Skip to content

← ArchiveTalk2024

Speed Up Animation Workflows With Maya's ML Deformer, Powered by Autodesk AI

Autodesk

Official Autodesk video introducing the Maya ML Deformer, which uses machine learning to approximate complex character deformations for interactive playback speeds during animation, blocking, and crowd work.

Abstract

Rigging supervisor Todd Whittep demonstrates Maya 2025.2's machine learning ML Deformer on a layered walrus-man rig, where segmented low-res geometry is skin clustered with corrective shapes and drives higher-resolution meshes via proximity wraps and blendshapes. He applies the ML Deformer, sets up the control collector by adding the translating and rotating joints that drive the skin cluster, and trains the system against an existing target mesh using a keyed range-of-motion rather than auto-generated random poses, exporting training data with offset delta mode and smoothing iterations. He then walks through the training UI parameters (batch size, epochs, validation ratio, learning rate, hidden layers, neurons per layer, dropout, and principal shapes accuracy), iterating across several solves to reduce crosstalk artifacts where one body part wrongly influences another. The talk shows how pose count drives deformation quality, how component tagging can split different body regions across separate ML solves, and how the trained deformer approximates costly muscle or cloth simulation results for major interactive playback speedups.

How to read this

Category
Production talk / tool walkthrough (ML-based deformation approximation)
Contributions
  • Demonstrates Maya 2025.2's ML Deformer learning a layered character rig's deformation for interactive playback
  • Shows a practical training pipeline: control collector setup, keyed range-of-motion training data, offset delta mode with smoothing
  • Walks through training UI parameters and component tagging to split body regions across separate solves and reduce crosstalk artifacts
Context
A vendor-tool realization of fast learned deformation approximation, in the lineage of Bailey et al.'s Fast and Deep Deformation Approximations, packaged inside Maya to replace costly muscle or cloth simulation at runtime.Builds on: Fast and Deep Deformation Approximations
Correctness
Studio/tool practice rather than peer-reviewed; quality is shown to depend on pose count and careful region splitting, and the demo itself notes crosstalk artifacts that require iterating across several solves, so results are approximations tied to the training range of motion.
Clarity
Very accessible as a screen-recorded demo; one pass conveys the workflow, with no formulation to chase.
How to read it
Watch for the data-prep and parameter choices (keyed ROM vs random poses, offset delta mode, epochs/neurons, component tagging); treat it as a how-to and revisit specific UI steps when you actually set up a solve, rather than for theory.

Builds on

Built upon by

Nothing yet.

Related work

Keywords

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →