Point to Mesh Distance

A point cloud is just a bag of coordinates until you know how each point relates to the surface it was scanned from. Point to Mesh Distance computes exactly that: for every point in your cloud, how far is it from a target mesh - and optionally which side it's on, inside or outside. The output is your point cloud back again with a per-point distance field attached, ready to colorize, filter, or feed into a reconstruction.

It's part of ComfyUI-GeometryPack ("ComfyUI libigl") by Andrea Pozzetti, and it's one of the pack's two distance nodes - this one is point-centric, its sibling Mesh to Mesh Distance is about whole-surface error.

How it works

You give it a target_mesh (the surface to measure against) and a pointcloud (which the code happily accepts even if it's actually a full mesh - it just uses the vertices). Two modes via distance_type:

• unsigned - every point gets the distance to the nearest surface point. Uses trimesh's closest_point query. Fast, robust, no assumptions about the target being closed.
• signed - points also get a sign: negative inside the mesh, positive outside. This is libigl's signed_distance, and it's the powerful one - signed distance is what SDF-based reconstruction and many sculpting/boolean pipelines want.

The sign_method combo only matters for signed mode and is where the fine control lives: default (fast and robust, libigl's choice), winding_number (accurate, handles non-watertight targets), fast_winding_number (the same but faster, slightly less exact), pseudonormal (the legacy test), and unsigned (which forces unsigned behavior even in signed mode). If your target mesh has holes, winding_number is the honest option; default can misclassify points near broken geometry.

The outputs are the pointcloud back (with the distance field added as a vertex attribute) and an info string summarizing the computation.

What you'd actually do with it

• Colorize and inspect - the distance field drops into the pack's field-aware preview nodes, so you can literally see how well a point cloud fits its target surface. Areas of high distance are your scan errors.
• Filter outliers - a point cloud where most points hug the surface but a few float far off? Threshold on the field to flag or drop them.
• Prep for reconstruction - signed distances to a reference are the standard input for making a clean SDF-derived surface.

Install and gotchas

Pack standard: search "GeometryPack" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-GeometryPack.git
cd ComfyUI-GeometryPack
pip install -r requirements.txt --upgrade
python install.py

Restart after. The pack installs via the experimental comfy-env + pixi route with a heavy isolated environment; libigl arrives with it, but signed mode will throw a clear "install libigl" error if it's ever missing.

The traps are conceptual more than technical. Signed distance assumes your target is a reasonably closed surface - against a half-open shell, the sign near the opening is unreliable, which is exactly the case winding_number exists for. And the two meshes must share a coordinate frame; if your point cloud is in scan space and the target in world space, the "distances" are nonsense and won't look obviously wrong. Normalize both, or don't normalize either.