA-L), vol. 5, no. 2, pp. 3227–3234, 2020

    2020

  9. [17]

    Simultaneous Enhancement and Super-Resolution of Underwater Imagery for Improved Vi- sual Perception,

    M. J. Islam, P . Luo, and J. Sattar, “Simultaneous Enhancement and Super-Resolution of Underwater Imagery for Improved Vi- sual Perception,” in Robotics: Science and Systems (RSS) , Corvalis, Oregon, USA, July 2020

    2020

  10. [18]

    Marvis: Motion & geometry aware real and virtual image seg- mentation,

    J. Wu, X. Lin, S. Negahdaripour, C. Ferm ¨uller, and Y. Aloimonos, “Marvis: Motion & geometry aware real and virtual image seg- mentation,” in IROS, 2024

    2024

  11. [19]

    3d reconstruction of underwater scenes using nonlinear domain projection,

    J. Wu, B. Yu, and M. J. Islam, “3d reconstruction of underwater scenes using nonlinear domain projection,” in IEEE Conference on Artificial Intelligence (CAI), 2023

    2023

  12. [20]

    Low-cost depth estimation and 3d reconstruction in scat- tering medium,

    J. Wu, “Low-cost depth estimation and 3d reconstruction in scat- tering medium,” Ph.D. dissertation, University of Florida, 2023

    2023

  13. [21]

    Enhancing underwater imagery using generative adversarial networks,

    C. Fabbri, M. J. Islam, and J. Sattar, “Enhancing underwater imagery using generative adversarial networks,” in ICRA, 2018

    2018

  14. [22]

    Fast underwater image en- hancement for improved visual perception,

    M. J. Islam, Y. Xia, and J. Sattar, “Fast underwater image en- hancement for improved visual perception,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3227–3234, 2020

    2020

  15. [23]

    Contrastive semi- supervised learning for underwater image restoration via reliable bank,

    S. Huang, K. Wang, H. Liu, J. Chen, and Y. Li, “Contrastive semi- supervised learning for underwater image restoration via reliable bank,” in CVPR, 2023

    2023

  16. [24]

    Histoformer: Histogram-based transformer for efficient underwater image en- hancement,

    Y.-T. Peng, Y.-R. Chen, G.-R. Chen, and C.-J. Liao, “Histoformer: Histogram-based transformer for efficient underwater image en- hancement,” IEEE Journal of Oceanic Engineering, 2024

    2024

  17. [25]

    Phaseformer: Phase-based attention mechanism for underwater image restoration and beyond,

    M. Khan, A. Negi, A. Kulkarni, S. S. Phutke, S. K. Vipparthi, and S. Murala, “Phaseformer: Phase-based attention mechanism for underwater image restoration and beyond,” arXiv preprint arXiv:2412.01456, 2024

    2024 arXiv

  18. [26]

    Wavelength-based attributed deep neural network for underwater image restoration,

    P . Sharma, I. Bisht, and A. Sur, “Wavelength-based attributed deep neural network for underwater image restoration,” ACM TOMM, 2023

    2023

  19. [27]

    Osmosis: Rgbd diffusion prior for underwater image restoration,

    O. B. Nathan, D. Levy, T. Treibitz, and D. Rosenbaum, “Osmosis: Rgbd diffusion prior for underwater image restoration,” in ECCV, 2024

    2024

  20. [28]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in ICCV, 2023

    2023

  21. [29]

    Sdedit: Guided image synthesis and editing with stochastic dif- ferential equations,

    C. Meng, Y. He, Y. Song, J. Song, J. Wu, J.-Y. Zhu, and S. Ermon, “Sdedit: Guided image synthesis and editing with stochastic dif- ferential equations,” arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  22. [30]

    Breathing new life into 3d assets with generative repainting,

    T. Wang, M. Kanakis, K. Schindler, L. Van Gool, and A. Obukhov, “Breathing new life into 3d assets with generative repainting,” in BMVC, 2023

    2023

  23. [31]

    Dginstyle: Domain-generalizable semantic segmen- tation with image diffusion models and stylized semantic control,

    Y. Jia, L. Hoyer, S. Huang, T. Wang, L. Van Gool, K. Schindler, and A. Obukhov, “Dginstyle: Domain-generalizable semantic segmen- tation with image diffusion models and stylized semantic control,” in ECCV, 2024

    2024

  24. [32]

    Consistencyˆ 2: Consis- tent and fast 3d painting with latent consistency models,

    T. Wang, A. Obukhov, and K. Schindler, “Consistencyˆ 2: Consis- tent and fast 3d painting with latent consistency models,” arXiv preprint arXiv:2406.11202, 2024

    2024 arXiv

  25. [33]

    Parametric shadow control for portrait generation in text-to-image diffusion models,

    H. Cai, T.-W. Huang, S. Gehlot, B. Y. Feng, S. Shah, G.-M. Su, and C. Metzler, “Parametric shadow control for portrait generation in text-to-image diffusion models,” arXiv preprint arXiv:2503.21943 , 2025

    2025 arXiv

  26. [34]

    Stablenormal: Reducing diffusion variance for stable and sharp normal,

    C. Ye, L. Qiu, X. Gu, Q. Zuo, Y. Wu, Z. Dong, L. Bo, Y. Xiu, and X. Han, “Stablenormal: Reducing diffusion variance for stable and sharp normal,” ACM Transactions on Graphics, 2024

    2024

  27. [35]

    Deblurring via stochastic refinement,

    J. Whang, M. Delbracio, H. Talebi, C. Saharia, A. G. Dimakis, and P . Milanfar, “Deblurring via stochastic refinement,” inCVPR, 2022

    2022

  28. [36]

    The power of context: How multimodality im- proves image super-resolution,

    K. Mei, H. Talebi, M. Ardakani, V . M. Patel, P . Milanfar, and M. Delbracio, “The power of context: How multimodality im- proves image super-resolution,” in CVPR, 2025

    2025

  29. [37]

    Ima- genet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Ima- genet: A large-scale hierarchical image database,” in CVPR, 2009

    2009

  30. [38]

    Diffusion models beat gans on image synthesis,

    P . Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” in NeurIPS, 2021

    2021

  31. [39]

    Computer modeling and the design of optimal under- water imaging systems,

    J. S. Jaffe, “Computer modeling and the design of optimal under- water imaging systems,” IEEE journal of oceanic engineering, vol. 15, no. 2, pp. 101–111, 1990. 10

    1990

  32. [40]

    U-net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolutional net- works for biomedical image segmentation,” inMICCAI, N. Navab, J. Hornegger, W. M. W. III, and A. F. Frangi, Eds., 2015

    2015

  33. [41]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in ICLR, 2021

    2021

  34. [42]

    Score-based generative modeling through stochastic differential equations,

    Y. Song, J. Sohl-Dickstein, D. P . Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in ICLR, 2021

    2021

  35. [43]

    Auto-encoding variational bayes,

    D. P . Kingma and M. Welling, “Auto-encoding variational bayes,” in ICLR, 2014

    2014

  36. [44]

    Latent consistency models: Synthesizing high-resolution images with few-step infer- ence,

    S. Luo, Y. Tan, L. Huang, J. Li, and H. Zhao, “Latent consistency models: Synthesizing high-resolution images with few-step infer- ence,” arXiv preprint arXiv:2310.04378, 2023

    2023 arXiv

  37. [45]

    Consistency models,

    Y. Song, P . Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” in NeurIPS, 2023

    2023

  38. [46]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in ICLR, 2022

    2022

  39. [47]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” in NeurIPS, 2022

    2022

  40. [48]

    Depth pro: Sharp monocular metric depth in less than a second,

    A. Bochkovskii, A. Delaunoy, H. Germain, M. Santos, Y. Zhou, S. R. Richter, and V . Koltun, “Depth pro: Sharp monocular metric depth in less than a second,” arXiv preprint arXiv:2410.02073, 2024

    2024 arXiv

  41. [49]

    Inherent optical properties of jerlov water types,

    M. G. Solonenko and C. D. Mobley, “Inherent optical properties of jerlov water types,” Appl. Opt., no. 17, pp. 5392–5401, 2015

    2015

  42. [50]

    A rapid scene depth estimation model based on underwater light atten- uation prior for underwater image restoration,

    W. Song, Y. Wang, D. Huang, and D. Tjondronegoro, “A rapid scene depth estimation model based on underwater light atten- uation prior for underwater image restoration,” in Advances in Multimedia Information Processing – PCM 2018 , 2018

    2018

  43. [51]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P . Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,” The international journal of robotics research, 2013

    2013

  44. [52]

    Usod10k: a new benchmark dataset for underwater salient object detection,

    L. Hong, X. Wang, G. Zhang, and M. Zhao, “Usod10k: a new benchmark dataset for underwater salient object detection,” IEEE transactions on image processing, 2023

    2023

  45. [53]

    Unsupervised underwater image restoration: From a homology perspective,

    Z. Fu, H. Lin, Y. Yang, S. Chai, L. Sun, Y. Huang, and X. Ding, “Unsupervised underwater image restoration: From a homology perspective,” in AAAI, 2022

    2022

  46. [54]

    Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement,

    W. Zhang, P . Zhuang, H.-H. Sun, G. Li, S. Kwong, and C. Li, “Underwater image enhancement via minimal color loss and locally adaptive contrast enhancement,”IEEE Transactions on Image Processing, vol. 31, pp. 3997–4010, 2022

    2022

  47. [55]

    Human-visual-system- inspired underwater image quality measures,

    K. Panetta, C. Gao, and S. Agaian, “Human-visual-system- inspired underwater image quality measures,” IEEE Journal of Oceanic Engineering, vol. 41, no. 3, pp. 541–551, 2015

    2015

  48. [56]

    Musiq: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y. Wang, P . Milanfar, and F. Yang, “Musiq: Multi- scale image quality transformer,” in ICCV, 2021, pp. 5148–5157

    2021

  49. [57]

    Indoor segmentation and support inference from rgbd images,

    P . K. Nathan Silberman, Derek Hoiem and R. Fergus, “Indoor segmentation and support inference from rgbd images,” in ECCV, 2012

    2012

  50. [58]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in CVPR Workshops, 2017

    2017

  51. [59]

    From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions,

    P . Young, A. Lai, M. Hodosh, and J. Hockenmaier, “From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions,” Transactions of the association for computational linguistics, vol. 2, pp. 67–78, 2014

    2014

  52. [60]

    Scene parsing through ade20k dataset,

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene parsing through ade20k dataset,” in CVPR, 2017

    2017

  53. [61]

    Learning-based inverse rendering of com- plex indoor scenes with differentiable monte carlo raytracing,

    J. Zhu, F. Luan, Y. Huo, Z. Lin, Z. Zhong, D. Xi, R. Wang, H. Bao, J. Zheng, and R. Tang, “Learning-based inverse rendering of com- plex indoor scenes with differentiable monte carlo raytracing,” in SIGGRAPH Asia, 2022

    2022

  54. [62]

    Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation,

    C. Sakaridis, D. Dai, and L. Van Gool, “Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation,” IEEE TP AMI, 2020

    2020

  55. [63]

    Marine video kit: A new marine video dataset for content-based analysis and retrieval,

    Q.-T. Truong, T.-A. Vu, T.-S. Ha, J. Loko ˇc, Y. H. W. Tim, A. Joneja, and S.-K. Yeung, “Marine video kit: A new marine video dataset for content-based analysis and retrieval,” in MultiMedia Modeling, 2023. Jiayi Wuis a Ph.D. student at the Perception and Robotics Group of the...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.