Convert DA3 Geometry to Mesh

The DA3 pipeline's payoff node. DA3GeometryToMesh takes the da3_geometry packet from DA3Inference and triangulates the depth map into a MESH - with your source image baked in as the base color texture. This is the DA3 family's answer to MoGe's point-map-to-mesh node, with one extra trick up its sleeve: in multi-view mode it aligns every frame into a shared world frame before meshing, so you can reconstruct a whole scene, not just one photo's worth of depth.

How it works

Each pixel of the depth map gets unprojected into 3D using the camera intrinsics (inverse-K), the result is filtered through the sky and confidence masks, and the surviving pixels are connected into a grid mesh - the same triangulation core MoGe uses. Then it does the coordinate conversion (OpenCV Y-down/Z-forward to glTF Y-up/Z-back) and hands you a MESH. Multi-view is where it gets interesting: the geometry packet carries extrinsics, so each frame is placed in a common world space before triangulation, which is how a stack of photos becomes one coherent reconstruction. Feed the MESH into Save 3D Model for a .glb.

The dials that matter

• batch_index (default 0) - which image of a batched geometry to mesh. Batches can't be stacked into one MESH because per-image vertex counts differ.
• decimation (1–8, default 1) - vertex stride; 1 is full resolution, 2 is half, and so on. Drop it for lighter meshes.
• discontinuity_threshold (0–1, default 0.04) - drops triangles whose 3×3 depth span exceeds this fraction. This is the anti-tearing dial at occlusion edges; 0 turns it off.
• confidence_threshold (default 0.1) - excludes low-confidence pixels, but only when the geometry has a confidence map, i.e. Small/Base models. 0 keeps everything.
• use_sky_mask (default on) - excludes sky pixels (sky ≥ 0.5) for Mono/Metric models.
• texture (default on) - carries the source image through as baseColor.

Where people get burned

Two things. First, mono mode has no intrinsics - the node falls back to an assumed ~60° FOV pinhole camera and logs a warning that 3D reconstruction may be inaccurate. Metric-Large gives you real metres, but the geometry is only as good as that assumed lens. Second, the empty-mesh error: it happens when sky/confidence filtering culls too much. The node tells you the fix in its own error - raise discontinuity_threshold, lower confidence_threshold, or disable use_sky_mask. Non-finite depth pixels get zeroed with a warning before unprojecting, so those don't silently corrupt the mesh. And as with everything monocular, occluded backsides simply aren't there - multi-view input is the only way around it.