Source-linked AI summary
MHR: Momentum Human Rig
Aaron Ferguson, Ahmed A. A. Osman, Berta Bescos, Carsten Stoll, Chris Twigg, Christoph Lassner, David Otte, Eric Vignola, Fabian Prada, Federica Bogo, Igor Santesteban, Javier Romero, Jenna Zarate, Jeongseok Lee, Jinhyung Park, Jinlong Yang, John Doublestein, Kishore Venkateshan, Kris Kitani, Ladislav Kavan, Marco Dal Farra, Matthew Hu, Matthew Cioffi, Michael Fabris, Michael Ranieri, Mohammad Modarres, Petr Kadlecek, Rawal Khirodkar, Rinat Abdrashitov, Romain Prévost, Roman Rajbhandari, Ronald Mallet, Russell Pearsall, Sandy Kao, Sanjeev Kumar, Scott Parrish, Shoou-I Yu, Shunsuke Saito, Takaaki Shiratori, Te-Li Wang, Tony Tung, Yichen Xu, Yuan Dong, Yuhua Chen, Yuanlu Xu, Yuting Ye, Zhongshi Jiang
TL;DR
Surface-derived joints and entangled shape and skeleton can limit control in parametric human models. MHR combines a decoupled skeleton and shape representation with a corrective-driven rig; evaluation reports lower fitting error with fewer components, while its expression and pose correctives remain independent of body shape.
Problem
Surface-derived joints and entangled shape and skeleton limit direct skeletal control and precise customization of body proportions and soft tissue.
Source: §1 · p. 2Method
Results
MHR demonstrates lower fitting error with fewer components and closer qualitative fits around articulated joints and shoulders.
Source: §4 · p. 8Takeaways & Limitations
MHR is presented as a flexible, expressive model for animation, vision, and AR/VR applications.
Source: §6 · p. 9MHR's facial expression and pose corrective blendshapes are independent of body shape; conditioning them on body shape is future work.
Source: §5 · p. 8
Abstract
from arXiv · showhide
We present MHR, a parametric human body model that combines the decoupled skeleton/shape paradigm of ATLAS with a flexible, modern rig and pose corrective system inspired by the Momentum library. Our model enables expressive, anatomically plausible human animation, supporting non-linear pose correctives, and is designed for robust integration in AR/VR and graphics pipelines.
1 Introduction
Existing parametric body models can entangle shape and skeleton, limiting control over skeletal attributes and body proportions. MHR builds on ATLAS with updates for artist and industry needs, including decoupled body structure, expression blendshapes, and a multi-resolution identity dataset.
1 Introduction
- Parametric body models map compact shape and pose descriptors to articulated meshes, supporting applications that depend on representing human form and movement.The introduction connects these models to avatar creation, motion capture, human-object interaction, and character synthesis.Source: §1 · p. 1
- Surface-derived joints and entangled shape and skeleton limit direct skeletal control and precise customization of body proportions and soft tissue.These limitations complicate keypoint fitting and production workflows.Source: §1 · p. 2
- ATLAS separates external shape from the internal skeleton and uses sparse, nonlinear pose correctives, but its FLAME-based expressions and skeleton are not suited to common artist workflows and added pose correctives.The cited concerns are expression compatibility and skeleton optimization for correctives.Source: §1 · p. 2
- MHR builds on ATLAS with a decoupled skeleton and mesh, semantic expression blendshapes, and a dataset supported at multiple levels of detail.The model is described as flexible and expressive, with a license supporting free experimentation.Source: §1 · p. 2
2 Related Work
Prior work expanded mesh deformation, body-part coverage, and pose-corrective representations, while persistent issues include shape-skeleton entanglement and limits in accessible, controllable rigging. MHR and ATLAS are presented as combining decoupled shape and skeleton with rich pose correctives and finger articulation.
3D Human Mesh Modeling
- Vertex-based models including SMPL Loper et al. (2015) expanded body representations, but most models still regress skeletal joints from the mesh surface, entangling shape and skeleton.Earlier work explored triangle-level pose and shape separation, and later models added hands, faces, sparse corrections, or broader shape spaces.Source: §2 · p. 2
Skeleton Models
- Biomechanics and medical-data approaches model anatomical skeletons, while graphics-oriented decoupled models may lack open licensing, expressive pose correctives, or fine-grained finger control.Some approaches rely on specialized simulation tools, and retargeting extracted skeletons to new poses can require additional optimization.Source: §2 · pp. 2–3
PoseCorrectiveDeformations
- MHR and ATLAS balance corrective flexibility and control with sparse, nonlinear pose correctives intended to limit spurious mesh dependencies while modeling complex deformations.Earlier corrective methods include local offsets, key-pose interpolation, PCA spaces, and learned mappings from joint rotations to mesh deformations.Source: §2 · p. 3
3 Model
MHR combines a decoupled surface and skeleton with distinct pose, expression, identity, and pose-corrective representations. Its implementation supports multiple levels of detail and graphics and machine-learning workflows.
3 Model
- MHR keeps the external surface decoupled from its internal skeleton, with LBS applying shape and pose inputs while the mesh remains aligned to a fixed skeleton.The model has 127 joints and supports six levels of detail.Source: §3 · p. 3
3.1 Pose Parametrization
- A 204-parameter vector maps linearly to joint settings, enabling restricted degrees of freedom and coordinated motion such as supplemental limb twists.Artist-defined skinning weights retain structure and locality, which optimized weights tended to lack.Source: §3.1 · p. 4
3.2 Facial Expressions
- MHR uses 72 artist-sculpted, sparse semantic expressions following FACS Ekman and Friesen (1978), rather than dense data-derived expression spaces.The authors report that these expressions simplify connections to synthetic data pipelines and help eliminate spurious pose movement in related work.Source: §3.2 · p. 4
3.3 Identity Space
- Identity shape is fixed across frames for a given skeleton and split into body, head, and hand subspaces for artist control and part-specific scan data.Soft masks precede separate PCAs, yielding 20 body, 20 head, and 5 hand components; training uses 7,110 body scans and separate hand and head datasets of 3,000 and 2,138 subjects.Source: §3.3 · pp. 4, 6
3.4 Pose Correctives
- Pose correctives combine local nonlinear joint-neighborhood embeddings with geodesic-initialized sparse activations to model pose-dependent mesh deformations.The correctives are trained on 26,000 scans: 13,000 full-body and 13,000 hand scans.Source: §3.4 · pp. 6–7
3.5 Implementation Details
- The Momentum implementation provides C++ and Python APIs, supports FBX and GLTF export, and integrates with PyTorch frameworks.Identity, expression, and corrective components are transferred across levels of detail using linear mappings or subdivision, depending on the level.Source: §3.5 · pp. 7–8
4 Evaluation
On 3DBodyTex, MHR achieves lower fitting error with fewer components and fits previously unseen identities; qualitative results are strongest around articulated extremities and shoulders.
4 Evaluation
- The evaluation fits body shape and pose by minimizing scan-to-model surface distances plus a head-and-hand landmark term.Hands, face, and hair are masked; optimization uses Adam for 2,500 iterations at a 0.01 learning rate.Source: §4 · p. 8
- MHR demonstrates lower fitting error with fewer components while representing posed body shapes for previously unseen identities.The reported average distance excludes face, hair, and hands.Source: §4 · p. 8
- Qualitatively, MHR fits articulated extremities such as elbows and knees and scan shoulders more closely.Source: §4 · p. 8
5 Discussion and Future Work
The discussion frames rig decisions as trade-offs, favoring a simpler skeleton and a four-joint influence limit in most LODs. It also identifies body-shape-independent correctives and omitted facial geometry as scope boundaries, alongside future work.
5 Discussion and Future Work
- MHR simplifies the ATLAS skeleton because extra glute and upper-back joints can improve LBS-only accuracy but make pose optimization harder.With pose correctives, the authors choose simplicity for the skeleton.Source: §5 · p. 8
- Figure 8 reports vertex-to-vertex error across fitting-component counts; Figure 9 shows two views of three scans alongside SMPL, SMPL-X, and MHR.Figure 9 also shows scan-model overlap, the model alone, and distance heatmaps with masked areas in black.Source: Figure 8 · p. 8, Figure 9 · p. 9
- Most LODs cap each vertex at four joint influences, while the highest L0 LOD exceeds that limit to avoid sharp skinning-boundary creases.The authors note that more influences can benefit LBS accuracy and smoothness.Source: §5 · p. 8
- MHR targets geometry in minimal scans and has no explicit eyeball geometry; an explicit mouth system for teeth and tongue is planned.Source: §5 · p. 8
- Facial-expression and pose-corrective blendshapes are independent of body shape, and conditioning them on shape remains future work.The authors note that shape-varying pose correctives can be more accurate and shape-dependent expressions can improve realism.Source: §5 · p. 8
- Planned directions include soft-tissue and clothing integration, real-time optimization and AR/VR deployment, and stylized characters.Source: §5 · p. 8
6 Conclusions
MHR combines ATLAS's decoupled skeleton/shape paradigm with a modern corrective-driven rig, yielding a flexible, expressive model for animation, vision, and AR/VR applications.
A Author Contributions
MHR is described as the culmination of more than nine years of work on multiple iterations of anatomically inspired body models developed at Meta.
A Author Contributions
- Aaron Ferguson contributed modeling, rigging, artistic direction, pose correctives, SAM3D integration, and ATLAS development.Source: A Author Contributions · p. 13
- Ahmed A. A. Osman contributed to statistical model development; Matthew Hu and Romain Prévost contributed hand and head scan registrations, respectively.Source: A Author Contributions · p. 13
- Carsten Stoll and Chris Twigg contributed to Momentum library development; Christoph Lassner contributed to software and pipeline development.Source: A Author Contributions · p. 13
- Eric Vignola contributed rigging, tooling, and artistic supervision; Jeongseok Lee and Ladislav Kavan contributed to Momentum and body pose correctives.Source: A Author Contributions · p. 13
- Fabian Prada contributed to ATLAS development, while Federica Bogo contributed to statistical modeling, scan registrations, and pipeline development.Source: A Author Contributions · p. 13