Compact flat-point GPU payload → frame cap 128 (denser); GLB built from points, verified == predictions_to_glb

Clarity: single trajectory line (not 64 frustums), default conf 60, honest shown/kept status

Sample frames across whole video (full-tour coverage); cap 64; downsample display to 2M

Depth->points: unproject depth (Depthmap mode); streaming model has enable_point=False

Preprocess without torchvision (ZeroGPU base lacks it); show_error=True

Community Gradio ZeroGPU demo of LingBot-Map streaming 3D reconstruction