Research map / 3D vision / Face and shape

Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation

NeurIPS 2026 · 3D vision › Face and shape · approach: Textured mesh generation · 0 citations (OpenAlex)

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.

Rigel3D jointly generates image-conditioned rigged 3D assets (mesh, skeleton, skinning weights) using coupled surface and skeleton structured latents, plus an open-vocabulary CLIP-space joint labeler.

Problem

Generated 3D assets are static and lack skeletons, joint hierarchies and skinning weights needed for animation; existing auto-rigging is post-hoc and joint labels are closed-set or generic.

Approach

Extend TRELLIS structured latents with a second, skeleton SLat over bone-voxels, decode mesh, autoregressive skeleton and point-bone attention skinning jointly, and train two-stage rectified-flow generators for both latent sets; joints are embedded into frozen OpenCLIP text space for open-vocabulary labeling.

Key results

• On ArticulationXL, J2J drops from 10.667 (AniGen) to 7.168 (32.8% relative reduction), J2B from 9.715 to 6.036, B2B from 8.601 to 5.976.
• On the Anymate test split, Rigel3D beats Anymate and Puppeteer on all reported skeleton and skinning metrics (e.g., J2J 7.364 vs 7.834 for Anymate).
• Joint labeling with autoregressive BFS ordering reaches 85.90% top-1 accuracy versus 81.31% for the non-autoregressive TRELLIS-SLat variant.

Surprising findings

• AniGen obtains slightly lower skinning errors than Rigel3D on ArticulationXL, and Anymate is slightly better on J2J-B on ModelsResource.
• Rigel3D is slightly worse than TRELLIS on CLIP similarity, LPIPS and Chamfer distance, although it improves FD_dinov2 and KD_dinov2.

Limitations

• Generated rigs may have missing joints, spurious branches or imperfect connectivity.
• Skinning is not supervised by motion sequences, so deformations may be unnatural under extreme poses.
• Open-vocabulary labels are ambiguous for repeated structures such as fingers, tails or appendages.

Experimental setup

Datasets: Anymate ArticulationXL ModelsResource

Benchmarks: Anymate test split ModelsResource test split ArticulationXL

Metrics: J2J J2B B2B L1 L2 KL PSNR LPIPS Chamfer distance F-score

Baselines: Anymate Puppeteer AniGen TRELLIS

Builds on: TRELLIS Anymate FlexiCubes RigNet

Closest papers by meaning

• Puppeteer: Rig and Animate Your 3D Models (NeurIPS 2025 · similarity 0.92)
• Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.88)
• MagicArticulate: Make Your 3D Models Articulation-Ready (CVPR 2025 · similarity 0.88)
• PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
• MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.88)
• 3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code (NeurIPS 2026 · similarity 0.87)