Abstract
CharacterMixer is a novel system designed for the interpolation of rigged 3D characters with disparate mesh and skeleton topologies. Traditional morphing techniques often focus exclusively on surface geometry, resulting in intermediate states that lack functional rigs and cannot be animated. CharacterMixer addresses this by maintaining a "unified skeleton" throughout the transition. By utilizing hierarchical skeleton correspondence and part-based Signed Distance Fields (SDFs), the system ensures semantic consistency and allows for real-time posing and animation of characters even during the morphing process.
Introduction
In the realms of 3D animation and game development, the ability to interpolate—or morph—between two distinct characters is a vital tool for creating hybrid creatures or smooth visual transitions. However, a significant gap exists between geometric interpolation and functional animation. Most existing methods treat characters as static meshes, ignoring the internal skeletons (rigs) that drive movement. When a rigged character morphs into another with a different structure, the underlying animation data is typically lost, leaving the user with an "unposable" intermediate model. CharacterMixer bridges this gap by introducing a rig-aware pipeline that preserves the animation structure throughout the interpolation sequence.
The Challenge: The Rigging Gap
The primary difficulty in interpolating rigged characters lies in the topological differences between skeletons. Mapping a humanoid skeleton to a quadruped or a multi-limbed monster involves complex structural changes that traditional vertex-mapping cannot resolve. Prior solutions generally required both characters to share an identical skeleton or ignored the rig entirely, leading to "ghosting" artifacts or the total collapse of the character's ability to move during the transition. CharacterMixer seeks to solve this by treating the character not as a single mesh, but as a collection of semantically linked parts.
Methodology
Hierarchical Skeleton Correspondence
The foundation of CharacterMixer is its ability to automatically find matches between source and target skeletons. The system employs a custom heuristic to handle various mapping scenarios:
- 1-to-1: Direct mapping (e.g., left arm to left arm).
- 1-to-many: Mapping a single bone to a more complex structure (e.g., a single head bone morphing into a multi-jointed neck and head).
- 1-to-void: Handling features that appear or disappear (e.g., a character with a tail morphing into one without).
The Unified Skeleton
To enable animation during the morph, the system constructs a "unified skeleton"—a proxy rig that exists at any point $t$ during the interpolation. This allows animators to manipulate the unified skeleton, with the system automatically transferring those poses back to the underlying source and target models. This ensures that a character can perform a run cycle or a jump while simultaneously changing its physical form.
Part-based Signed Distance Fields (SDFs)
Rather than interpolating the entire mesh globally, CharacterMixer segments the mesh into parts based on bone influence. Each part is converted into a Signed Distance Field (SDF). By performing blending in the local space of each bone, the system prevents common artifacts like "melting" or limb detachment. To maintain the relative proportions of features like claws or facial features, the system utilizes a specialized bounding box mapping technique within the bone's local coordinate system.
Experimental Evaluation
The researchers evaluated CharacterMixer using 36 character pairs from the RigNet dataset. These models, sourced from Models Resource, represent a wide array of topologies, including bipeds, quadrupeds, and fantastical creatures. The goal was to test the system's robustness across diverse skeletal structures and mesh densities.
Results and Discussion
Qualitative and Quantitative Performance
CharacterMixer demonstrated significant improvements over state-of-the-art methods such as ConvWasser (optimal transport) and NeuroMorph (learning-based). Qualitatively, the system produced fewer distortions and maintained better structural integrity. Quantitatively, the custom 3D skeleton matching heuristics achieved 86% accuracy, outperforming standard 2D-centric algorithms.
Computational Efficiency
The system generates a posed, interpolated character in approximately 83 seconds on average. While not yet instantaneous, the authors introduced an "advection" method that improves speed by 17% compared to standard voxelization with negligible impact on visual quality. Currently, 70% of the processing time is dedicated to the voxelization process (converting meshes to SDFs).
Limitations and Future Work
Despite its successes, CharacterMixer has areas for improvement. The SDF-based approach can occasionally lose fine-grained surface textures or sharp geometric details. Additionally, while the system is robust, "extreme" correspondence cases—such as morphing a snake into a spider—may still require manual intervention via the provided UI. Future research will focus on integrating Neural SDFs to move toward real-time performance and implementing more sophisticated motion retargeting to handle drastically different limb orientations more fluidly.
Conclusion
CharacterMixer represents a significant step forward in 3D character modeling by making interpolation "rig-aware." By maintaining a functional skeleton throughout the morphing process, it allows for a seamless blend of geometry and animation, providing artists and developers with a powerful new tool for dynamic character creation.
Summary updated