New ML Regression Linear SOP allows training and inferencing using linear models entirely within SOPs. This fits a linear model to your training data.

New ML Regression Kernel SOP allows training and inferencing using kernel ridge regression entirely within SOPs. This method works by computing optimal weights for kernel functions centered at data points. This approximates your training data using a weighted combination of kernel functions centered at your data points.

New ML Regression Kernel TOP trains a model using kernel ridge regression. This trains the same kind of model as ML Regression Kernel SOP, but allows it to export as an ONNX model to use with the ML Regression Inference SOP.

The new nodes ML Example Raw and ROP ML Example Raw Output allow faster and more direct writing of data sets to disk for training.

Faster performance for ML Example, ML Example Decompose, and ML Extract Example using dedicated C++ core nodes. This speeds up the processing of data sets.

ML Train Regression improvements:

New ML Pose Deserialize SOP node helps to convert a rig pose that was previously serialized using ML Pose Serialize back into a rig pose.

New ML Train Deformer recipe allows you to train your own ML Deformer model, which can give you more realistic animation than traditional animation methods such as linear blend skinning.

New ML Deform SOP allows a model trained using the ML Train Deformer recipe in SOPs.

New APEX Add ML Deformer SOP allows a model trained using the ML Train Deformer recipe in APEX rig.

New Armature Deform SOP and Armature Capture SOP allow you to deform organic tissue repeatedly using skeleton poses. You can use these nodes to generate training data for an ML deformer and other applications outside of ML.

ML Pose Serialize SOP and ML Pose Deserialize have new options to control whether local or world transforms are serialized/deserialized.

New ML Train Style Transfer TOP trains a machine learning model for doing style transfer between two classes of images. The resulting model can be used with ONNX Interface SOP or ONNX Inference COP .

New ML Train OIDN TOP is a wrapper around the OIDN training script used to train an OIDN denoising filter model on preprocessed training and validation datasets.

New ML Preprocess OIDN TOP is a wrapper around the OIDN preprocessing script used to preprocess training and validation datasets compliant with the OIDN dataset naming scheme.

Added the PDG ML Training Monitor Panel, which can inspect the training progress and plots of the ML Train Style Transfer TOP and ML Train Regression TOP .

New ML Train Volume Upres recipe allows you to train your own pyro upres model. A trained model can be used to make a lowres pyro sim look more like a highres pyro sim, without actually running a highres pyro sim.

New ML Volume Tile Component allows you to crop tiles from a given volume at one or multiple provided point positions. This is useful for creating training data from 3D volumes. This node is part of the ML Train Volume Upres recipe.

New ML Volume Tile Inference allows you to run an ONNX model on a volume without creating any seams by splitting it up into tiles.

New ML Volume Upres allows you to run an ONNX model that was trained using the new ML Train Volume Upres recipe.

The Sparse Billowy Smoke shelf tool was updated to use ML Volume Upres node to upscale and sharpen the smoke simulation using a pretrained ONNX machine learning model. It creates thick cloud-like smoke rising from a spherical base and is useful as a starting point for simulating scenarios such as chimney smoke or cooling tower steam.

New Neural Point Surface SOP takes a point cloud as input and reconstructs a VDB surface from it. It contains multiple specialized pretrained models to yield specific looks depending on the material being surfaced. In general, this node should allow users to reconstruct smooth surfaces while preserving the sharp, high frequency, details described by the point cloud.

Options to use the Neural Model are available on the MPM Surface SOP.