UPhy-Diff: A Text-Guided Multimodal Physical Diffusion Model for Underwater Image Enhancement

Abstract

Multifaceted environmental degradation inherently obscures critical visual information in underwater imagery. Recently, most underwater image enhancement methods (UIE) neglect physical imaging and multi-modal information compensation. Simultaneously, a domain gap exists between reference images and real-world natural images. To address these limitations, a text-guided multimodal underwater physical diffusion framework (UPhy-Diff) is presented. Specifically, pre-trained foundation models are utilized to generate textual descriptions and relative depth maps, enabling the creation of UIE triplets and expanding diffusion models to RGBD generation tasks. Subsequently, a noise estimation network decoupled from the guiding image, termed Struct-UNet, is proposed. Beneficial features are captured based on the state space model (SSM). Ultimately, during the diffusion sampling stage, physical optimization is incorporated to achieve physically-aware fine-tuning. This compensates to a certain extent for the inter-domain gap caused by reference images and pre-trained large models. Thus, the model is capable of generating high-quality underwater images that align better with human visual perception. Comprehensive experimental evaluations confirm the efficacy and robustness of the proposed method across diverse color distortions, low-light conditions, and underwater haze degradation scenarios.