Designing Animatable Stylized Characters with Multi-Layer 3D Gaussians

In this work, we introduce a unified framework that produces fully rigged, animatable characters from a single image, while supporting additional references when available.

November 16, 2026
ACM SIGGRAPH Asia (2026)
 

 

Authors

Wenqing Wang (DisneyResearch|Studios/ETH Zurich)

Clara Fernandez (DisneyResearch|Studios)

Yang Zhang (DisneyResearch|Studios)

Abdelaziz Djelouah (DisneyResearch|Studios)

 

 

Designing Animatable Stylized Characters with Multi-Layer 3D Gaussians

Abstract

Animatable stylized characters are central to animation pipelines, yet transforming a few hand-drawn references into a rigged, editable, and renderable asset remains labor-intensive. Existing generative methods tend to deviate from the original artist-provided designs, while single image based methods offer unclear paths for integration into artist workflows. We introduce a unified framework that produces fully rigged, animatable characters from a single image, while supporting additional references when available. The core representation couples a deformable mesh with a multi-layer Gaussian field, enabling layered rendering, stable articulation, and rich view-dependent appearance. Fitting is accelerated by generative priors in 3D and image space. Beyond single-image animation, the system supports an artist-in the-loop workflow for assembling or extending a character sheet through iterative image edits that are straightforward to model and control in a Gaussian formulation. Our design allows for decoupled modules that learn complex appearance phenomena, such as line work and lighting. We report results that match or exceed prior approaches while better matching goals of alignment and controllability, for stylized 4D character synthesis.

Copyright Notice