ML Deformer H20.5
Film/TV

This is an example of machine learning (ML) entirely built using the example-based ML nodes that are released with Houdini 20.5. This scene demonstrates how a variety of ML setups can now be built entirely from within Houdini just by putting down ML nodes and setting parameters, without the need to write custom training scripts. The example-based ML nodes consist of eleven SOP nodes that allow for data set generation, pre-processing, and inferencing. In addition, there is a single ML TOP node, which is a generic training node.

The ML Deformer setup demonstrates how ML can be applied to train a deformer that creates noticeably more realistic results than regular linear blend skinning. Randomly sampled poses are combined with the results of quasi-static simulation on those poses. This results in a data set that is used for training. The trained model can effectively approximate the results a quasi-static simulation would give for unseen poses that were not part of the data set.

This file demonstrates the same basic idea as the ML Deformer content library example that was released with Houdini 20.0. However, the stages of the data set generation are significantly simpler than before, due to the new example-based ML nodes. In addition, a separate ML training script is no longer needed. Instead Houdini's new built-in training node is used.

Other new features that were not in the 20.0 ML Deformer include:

• Support for constraining parts of the skin
• Additional visualizations
• A minotaur model with corresponding animations
• A skinning scene and asset that allow you to pose a trained model using APEX

Thanks to Cody Spahr for prepping the Minotaur example for this demo.