← ArchivePaper2023
Bodyopt: A Character Deformation Pipeline for Avatar: The Way of Water
Christoph Sprenger, Tobias Mack, Alexey Stomakhin, Florian Fernandez
ML-driven character deformation pipeline used in Avatar 2 production, combining optimization-based and learned approaches for hero characters.
Abstract
We present Bodyopt, a character skin deformation framework developed for Avatar: The Way of Water. Our approach aims to learn the skin deformations from a given dataset and reproduce them reliably during shot production. In conjunction with the kinematic skeleton, we employ muscle fibers as an additional anatomical basis, where their length changes serve as a parametrization for the non-linear deformation components. We provide a novel way of curating the dataset to minimizing differences between similar poses, which would otherwise lead to a quality loss in the reconstruction. Our approach also handles runtime skin dynamics and utilities for artists to transfer deformations to new character types as well as extra modifiers for secondary motions like breathing. Additionally, we close the gap between final skin deformation and the representation used in Animation by providing a fast proxy solution that is based on the same input data.
How to read this
- Category
- Production talk: a character deformation pipeline
- Contributions
- Demonstrates Bodyopt, a framework that learns hero-character skin deformations from a dataset and reproduces them in shot production
- Uses muscle-fiber length changes alongside the kinematic skeleton to parametrize non-linear deformation, plus runtime skin dynamics and breathing modifiers
- Provides dataset curation to minimize differences between similar poses and a fast animation-side proxy from the same input data
- Context
- Part of the Avatar: The Way of Water production toolset alongside Weta's Loki framework, combining optimization-based and learned deformation for hero characters.Builds on: Loki: A Unified Multiphysics Simulation Framework for Production
- Correctness
- Studio practice, not peer-reviewed; results are production-proven, and the reported quality gains (for example from pose-difference-minimizing curation) come from production use rather than formal benchmarks.
- Clarity
- Accessible talk-level material; a first pass conveys the muscle-fiber parametrization and proxy idea.
- How to read it
- Read once for the muscle-fiber parametrization, the dataset-curation trick, and the animation proxy; a second look helps if you build ML-driven deformation pipelines.
Built upon by
Nothing yet.
Related work
- Interactive Skeleton-Driven Dynamic Deformations 2002 / SIGGRAPH
- Physically Based Rigging for Deformable Characters 2005 / SCA
- Data-Driven Physics for Human Soft Tissue Animation 2017 / SIGGRAPH
- A Neural Network Model for Efficient Musculoskeletal-Driven Skin Deformation 2024 / SIGGRAPH
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 →