The underwater images always suffer from low contrast and color bias due to the scattering water medium. Such degradations impair the image quality and therefore have adversarial effects on the underwater visual-based tasks. In recent years, many learning-based underwater image enhancement (UIE) methods have been proposed and have achieved great progress. However, in this work, we show that although good visual quality has been reached, the enhanced images do not contribute as expected to downstream 3D tasks like 3D Gaussian splatting (3DGS) due to the quality instability of the train labels that impairs the multi-view consistency. To solve this problem, A diffusion-based UIE is proposed to support 3DGS-based novel views synthesis with high visual fidelity. This is achieved by a physics-based label generation from continuous video dataset, which enables the model to witness stable and physically-aligned air-water sequential pairs to preserve multi-view consistency. In addition, a frequency fusion strategy is also proposed to compensate for the reconstruction loss of latent-space diffusion learning, where the low-frequency part representing the correct light field and the high-frequency part preserving texture details are combined. The results in both public dataset and our field experiment demonstrate the superiority of the proposal, showing a significant improvement qualitatively and quantitatively.