Full measurements and figures on Digital Garden
In Houdini FX 22.0.429, I compared two networks on the same 100,000 UV samples. Both predicted RGB, specular strength and a mask. Giving the second network eight frequency bands reduced the five-channel RMSE on held-out samples from 0.26167 to 0.01507. The mask showed the largest visible change.
On a procedural heightfield with 1,000 training points, nine frequency bands increased the error from 5.56 to 10.42 world units. A direct evaluation of the terrain formula came within 0.0133 units of the reference. With a one-unit tolerance, I kept the formula.
The export had a separate test. Feeding the same 10,000 coordinates to PyTorch and Houdini produced a maximum difference of 0.0000363. A close match there tells me the export preserved the network, even when the network was a poor fit for the terrain.
Codex assisted with the experiments and writing. Read the complete Houdini study.

Top comments (0)