cs.CVOct 21, 2026

Rethinking Underwater Image Enhancement: Learning through Text-Guided Degradation Curriculum

Authors: Hongri Tian, Wenxue Li, Tian Ye, Sixiang Chen, Zongxian Su, Lujia Wang, Xianying Wang, Xinhu Zheng, 

Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · China Geological Survey, Beijing, China. · Director of Applied Science, Amazon.com, Inc., U.S.. · The Hong Kong University of Science and Technology, Hong Kong SAR, China.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Existing paired clean-degraded datasets have advanced learning-based underwater image enhancement, but they remain either limited in scale or difficult to acquire, and still fail to adequately cover the diverse degradations encountered in real-world underwater environments. Such scarce and insufficient data ultimately restrict the generalization of underwater image enhancement models, highlighting the need for realistic and controllable degradation modeling to better capture real underwater conditions and improve enhancement robustness. To this end, we propose a unified degradation-aware training framework for the underwater image enhancement task. A Text-Conditioned Degradation Augmentation Pipeline uses a text-guided diffusion model to synthesize diverse and semantically consistent degraded samples, providing paired data to support the subsequent curriculum learning strategy. The proposed Progressive Degradation Curriculum organizes training across multiple attributes and difficulty levels to promote balanced learning, and a Transmission-Guided One-Step Diffusion Model leverages this progressive process to better address the wide range of degradation severity and restoration ambiguity in underwater images. Extensive experiments on multiple underwater benchmarks demonstrate that our proposed method achieves superior performance. Our code and models will be released upon the publication of this work.

Figures & tables

Figure 1: Top : Illustration of three common routes for constructing paired underwater datasets. Medium : Proposed text-driven degradation diffusion paradigm for controllable underwater degradation synthesis. Bottom : Quantitative results of our method on performance improvement and synthesis quality.
Figure 2: Overview of the proposed Text-Conditioned Degradation Augmentation (TCDA) pipeline. It consists of four stages: (1) Degradation Captioning via MLLMs for generating fine-grained textual descriptions of underwater degradations, (2) Flow-Based Degradation Modeling for learning text-conditioned mappings between clean and degraded representations, (3) Prompt-Space Augmentation for enriching degradation prompts with varying difficulty levels, and (4) Text-Conditioned Degradation Synthesis for controllable degradation generation to support curriculum learning.
Figure 3: Overview of the proposed UIE framework. The model learns progressive degradation-aware enhancement through curriculum sampling and transmission-guided one-step diffusion, combining physical priors and text-conditioned cues for controllable restoration with greater inference efficiency than multi-step diffusion methods.
Figure 4: Examples of learned degradation generation results. Given a clean image and a textual prompt explicitly describing degradation attributes (e.g., turbidity, color cast, lighting, haze), our model synthesizes degraded counterparts that accurately reflect the specified visual conditions and severity levels.
MethodsUIEBC60U45
SSIM ↑\uparrowPSNR ↑\uparrowLPIPS ↓\downarrowUCIQE ↑\uparrowUIQM ↑\uparrowUCIQE ↑\uparrowUIQM ↑\uparrowUCIQE ↑\uparrowUIQM ↑\uparrow
UIEC 2 - Net Wang et al. (2021)0.88017.9610.2520.5263.7660.4872.1490.5354.231
+ TCDA0.880 (+0.000)19.020 (+1.059)0.229 (-0.023)0.550 (+0.024)3.695 (-0.071)0.525 (+0.038)2.146 (-0.003)0.554 (+0.019)4.156 (-0.075)
SCNet Fu et al. (2022)0.90719.5600.1950.5853.8690.5662.4980.5704.155
+ TCDA0.909 (+0.002)20.998 (+1.438)0.172 (-0.023)0.593 (+0.008)3.925 (+0.056)0.551 (-0.015)2.452 (-0.046)0.578 (+0.008)4.163 (+0.008)
FiveA+ Jiang et al. (2023)0.90318.6710.2600.5583.8110.5312.2290.5464.186
Table 1: Quantitative comparison on UIEB, C60, and U45. Each method is evaluated with its original training data and with TCDA augmentation. Parentheses show the performance change. Best results are in bold.
Figure 5: Visual comparisons with state-of-the-art UIE methods on the UIEB dataset. Our method consistently outperforms state-of-the-art UIE methods for underwater images with varying degrees of degradation.
Backbone (Single-step) Encoder-lora (Trainable) Skip Conn. UDCP SSIM↑\text{SSIM}\uparrow PSNR↑\text{PSNR}\uparrow LPIPS ↓\downarrow ✓ 0.697 20.374 0.226 ✓ ✓ 0.701 20.516 0.228 ✓ ✓ 0.879 21.266 0.153 ✓ ✓ 0.700 20.436 0.233 ✓ ✓ ✓ 0.912 21.767 0.139 ✓ ✓ ✓ ✓ 0.917 21.723 0.135
Table 2: Ablation study on the key components of TG-Diff.
Figure 6: Qualitative ablation study of the proposed TG-Diff components on the UIEB dataset. The full model restores clearer structure, more natural illumination, and balanced color rendition.
Methods SSIM ↑\uparrow PSNR ↑\uparrow LPIPS ↓\downarrow UCIQE ↑\uparrow UIEC 2 -Net Wang et al. (2021) 0.857 20.570 0.255 0.552 SCNet Fu et al. (2022) 0.872 20.737 0.215 0.584 FiveA+ Jiang et al. (2023) 0.876 21.655 0.241 0.570 WF-Diff Zhao et al. (2024) 0.895 23.346 0.165 0.590 CE-VAE Pucci and Martinel (2025) 0.851 25.535 0.190 0.578 Phaseformer Khan et al. (2025) 0.891 25.483 0.135 0.579 TG-Diff (Ours) 0.904 25.734 0.130 0.592
Table 3: Quantitative comparison on LSUI Peng et al. (2023) . Best results are in bold.
Figure 7: Visualization of degradation generation across different difficulty levels. Conditioned on text prompts describing easy, moderate, and hard degradations, the model generates visually consistent degradation patterns with increasing severity.
Methods SSIM ↑\uparrow PSNR ↑\uparrow LPIPS ↓\downarrow UCIQE ↑\uparrow TG-Diff + Easy only 0.901 21.474 0.149 0.603 TG-Diff + Moderate only 0.905 21.392 0.147 0.608 TG-Diff + Hard only 0.907 21.404 0.150 0.603 TG-Diff + Random Sample 0.925 22.248 0.129 0.608 TG-Diff + Curriculum 0.927 22.341 0.129 0.613
Table 4: Ablation of degradation-difficulty sampling. Curriculum learning yields the best balanced performance.
MethodSSIM ↑\uparrowPSNR ↑\uparrowLPIPS ↓\downarrowUCIQE ↑\uparrow
w/ UWCNN0.85519.4910.1850.580
w/ TCDA (Ours)0.90521.3920.1470.608
Table 6: Comparison between UWCNN-based augmentation and TCDA (moderate only).
Figure 8: Trends of non-reference underwater image quality metrics under different degradation difficulty level settings. Both UCIQE and UIQM consistently decrease as the degradation difficulty increases from easy to hard.
More in this topic

Underwater Image Enhancement

Underwater image enhancement restores color, contrast, visibility, and detail degraded by water absorption, scattering, haze, and artificial light.

See all 37 papers

Explore similar work

CardsList

  1. RPL-UIE: Reliable Prior Learning for Underwater Image Enhancement

    Jul 31, 2026Yifan Chen, Jiaming Liu, Ye Zheng +2Diffusion Models for Image RestorationUnderwater Image Enhancement

    College of Future Information Technology, Fudan University, Shanghai 200433, China · Institute of Artificial Intelligence (TeleAI), China Telecom, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China +1

  2. Degradation-Guided Underwater Image Restoration with Task-Oriented Latent Control

    Aug 9, 2026Xu Zhang, Xuhui Cao, Kangzhe Yuan +5Underwater Image EnhancementUnderwater Imaging

    National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · School of Computer and Artificial Intelligence, Zhengzhou University · Aerospace Information Research Institute, Henan Academy of Sciences +2

  3. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM

    Mar 13, 2026Guodong Fan, Shengning Zhou, Genji Yuan +3Geometry-Guided Image RestorationImage Enhancement

    Shandong Technology and Business University, Yantai, China · Shandong University of Finance and Economics, Jinan, China · Dalian Maritime University, Dalian, China