ScoreDiff: Decoupled score decomposition for Physics-guided underwater image restoration

Underwater image restoration is a challenging, ill-posed problem due to severe physical degradation and information loss. While diffusion models provide strong generative priors, existing physics-guided approaches suffer from representational entanglement, where a shared network must simultaneously learn content generation and degradation modeling without explicit task separation. To address this, we present ScoreDiff, a unified framework based on decoupled score decomposition that separates semantic prior learning from physics-guided correction through dedicated optimization pathways. Specifically, an Asymmetric Dual-Branch Architecture with stop-gradient optimization structurally isolates the generative prior from physical consistency constraints, treating the physics branch as a residual refinement mechanism. To handle unreliable physical estimates in complex scenes (e.g., dynamic caustics), we introduce an Uncertainty-Aware Arbitration Module that treats inter-branch disagreement as a proxy for local guidance reliability, dynamically suppressing erroneous corrections where the optical model fails. Additionally, to bridge the resolution gap in latent diffusion, we propose a Content-Consistent Decoder that mitigates latent-space detail loss by injecting pixel-space structural features. Extensive experiments on paired and unpaired benchmarks demonstrate state-of-the-art performance, with particular robustness in challenging aquatic environments exhibiting non-uniform illumination and scattering.