Simulation can be used to obtain deformations that look more realistic than linear blend skinning. However, linear blend skinning is much faster. Can we learn from simulated poses to improve on linear blend skinning? For an example that shows how linear blend skinning can be improved on by learning from random poses, download the ML Deformer files from the Content Library.

This setup uses a TOP network to control all the stages of the Machine Learning: data generation, preprocessing, and training. It demonstrates the use of several new Houdini 20 features that support ML, including the new ONNX SOP, the new Principal Component Analysis SOP, and enhancements to the Python Script TOP.

The ML Terrain example on the Content Library demonstrates how you can use TOPs to generate terrain data, and then use that data to train an ML model that’s able to produce new terrain given a simple sketch from an artist. The model captures the style of the terrain that it’s trained on. For example, if the input terrain data consists of desert dunes, the model learns to produce terrain in that style. The trained model is evaluated by Houdini using the new ONNX SOP.

This demo is intended to be a full walkthrough of the process from creating an ML model and using it directly in Houdini. It is also built with exploration in mind. Plug in your own terrain erosion HDAs, or use heightfield patterns. Use higher or lower resolution to compare results. Test against massive terrain, or work with small fine tune details. Train a moon crater ML model or alien world, or simply just use existing real world lidar data.

The ONNX Machine Learning Terrain demo comes in two parts. Download the files here.