Obeida Zakzak

Here's a possible method without having to loop over samples, which executes fast on 13M points.

Couldn't find your model to test over it, as it wasn't attached on your post.

1. Create f@h attribute and zero-out altitude on points
2. Use a Fuse SOP to snap points and average their f@h value : this will result with an average along Y-axis
3. Transfer f@h from fused/averaged points to the zero-altitude points
4. Copy back f@h to original points

In Houdini 20, SideFX introduced a new syntaxe for OpenCL, similar to VEX with @-syntaxe, and they provided a documentation for this one https://www.sidefx.com/docs/houdini/vex/ocl.html [www.sidefx.com]

So it might help in case you are familiar with VEX.

You can find an example with Heightfields (which are 2D volumes) in the Content Library https://www.sidefx.com/contentlibrary/ml-terrain/ [www.sidefx.com]

This node expects Houdini's native volumes, not VDBs, so in your scene you can use a VDB Convert node to convert VDB to Volume and be able to use ONNX Inference node.

However, the onnx model you are using seems to output 1D volumes, so you need to copy data to 3D volume using 3D to 1D index conversion, I used formula from this site to do the conversion and copy results from ONNX node to expected output volumes https://ennogames.com/blog/3d-and-2d-coordinates-to-1d-indexes [ennogames.com]

Also added an example without VDBs, by creating either a 2D (image) or 3D (smoke) volume.