SLURPP: Single-Step Underwater Restoration

• SLURPP is a physically grounded method for underwater restoration that jointly recovers clear images and dense water-medium effects using a reformulated image formation model.
• The approach leverages a dual-branch latent diffusion architecture with inter-branch cross-attention to combine pretrained scene and depth priors effectively.
• Empirical results show that SLURPP achieves state-of-the-art performance with a PSNR of 25.66 dB at over 200× faster runtime than prior diffusion-based methods.

Searching arXiv for the specific SLURPP paper and closely related underwater restoration work. Search query: "SLURPP underwater restoration latent diffusion" SLURPP, short for Single-step Latent Underwater Restoration with Pretrained Priors, is a method for underwater image restoration that jointly estimates a clear scene image and dense water-medium effects from a degraded underwater observation. It is motivated by the view that underwater image formation is governed not only by generic image degradation but by the interaction of scene radiance with depth- and wavelength-dependent attenuation and backscatter. In that formulation, SLURPP combines pretrained latent diffusion priors, an explicit scene-medium decomposition, and a single-forward-pass restoration pipeline, and is reported to achieve state-of-the-art performance on both synthetic and real-world benchmarks while being over $200\times$ faster than prior diffusion-based methods (Wu et al., 10 Jul 2025).

1. Physical formulation and problem setting

SLURPP is built around the revised Jaffe–McGlamery style underwater image formation model

$$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$

where $c \in \{R,G,B\}$, $I$ is the observed underwater image, $J$ is the clear image, $z$ is scene distance or depth, $B^\infty$ is the background light, and $\beta^D$ and $\beta^B$ are attenuation and backscatter coefficients. The paper rewrites this as

$I_c = J_c \cdot T_c + B_c$

with

$$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$0

This reformulation makes the restoration target a dense decomposition into a clear image $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$1, transmission $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$2, and backscattering $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$3 (Wu et al., 10 Jul 2025).

A central design choice is that SLURPP predicts two dense 3-channel images, $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$4 and $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$5, rather than estimating scalar water parameters or explicit depth as the primary restoration target. The stated reasons are that diffusion backbones are naturally suited to dense image-like outputs, that this avoids simplifying assumptions such as homogeneous media or $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$6, and that dense medium maps are bounded image-like quantities that are more stable than regressing raw depth over very large ranges. This suggests that the method is intended as a physically grounded restoration system rather than a generic image-to-image translator.

The decomposition also encodes the paper’s interpretation of underwater degradation: direct scene radiance is attenuated exponentially with depth, while an additive veiling component grows with depth. In that sense, SLURPP treats underwater restoration as the joint recovery of scene content and water-medium structure, not merely the inversion of a color cast.

2. Architecture and single-step latent diffusion design

SLURPP operates in latent space. An underwater RGB image is first encoded by the frozen VAE encoder from Stable Diffusion v2, after which the model branches into a scene branch that predicts the latent for the clear image $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$7, and a medium branch that predicts latents for transmission $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$8 and backscattering $$I_c = J_c \cdot e^{-\beta^D_c z} + B_c^\infty \cdot (1 - e^{-\beta^B_c z}),$$9. The architecture is described as a dual-branch architecture with inter-branch cross-attention, so that scene and medium estimates can inform one another (Wu et al., 10 Jul 2025).

The branch initialization is task-aligned. The scene branch is initialized from Stable Diffusion v2 to exploit natural-image priors, while the medium branch is initialized from Marigold, a pretrained diffusion-based monocular depth model, to exploit geometry and depth priors. After latent prediction, the $c \in \{R,G,B\}$0 and $c \in \{R,G,B\}$1 latents are decoded using the standard Stable Diffusion decoder, whereas the clear scene latent $c \in \{R,G,B\}$2 is decoded using a cross-latent decoder.

The “single-step” aspect refers to how the latent diffusion backbone is repurposed. Rather than adding Gaussian noise to output latents and training a UNet to predict noise across many diffusion steps, SLURPP feeds the diffusion UNet the input latent concatenated with a zero image, interpreted as the mean of the pure Gaussian noise distribution, and trains the network to directly predict the output latent image. This removes iterative denoising at inference time and permits direct supervision in decoded RGB space.

The paper explicitly argues that underwater restoration is a narrower conditional prediction problem than text-to-image generation, so a one-step mapping is sufficient and can outperform iterative denoising. That claim is operational rather than purely conceptual: the reported one-step model exceeds the authors’ own iterative latent-loss baseline, including a version using 50 denoising steps. A plausible implication is that the pretrained latent prior is being used more as a structured conditional feature extractor than as a stochastic sampler.

Two-stage training is used. In stage 1, the two UNet branches and their cross-attention are trained to predict scene and medium latents from the underwater input while the VAE encoder and decoder remain frozen. In stage 2, only the decoder and the added zero-convolution skip connections of the cross-latent decoder are fine-tuned to preserve high-frequency details such as text and facial structure (Wu et al., 10 Jul 2025).

3. Objectives, decomposition consistency, and synthetic supervision

The basic image-space objective applied after decoding is

$c \in \{R,G,B\}$3

The total training loss is

$c \in \{R,G,B\}$4

with

$c \in \{R,G,B\}$5

The additional $c \in \{R,G,B\}$6 term is a self-supervised reconstruction loss obtained by recombining the predicted outputs according to

$c \in \{R,G,B\}$7

The same image loss is applied between the recomposed image and the input underwater image (Wu et al., 10 Jul 2025).

This reconstruction term is used to enforce joint physical consistency between predicted scene and medium components. The paper states that it helps bridge the synthetic-to-real domain gap because even if synthetic labels are imperfect, the outputs are constrained by the underwater formation model. In the reported ablation, removing this reconstruction loss lowers synthetic PSNR from 25.66 dB to 24.80 dB, a difference of 0.86 dB.

Because paired real underwater and clear data are scarce, SLURPP is trained with a synthetic data generation pipeline based on terrestrial image datasets, specifically DIV2K, Flickr, ADE20K, InteriorVerse, and Dark Zurich. The formation model requires metric depth, attenuation coefficients, and background light. To obtain these, the pipeline uses Depth Pro for metric monocular depth, samples attenuation coefficients according to real-world optical measurements of Jerlov water types at 600 nm, 525 nm, and 475 nm, and estimates background light using ULAP, with ULAP’s depth component replaced by Depth Pro. The extracted background-light estimates are clustered by K-means ($c \in \{R,G,B\}$8) in Lab space using the $c \in \{R,G,B\}$9 channels. The resulting synthetic data provide supervision for $I$0, $I$1, $I$2, $I$3, and $I$4 (Wu et al., 10 Jul 2025).

The synthetic pipeline is itself part of the method’s contribution. The supplementary ablation reports that training on the Osmosis terrestrial data with sampled real-world water values yields 25.76 dB, whereas using random water values yields 24.87 dB. This suggests that realistic water-parameter sampling contributes materially to restoration quality, although the architecture remains stronger than the compared prior method even under weaker data conditions.

4. Empirical performance and benchmark results

SLURPP is evaluated on two real-world underwater image benchmarks, USOD10K and UIEB, using the no-reference metrics UIQM and MUSIQ, and on a synthetic benchmark derived from NYUv2 with simulated underwater degradation. On the real datasets, the paper reports the best values among the listed methods for both datasets; on the synthetic benchmark, it reports the best PSNR and SSIM and an LPIPS of 0.05 (Wu et al., 10 Jul 2025).

Benchmark	Metrics	SLURPP
USOD10K	UIQM / MUSIQ	3.152 / 70.110
UIEB	UIQM / MUSIQ	3.340 / 72.457
Synthetic NYUv2-based benchmark	PSNR / SSIM / LPIPS	25.66 / 0.95 / 0.05

On the synthetic benchmark, the most direct comparison is to Osmosis, which is reported at 22.74 dB PSNR, 0.89 SSIM, and 0.06 LPIPS, versus 25.66 dB, 0.95, and 0.05 for SLURPP. This is the source of the paper’s claim of approximately $I$5 dB PSNR improvement. The same comparison frames the method’s runtime claim: Osmosis requires more than 200 seconds per image on an A6000 GPU and is limited to $I$6, whereas SLURPP requires 0.75 seconds for a $I$7 image and can restore up to $I$8.

The paper also reports medium-prediction accuracy on synthetic NYU data. For transmission, Osmosis obtains 13.97 PSNR and 0.207 MAE, while SLURPP obtains 24.69 PSNR and 0.060 MAE. For backscatter, Osmosis obtains 23.08 PSNR and 0.076 MAE, while SLURPP obtains 32.37 PSNR and 0.024 MAE. These numbers are presented as evidence that direct dense medium prediction can be more accurate than a depth-first restoration strategy.

Qualitatively, the paper states that SLURPP better handles distant and heavily degraded regions, avoids red-shift artifacts seen in baselines, and produces medium maps whose spatial variation aligns with scene depth. A plausible implication is that the decomposition $I$9 provides a stronger inductive bias for scenes with large depth variation than black-box restoration objectives.

5. Ablations, comparative analysis, and distinctive features

Several ablations isolate the contribution of SLURPP’s components. In the supplementary table, an iterative latent loss (50 steps) baseline reaches 23.46 PSNR, 0.89 SSIM, and 0.08 LPIPS; adding the cross-latent decoder raises this to 24.62, 0.93, and 0.05; the full single-step SLURPP reaches 25.66, 0.95, and 0.05. The paper therefore argues that the one-step design is not only faster but also better than the iterative latent-loss alternative (Wu et al., 10 Jul 2025).

The cross-latent decoder contributes an additional 0.67 dB on the synthetic benchmark: 24.99 dB without it versus 25.66 dB with it. Its stated purpose is to recover high-frequency detail that the vanilla Stable Diffusion decoder may blur or hallucinate. The decoder uses zero-convolution skip connections that transfer high-frequency information from the underwater input through the encoder to the clear-output decoder.

The paper also compares the full system to variants without cross-attention and without the cross-latent decoder. The reported values are 25.06 dB for Ours (w/o cross-attention), 24.99 dB for Ours (w/o cross-latent decoder), and 25.66 dB for the full model. This pattern is used to support the claim that both scene-medium interaction and decoder refinement contribute to final quality.

Relative to prior diffusion-based underwater restoration, the paper highlights five distinguishing properties: latent-space rather than pixel-space operation, one-step rather than iterative denoising, explicit scene-medium decomposition rather than depth-first restoration, direct image-space supervision, and higher practical resolution. It also distinguishes SLURPP from single-step distillation approaches such as progressive distillation, consistency models, and latent consistency models by noting that SLURPP is trained directly for one-step prediction from the start, rather than distilled from a multi-step teacher.

6. Limitations, scope, and position within underwater restoration

The paper is explicit about several limitations. First, SLURPP is not real-time, even though it is substantially faster than previous diffusion-based methods. Second, it is a single-image method and includes no temporal consistency mechanism, so underwater video applications can exhibit frame-to-frame flicker. Third, the synthetic training data still simplify real underwater physics and do not fully capture effects such as non-uniform media and mismatched scattering/attenuation assumptions. Fourth, turbidity is not explicitly modeled, and severe turbidity can still lead to blur. Finally, the method assumes that the dense decomposition

$J$0

is sufficiently expressive for most degradations, although the paper notes that this may not capture all complex underwater effects (Wu et al., 10 Jul 2025).

Within the underwater restoration literature, SLURPP is best understood as a physically grounded latent-restoration framework that repurposes pretrained priors for both scene appearance and medium geometry. Its novelty does not reduce to “using diffusion for underwater images.” Rather, the method combines latent-space prediction, direct one-step inference, explicit decomposition into $J$1, $J$2, and $J$3, task-aligned pretrained initialization, and a synthetic pipeline with dense degradation annotations. This suggests a broader methodological position: strong pretrained image priors can be made more effective for restoration when paired with explicit forward models and decomposition targets instead of purely implicit enhancement objectives.

The reported empirical outcome is correspondingly twofold. On the one hand, SLURPP improves quality on synthetic and real benchmarks, including a gain from 22.74 dB to 25.66 dB over the compared diffusion baseline on the synthetic benchmark. On the other hand, it reduces diffusion-based restoration runtime from more than 200 seconds to 0.75 seconds in the stated A6000 setting. In the paper’s framing, those results support the view that underwater restoration benefits from combining pretrained latent diffusion priors with an explicit model of attenuation and backscattering rather than treating restoration as an unconstrained image-enhancement problem.