Neural Terrain Paint
Film/TV

This project showcases new machine learning workflows for Heightfields in Houdini 21, featuring an interactive hda Neural Terrain Painting hda, a streamlined training hda for generating models, and an example scene demonstrating how ML can be combined with heightfield SOPs to create large, visually rich landscapes.

The Neural Terrain Painting hda lets users sculpt heightfields using a paint brush, with real-time deformation driven by pre-trained ML models via the onnx Inference SOP. For non-painting workflows, you can simply input a heightfield directly as the system supports both methods seamlessly. The hda also includes helpful parameters to refine ML inference results based on the current terrain setup to yield better results, as well as standard brush controls.

Note: The hda includes three downloadable example onnx files, but will need to be downloaded from its parameter interface.

The included ML Train Terrain hda offers a quick and intuitive way to train your own ML models from custom heightfield edits within Houdini. These models plug directly into the Neural Terrain Painting hda, making a fast and flexible setup. The hda also comes with a pre-built erosion network and a corresponding trained model to demonstrate the full process.

Lastly, the project includes a full example scene that combines newly trained models, created using the ML Training hda, with official SideFX ML terrain models, which can be downloaded via the SHFS (Shared File Server) integration built into the HDA.