acceptodds
Under review as a conference paper at ICLR 2027

Physical Degradation-Aware Generative Prior Optimization for Zero-Shot Underwater Image Restoration

Abstract

Single-image underwater restoration is ill-posed: wavelength-dependent attenuation and scattering produce spatially varying structural and chromatic degradation. Learned underwater restoration methods typically depend on training across underwater image collections, whereas generic generative priors provide no explicit guidance for such nonuniform degradation. We propose Physical Degradation-Aware Generative Prior Optimization (PD-GPO), a zero-shot framework that uses a spatial and a chromatic degradation field to guide image-specific latent optimization under a frozen generative prior. A physical degradation field estimator (PDFE), trained only on synthetic degradations of natural RGB-D images, remains frozen and estimates both fields from each generated candidate. The spatial field measures the local degree of structural degradation; it guides contrast initialization and selects the support on which the chromatic field regularizes channel ratios. A frozen DINOv2 feature-consistency loss limits generative drift. We also introduce MIUD, an 850-image multi-region in-situ underwater benchmark spanning depths from several meters to over 4,000 m in the Beibu Gulf, the South China Sea, and the Eastern Indian Ocean. Extensive experiments demonstrate that PD-GPO outperforms existing zero-shot methods on most evaluated metrics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-shot underwater image restoration, Spatial–chromatic degradation fields, Multi-region in-situ underwater dataset.

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 2 reviews, above 27% of the others.
Reviewer_6ZzDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Physical Degradation-Aware Generative Prior Optimization (PD-GPO), a zero-shot approach for underwater image restoration. It estimates two degradation fields: a spatial field describing contrast loss and a chromatic field describing color attenuation. These fields are predicted by a network trained on synthetic data. The spatial field guides initial contrast enhancement, after which the image latent is optimized using a frozen FLUX.2 generative model. During optimization, the network repeatedly re-estimates the degradation fields to guide color correction, while a DINOv2 feature loss helps preserve scene content.

The main contribution is combining spatial and chromatic degradation guidance with generative optimization through iterative field re-estimation. The authors also introduce a new dataset of 850 underwater images. Experiments on three public benchmarks and a subset of the dataset show strong results compared with other zero-shot methods. Ablations support the benefits of guided contrast enhancement and selective color correction, but do not establish the advantage of repeatedly updating the degradation fields.

Critical strengths
  • The paper provides a coherent framework in which two complementary degradation fields are used for different restoration operations. This is better motivated than applying a generic global enhancement loss to a generative model.

  • The method shows consistent improvements over competing zero-shot approaches across multiple benchmarks, including strong gains in PSNR and LPIPS. The ablations support the benefits of spatially guided initialization and selective chromatic correction. The authors also introduce MIUD, a dataset of 850 underwater images collected across different ocean regions, scene types, and depths, to evaluate performance under more diverse real-world conditions.

Critical weaknesses
  • The approach builds largely on existing methods: generative latent optimization, AutoACE, wavelet-based features, DINOv2 perceptual constraints, and physical degradation estimation. The main novelty lies in combining these components and repeatedly estimating degradation fields during optimization. However, the paper does not demonstrate a clear advantage over simpler alternatives using fixed fields, contrast- or depth-based spatial masks, or conventional color balancing. Thus, the central methodological contribution is not sufficiently isolated experimentally.

  • The physical degradation estimator is trained on synthetic data generated using the authors' model. Its high validation accuracy shows that it can recover fields produced by this model, but not necessarily that these fields reflect real underwater degradation. Furthermore, the chromatic loss encourages balanced channel ratios, which does not necessarily correspond to recovering true underwater radiance. Naturally colored scenes, varying illumination, and severe attenuation may make this constraint unreliable. Since the physical interpretation is central to the method, its validity on real underwater images needs stronger evidence.

Questions for authors
  • My main question concerns the dynamic field estimation. How does repeatedly re-estimating the degradation fields during optimization compare with estimating them once from the input and keeping them fixed? A controlled ablation would help clarify the benefit of this iterative update.

  • How important is the accuracy of the physical degradation estimates? Could similar improvements be achieved with simpler local contrast measures for spatial guidance and conventional selective color balancing, rather than learned degradation fields?

  • How well does the method generalize to different underwater conditions? Does PD-GPO maintain its advantage over underwater-trained methods when water type, illumination, scattering, or depth changes? Results broken down by degradation condition, ideally on the full MIUD dataset, would help clarify this.

Reviewer_VLccOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PD-GPO, a zero-shot underwater image restoration framework that integrates spatial-chromatic degradation estimation with a frozen generative prior. A degradation field estimator trained on synthetic RGB-D data guides contrast initialization and chromatic regularization during latent optimization, while DINOv2 features preserve image structure and semantics. The authors also introduce an 850-image underwater dataset (MIUD) and report competitive results against existing zero-shot methods.

Critical strengths
  1. The paper addresses the challenging problem of zero-shot underwater image restoration without relying on paired or unpaired real underwater training datasets.
  2. The combination of spatial and chromatic degradation estimation provides an intuitive way to incorporate degradation-aware guidance into generative image restoration.
  3. The evaluation includes multiple public underwater datasets and an additional independently collected dataset, with comparisons against several relevant zero-shot approaches.
Critical weaknesses
  1. Limited methodological novelty beyond existing generative prior optimization frameworks. The proposed framework largely builds upon existing generative prior optimization methods, particularly DIIP, while incorporating established components such as FLUX, DINOv2, and AutoACE. The spatial–chromatic guidance introduces only incremental modifications.
  2. The claimed physical degradation awareness is not supported by a rigorous physical reconstruction constraint. The physical degradation model is used only to generate synthetic training data rather than enforcing reconstruction consistency during restoration. Moreover, the estimated degradation fields are quantitatively validated only on synthetic data, leaving their physical reliability in real underwater scenes unclear.
  3. The chromatic regularization objective relies on a questionable assumption. The chromatic loss encourages aggregated channel ratios to approach unity, but this assumption lacks sufficient theoretical justification. Such a constraint may introduce inaccurate color compensation in scenes with nonuniform color distributions or severe wavelength-dependent attenuation.
  4. High computational complexity. The method requires 1,000 latent optimization iterations and approximately 181 seconds per 512 × 512 image on an RTX 3090. This computational burden limits its practical applicability, while no convergence analysis or iteration-efficiency study is provided.
Questions for authors

Please refer to the critical weaknesses outlined above for further details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.