Research map / 3D vision / 3D generation
MagicArticulate: Make Your 3D Models Articulation-Ready
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.
Introduces MagicArticulate, a two-stage framework that auto-regressively generates skeletons (OPT-350M over tokenized bone coordinates conditioned on point-cloud shape tokens) and predicts skinning weights via functional diffusion with volumetric geodesic priors, plus the 33k-model Articulation-XL dataset.
Problem
Automatically converting static 3D models into articulation-ready assets (skeletons and skinning weights), which is labor-intensive manually and lacks large-scale benchmarks for learning-based methods.
Approach
Reformulates skeleton generation as autoregressive sequence modeling of bones (discretized joint coordinates) conditioned on shape tokens, and predicts skinning weights as residuals over volumetric geodesic priors using a functional diffusion process, with no orientation assumptions.
Key results
- Skeleton generation on Articulation-XL: CD-J2J 2.586 vs RigNet 7.478 and Pinocchio 8.360 (x10^-2).
- Skinning weights on Articulation-XL: avg L1 0.337 vs RigNet 0.698 and GVB 0.724; precision 80.7% vs 72.4% (RigNet).
- Cross-dataset generalization: trained on Articulation-XL and tested on ModelsResource, CD-J2J 4.103 vs RigNet trained in-domain 4.143 (x10^-2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MagicArticulate instead of RigNet | Skeleton generation · Articulation-XL | CD-J2J: 7.478 -> 2.586 | Helped |
| Use MagicArticulate instead of RigNet | Skinning weight prediction · Articulation-XL | Precision: 72.4% -> 80.7% | Helped |
| Use MagicArticulate instead of RigNet | Skinning weight prediction · ModelsResource | Precision: 77.1% -> 82.1% | Helped |
| Use MagicArticulate instead of RigNet | Skeleton generation trained on Articulation-XL and tested on ModelsResource · ModelsResource | CD-J2J: 7.132 -> 4.103 | Helped |
| Add VLM-based data filtering (ablation) | MagicArticulate skeleton generation · Articulation-XL | CD-J2J: 2.982 -> 2.586 | Helped |
| Balance training data across categories (ablation) | MagicArticulate skeleton generation · Articulation-XL | CD-J2J: 2.586 -> 2.691 | Hurt |
| Add volumetric geodesic distance initialization (ablation) | MagicArticulate skinning weight prediction · ModelsResource | Precision: 81.5% -> 82.1% | Helped |
| Add volumetric geodesic distance initialization (ablation) | MagicArticulate skinning weight prediction · ModelsResource | Recall: 77.7% -> 81.6% | Helped |
| Normalize skinning weights and geodesic distances to [−1, 1] before noise addition (ablation) | MagicArticulate skinning weight prediction · ModelsResource | avg L1: 0.436 -> 0.398 | Helped |
| Add global shape features (ablation) | MagicArticulate skinning weight prediction · ModelsResource | avg L1: 0.412 -> 0.398 | Helped |
Surprising findings
- Balancing category distribution by replicating non-human data gave no improvement over the original training, attributed to humanoid dominance in the test set.
- RigNet trained on Articulation-XL and tested on ModelsResource degrades to worse than template-based Pinocchio.
Limitations
- Skeleton generation prioritizes accuracy with a cap of 100 bones; models with more than 100 bones were filtered out.
- Skinning weight evaluation on Articulation-XL excluded samples with more than 55 joints.
- Category balancing did not help, and the test set is dominated by humanoid data.
Experimental setup
Datasets: Articulation-XL ModelsResource Objaverse-XL
Benchmarks: Articulation-XL ModelsResource
Metrics: CD-J2J CD-J2B CD-B2B precision recall avg L1
Baselines: Pinocchio RigNet GVB SkinningNet
Builds on: MeshGPT functional diffusion DDPM RigNet
Closest papers by meaning
- Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation (NeurIPS 2026 · similarity 0.88)
- Puppeteer: Rig and Animate Your 3D Models (NeurIPS 2025 · similarity 0.88)
- Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters (CVPR 2025 · similarity 0.87)
- Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model (ICLR 2025 · similarity 0.87)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.87)
- LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example (CVPR 2024 · similarity 0.86)
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.