Mini toolsLingBot Depth
By Robbyant
Model family: LingBot
LingBot-Depth is a ViT Large/14 based depth completion model trained via masked depth modeling: sensor depth inaccuracies are treated as masked signals reconstructed using RGB visual context. It ships in two variants: a general purpose pretrained model trained on 10M RGB-D samples for depth completion, refinement, and point cloud generation, and a post trained DC variant optimized for sparse depth completion from SfM or SLAM points, handling inputs with under 5 percent valid pixels. The architecture pairs a ViT Large/14 encoder with separated RGB and depth patch embeddings and a ConvStack decoder with hierarchical upsampling, about 300M parameters total. It preserves metric scale for downstream robot learning and 3D vision use, and was trained with an automated data curation pipeline; the team released 3M RGB depth pairs alongside the model.
3d
Overview
LingBot-Depth is a depth completion model that refines noisy or incomplete depth sensor data into high quality, metric accurate 3D measurements. It uses masked depth modeling to jointly align RGB appearance and depth geometry in a unified latent space, producing dense depth maps and point clouds for robot learning and 3D vision applications.
Last updated: July 8, 2026
