BenthicFlow: Generating Extensible Underwater Environments via Flow Matching

Computer vision and marine robotics suffer from a severe shortage of high-quality 3D underwater data because surface-trained models fail to generalize to aquatic scenes, and traditional structure-from-motion pipelines frequently break down in low-texture, turbid environments. Existing generative approaches often stitch independently generated tiles post hoc using separate models, leading to boundary artifacts and limited global coherence across geographically distinct sites. This paper introduces BenthicFlow, a unified framework designed to synthesize large-scale, spatially extensible, and geographically diverse 3D underwater environments directly from survey images.

BenthicFlow operates within a compressed latent space managed by a multi-modal autoencoder called d-RAE, which jointly encodes RGB images and depth maps into aligned latent tokens. Instead of relying on traditional denoising diffusion models or separate post-hoc in-painting networks, BenthicFlow employs a conditional flow-matching model trained to regress a velocity field over small latent windows. To achieve unbounded spatial extent, the authors adapt a MultiDiffusion-inspired sampling strategy that processes overlapping windows concurrently during the generative trajectory. By averaging the predicted velocities of overlapping tokens at every integration step, the model naturally reconciles window boundaries and maintains global semantic consistency under a shared conditioning framework.

Once the large latent canvas is fully denoised, the d-RAE decoder maps the latents back into a continuous RGBD mosaic of arbitrary size. This 2.5D observation is subsequently converted into an explicit 3D benthic environment by unprojecting pixels into point clouds and representing them as surface-aligned Gaussian surfels. Rather than using isotropic volumetric blobs, each surfel is oriented according to the local surface normal, with in-plane extents adjusted anisotropically based on inter-pixel spacing to prevent gaps at depth discontinuities. This allows BenthicFlow to render coherent, large-scale novel views that preserve site-specific visual appearances across different marine survey regions.