naver/anny

What it solves

Anny provides a differentiable human body mesh model that can represent a wide range of human shapes, from infants to elders, using a consistent topology and parameter space. It offers an open-source, free alternative to proprietary parametric human models.

How it works

Built with PyTorch, Anny uses a parametric approach to generate 3D meshes based on several input types:

• Skeletal Rig Pose: Controls the articulation and posture of the body.
• Phenotype Parameters: High-level shape controls (ranging from 0 to 1) to define the general body type.
• Local Shape Changes: Specific adjustments (ranging from -1 to 1) for detailed body modifications, such as pregnancy.
• Facial Actions: Blendshapes that control facial expressions (ranging from 0 to 1).

The model supports multiple mesh topologies (including native Anny, MakeHuman, and interoperable options like SMPL-X and SOMA) and different skeletal rigs to balance between complexity and performance.

Who it’s for

• 3D artists and researchers working with human digital doubles.
• Developers building applications that require differentiable human body modeling.
• AI researchers focusing on computer vision and human pose/shape estimation.

Highlights

• Differentiable: Written in PyTorch, making it suitable for optimization and learning tasks.
• Versatile Shape Range: Capable of modeling humans of all ages and body types.
• Interoperable: Supports multiple topologies for compatibility with ecosystems like SMPL-X and SOMA.
• Comprehensive Control: Includes specific models for full body, hands, and faces.