Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization

TL;DR

Auto-Connect presents a connectivity-preserving rigging framework that explicitly encodes skeletal topology via a tokenization scheme and RigFormer. A topology-aware reward and reward-guided Direct Preference Optimization (DPO) refine topology beyond next-token prediction, while a geodesic-aware bone probability module improves skinning by selecting influential bones based on geodesic cues. The method achieves superior joint localization accuracy, topological consistency, and deformation quality across benchmarks, outperforming state-of-the-art rigging and skinning approaches. This pipeline enables robust, animation-ready rigs across diverse character morphologies, with potential to streamline professional 3D animation workflows.

Abstract

We introduce Auto-Connect, a novel approach for automatic rigging that explicitly preserves skeletal connectivity through a connectivity-preserving tokenization scheme. Unlike previous methods that predict bone positions represented as two joints or first predict points before determining connectivity, our method employs special tokens to define endpoints for each joint's children and for each hierarchical layer, effectively automating connectivity relationships. This approach significantly enhances topological accuracy by integrating connectivity information directly into the prediction framework. To further guarantee high-quality topology, we implement a topology-aware reward function that quantifies topological correctness, which is then utilized in a post-training phase through reward-guided Direct Preference Optimization. Additionally, we incorporate implicit geodesic features for latent top-k bone selection, which substantially improves skinning quality. By leveraging geodesic distance information within the model's latent space, our approach intelligently determines the most influential bones for each vertex, effectively mitigating common skinning artifacts. This combination of connectivity-preserving tokenization, reward-guided fine-tuning, and geodesic-aware bone selection enables our model to consistently generate more anatomically plausible skeletal structures with superior deformation properties.

Figures (12)

• Figure 1: Overview of the Auto-Connect. The pipeline consists of three main stages. In the Rigging Pre-training stage, a point cloud sampled from the input 3D mesh is processed by the shape encoder to extract geometric features, which are subsequently fed into our autoregressive RigFormer to generate a token sequence. The generated sequence is then processed using our connectivity-preserving detokenization to construct the skeleton tree. In the Rigging Post-training stage, preference pairs are constructed using our topology-aware reward criterion, and RigFormer is fine-tuned with our reward-guided DPO for preference-driven optimization. Finally, in the Skinning Weight Prediction stage, the generated skeleton and mesh vertices serve as input. Our geodesic-aware bone probability prediction module is employed to implicitly determine the most influential bones to predict the skinning weights, enabling mesh animation.
• Figure 2: Connectivity-preserving tokenization process. The number indicates the joint indices.
• Figure 3: Examples of the collected preference pairs. Skeleton trees with higher reward exhibit superior topology and better align with human intuition, making them the preferred choice.
• Figure 4: Qualitative comparison of rigging result on Art-XL2.0 (top) and MR (bottom). Our connectivity-preserving representation effectively captures intrinsic skeletal topology, and reward-guided fine-tuning enables the generated skeletons to better align with artistic aesthetics. Additional results are provided in the appendix \ref{['More Rigging Result']}.
• Figure 5: Qualitative comparison of skinning result on Art-XL2.0 (top) and MR (bottom). Models marked with ${}^\ast$ were trained using our geodesic-aware bone probability prediction module, which effectively mitigates the L1-norm error and enhances skinning performance. Additional results are provided in the appendix \ref{['More Skinning Results']}.