<p>Underwater depth estimation is critical in underwater exploration, providing ranging information at a low cost. However, it remains inadequate due to physical constraints, which leads to data scarcity and the absence of high-quality benchmarks. Given the inherent challenges of light attenuation and backscatter in water, acquiring clear images or precise depth is notably difficult. To mitigate this issue, existing methods prefer the self- or unsupervised paradigms while the performance lags due to domain gaps and looser constraints. In this paper, we propose a simple yet effective solution Atlantis++, to enable the supervised underwater depth estimation. Specifically, we propose to create vivid non-existent underwater scenes with terrestrial depth, through the innovative generative diffusion models. We introduce a specialized Depth2Underwater ControlNet by training on prepared {Underwater, Depth, Text} data triplets, to flexibly accommodate underwater photorealism. Our method enables terrestrial depth estimation models to achieve considerable improvements on unseen underwater scenes, surpassing their terrestrial pretrained counterparts both quantitatively and qualitatively. We also show that the improvements can help downstream applications such as underwater image enhancement and Visual SLAM. Moreover, we present a large-scale high-quality underwater depth estimation benchmark, featuring controlled turbidity levels and color casts. We conduct comprehensive evaluations of existing monocular depth estimation methods to have a better understanding of the unique challenges in underwater depth estimation.</p>