Houdini 21 offers a Neural Point Surface node that uses pre-trained machine learning models to convert point clouds — clusters of 3D data points produced by simulations or 3D scans — into smooth, detailed meshable surfaces. The feature ships with four built-in models tuned for different materials, such as liquids and granular matter.

The Neural Point Surface node takes a point cloud as input and runs ONNX model inference to reconstruct a VDB surface, producing crisper high-frequency detail and better temporal stability than the previous analytic particle fluid surface method. SideFX provides four pre-trained models (including Liquid, Granular, Balanced, and Smooth), and the training setup is downloadable from SideFX's Content Library so users can train custom models. The node is GPU-accelerated and is embedded in the existing Particle Fluid Surface and MPM Surface nodes in Houdini 21.