Research map / Probabilistic and scientific modeling / Molecules and proteins
Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Auto-Connect serializes skeleton trees with connectivity-encoding special tokens (E1/E2) in an autoregressive RigFormer, refines it with topology-aware reward-guided DPO, and adds a geodesic-aware top-k bone selector for skinning.
Problem
Automatic rigging methods predict joints/bones without reliably encoding skeletal connectivity, yielding topological errors (missing chains, disconnected or excess joints) and skinning artifacts from static k-nearest bone selection.
Approach
A BFS-ordered skeleton tokenization with special endpoint tokens for each parent's children and each hierarchy level encodes connectivity directly; a topology-aware reward (CD, TED, Hierarchical Jaccard) drives reward-weighted DPO with an SFT term; and a geodesic-feature MLP learns the top-k influencing bones per vertex.
Key results
- On Art-XL2.0 rigging, Auto-Connect reaches CD-J2J 2.572% and IoU 82.806%, beating the best baseline Magic-spatial (3.041%, 78.675%).
- Reward-guided DPO improves Art-XL2.0 IoU from 82.183% to 82.806% and reduces CD-J2J from 2.695% to 2.572%.
- Adding the geodesic-aware bone module improves HeterSkinNet on Art-XL2.0 from 88.16% to 89.38% precision and reduces avg L1 from 0.43 to 0.42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reward-guided DPO post-training (ablation) | Auto-Connect · Art-XL2.0 | IoU: 82.183% -> 82.806% | Helped |
| add reward-guided DPO post-training (ablation) | Auto-Connect · MR | IoU: 72.213% -> 73.108% | Helped |
| add connectivity-preserving tokenization (ablation) | RigFormer · Art-XL2.0 | topological quality | Helped |
| remove data augmentation (ablation) | Auto-Connect · unseen walking pose | skeleton alignment | Hurt |
| add geodesic-aware bone probability prediction module | HeterSkinNet · Art-XL2.0 | avg L1: 0.43 -> 0.42 | Helped |
| add geodesic-aware bone probability prediction module | HeterSkinNet · MR | avg L1: 0.34 -> 0.28 | Helped |
Surprising findings
- UniRig has comparable token length (140.26) to the proposed tokenizer (142.01), so the tokenization gain is mainly in connectivity rather than sequence compression.
Limitations
- Fails to generalize to snake-like meshes due to lack of such training samples.
- Relies on heuristic-free but fixed-threshold preference pair construction (threshold 3, margin 0.5) and a reward that is hand-designed.
- Baselines MagicArticulate and UniRig are evaluated with public pretrained weights, and MagicArticulate and UniRig cannot be retrained on MR, limiting controlled comparison.
Experimental setup
Datasets: Articulation-XL2.0 ModelsResource
Benchmarks: Articulation-XL2.0 ModelsResource
Metrics: CD-J2J CD-J2B CD-B2B IoU Precision Recall avg L1 avg Dist
Baselines: RigNet MagicArticulate (Magic-hier) MagicArticulate (Magic-spatial) UniRig NeuroSkinning HeterSkinNet
Builds on: RigNet MagicArticulate UniRig Michelangelo DPO
Closest papers by meaning
- BlockGCN: Redefine Topology Awareness for Skeleton-Based Action Recognition (CVPR 2024 · similarity 0.85)
- KITRO: Refining Human Mesh by 2D Clues and Kinematic-tree Rotation (CVPR 2024 · similarity 0.82)
- HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset (CVPR 2025 · similarity 0.82)
- Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization (NeurIPS 2025 · similarity 0.81)
- Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.81)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.