Continuous Skinning Fields for Animatable 3D Shapes

24 concepts19 papers1 proposal

In traditional computer graphics, animating an articulated 3D character relies on Linear Blend Skinning (LBS). In this classical paradigm, vertices on a discrete surface mesh are bound to an underlying skeletal structure via a fixed set of skinning weights. While highly effective for explicit polygonal meshes, this approach breaks down when applied to modern implicit neural representations—such as Neural Radiance Fields (NeRFs), Signed Distance Fields (SDFs), and 3D Gaussian Splatting. Because these representations define geometry and appearance as continuous functions across 3D space rather than discrete surfaces, they lack explicit vertices to rig.

To bridge this gap, a powerful paradigm has emerged: continuous skinning fields (also known as volumetric neural skinning). Instead of assigning weights to discrete points, these methods define skinning weights as a continuous function over the entire 3D space. This shift allows unstructured primitives, volumetric densities, and implicit surfaces to be seamlessly deformed and animated according to a kinematic chain.

Mapping Space via Inverse Skinning

The primary mechanism that makes continuous skinning fields work is the establishment of a coordinate mapping between a deformed "world" space and a static "canonical" rest pose.

When rendering or evaluating an articulated neural model in a specific pose, a query point is sampled in the world space. To determine its physical properties (such as occupancy, color, or signed distance), the system must "unpose" or map this point back to the canonical space where the template or neural network is defined. As demonstrated by Marko Mihajlovic et al. in LEAP, a neural network can be trained to predict volumetric skinning weights for any spatial coordinate, enabling the continuous deformation of implicit shapes across highly articulated poses.

This inversion process is formalized by Mohammad Norouzi et al. in NiLBS, which represents skinning weights as a continuous volumetric field to map arbitrary spatial coordinates back to a canonical rest pose. This allows for fast queries of pre-computed spatial data. Similarly, Wenbin Lin demonstrates in EditableNeRF that a weight estimation network can predict the influence of sparse 3D key points across a continuous volume, adapting LBS to enable localized deformation control within neural fields.

Parameterizing the Field: MLPs vs. Voxel Grids

To implement a continuous skinning field, researchers parameterize the weight function using either coordinate-based neural networks or interpolated voxel grids.

Coordinate-based multi-layer perceptrons (MLPs) treat skinning weights as a continuous function of spatial coordinates and skeletal pose. For instance, Zijian Dong et al. in PINA and Sijing Wu et al. in GANHead utilize neural networks to learn continuous LBS weights that adapt directly to the specific topology of implicit neural representations.

However, relying solely on high-capacity neural networks can introduce training instabilities or high computational overhead during rendering. To address this, some frameworks opt for explicit volumetric structures. Guikun Chen et al. in TAGA employ a low-resolution voxel grid to store skinning weights, using trilinear interpolation to ensure spatially smooth transitions between body segments while improving training stability. A similar voxel-based grid interpolation strategy is leveraged by Chandradeep Pokhariya et al. in MANUS to assign smooth deformation weights to unstructured 3D Gaussian primitives, avoiding the discontinuities associated with discrete mesh vertices.

Deforming Loose Geometry and Complex Dynamics

One of the greatest strengths of a volumetric skinning field is its capacity to handle loose-fitting geometry, such as clothing, skirts, or hair, which do not strictly follow the underlying skeletal motion.

In traditional setups, points that float away from the body surface lack a reliable skeletal binding. Volumetric skinning fields solve this by diffusing or interpolating weights into the surrounding 3D space. As shown by Igor Santesteban et al. in Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On, volumetric diffusion of skinning weights allows loose-fitting layers to inherit the motion characteristics of the underlying body consistently. To capture the complex dynamics of loose clothing like skirts without tearing artifacts, Fangzhou Hong et al. in Garment4D interpolate skinning weights from the K-nearest neighbors of the body surface, preventing discontinuous stretching.

Furthermore, when modeling humans in challenging clothing, the authors of Neural Point-based Shape Modeling of Humans in Challenging Clothing predict per-point transformation weights via a neural network, allowing surface points to be influenced by multiple joints simultaneously. This flexibility is echoed by Shaofei Wang et al. in ARAH, where learned skinning weights enable articulated Signed Distance Fields to represent complex, non-rigid garment deformations that diverge from rigid skeletal motion.

Proximity Disambiguation and Resolution Independence

Beyond enabling the animation of implicit shapes, continuous skinning fields solve two fundamental limitations of classical geometric skinning: semantic proximity errors and resolution dependence.

First, traditional skinning methods often rely on simple geometric distance to assign weights. This causes failures when two distinct body parts come into close contact—for example, a hand moving near a torso might incorrectly inherit transformations from the torso bones. Continuous neural fields overcome this by learning semantic blending weights. As demonstrated by Yuxin Yao et al. in DynoSurf, a neural network can learn to distinguish between spatially proximal but semantically distinct components, ensuring clean separation during complex interactions.

Second, because a neural skinning field is a continuous function, it is entirely independent of the underlying geometric resolution. In VINECS, Vladislav Golyanik et al. leverage a coordinate-based neural network to represent skinning weights as a function of 3D coordinates and pose. This formulation allows a model to be trained on a low-resolution proxy and directly applied to high-fidelity cinematic meshes without retraining or manual weight painting.

This resolution-agnostic property is also vital for unstructured data. In Differentiable Robot Rendering, an implicit LBS network predicts weights for arbitrary spatial coordinates, allowing unstructured primitives like point clouds or Gaussians to deform realistically according to a kinematic chain. Finally, frameworks like MagicArticulate, developed by Chaoyue Song et al., generalize this by treating skinning weights as a continuous function refined via diffusion, capturing complex relationships between arbitrary geometries and underlying skeletal structures.

Key Foundational Work

LEAP: Learning Articulated Occupancy of People (2021) — contributed 3 concept elements to this direction.
NiLBS: Neural Inverse Linear Blend Skinning. (2020) — contributed 2 concept elements to this direction.
Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On (2021) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_skinning_fields_a8113945,
  title = {continuous skinning fields},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-skinning-fields/rdr_a8113945f5f445889ea383e06e646db8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_skinning_fields_a8113945_v1,
  title = {continuous skinning fields},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-skinning-fields/rdr_a8113945f5f445889ea383e06e646db8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake