Fast Deformation Method in Maya

Idea

The idea was inspired by Ziva Dynamics work to reduce the complexity of film rigs i.e. bones influence and deformation but maintaining the same visual quality

Use cases:

Realtime rendering
Procedural animation
Smaller scene size
Faster loading, less computation
To be able to run on low-end systems i.e. Mobile phones

Solution

The solution to this problem was suggested by Bailey in the Fast and Deep Deformation Approximation paper.

The main idea of the paper was to learn from the target rig and approximate vertex positions in your current rig.

The paper explains the following concepts:

One neuron per Joint
Hidden layer of neuron where each neuron influence each other result
Input is multiple sample of joints transformation
Network will be created based on above input
The above network is then trained
And result is the predicted vertex position

Implementation

We need the following software and tools to implement this paper:

Tensorflow, machine learning library
Maya 2022, supports for python3
Google Colab to train on cloud
Film rig to work on

Algorithm

Make a copy of the target rig with only one vertex influence per joint
For each frame of sample animation
a.For each vertex, make a record on how far it is as compared to where it is meant to be in the target rig
b.For each joint, make a record of rotation and translation values
Data for training: a & b
Model is trained: Guess correct vertex offset values
Move vertices to the correct positions