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

How to read this

Category

joint generative model for rigged 3D asset synthesis

Contributions

• Introduces Rigel3D, which jointly generates mesh geometry, skeleton topology, joint coordinates, and skinning weights together, instead of the standard post-hoc pipeline of generating a static shape then auto-rigging it afterward
• Extends TRELLIS's Structured Latent representation with two coupled latents, a surface latent for geometry and appearance and a skeleton latent for articulation, decoded by a rig-aware autoencoder with a skeleton-conditioned mesh decoder, an autoregressive skeleton decoder, and an attention-based skinning decoder
• Adds an open-vocabulary joint labeling module that embeds generated joints into a shared vision-language space, so arbitrary generated skeletons can match existing retargeting templates rather than a fixed closed-set label list
• Reports state-of-the-art results over prior auto-rigging baselines on two datasets, Anymate and ModelsResource

Context

Rigel3D explicitly positions itself against classical and learning-based auto-rigging methods such as Pinocchio, RigNet, TARig, and recent template-free riggers, all described as post-processing steps applied after shape generation, arguing that separation breaks down for generative outputs whose geometry can diverge far from a downstream rigger's training distribution. It builds directly on TRELLIS's sparse-voxel Structured Latent representation, extending it from a shape-only representation to a coupled shape-plus-skeleton one.

Correctness

The claim of outperforming prior auto-rigging baselines rests on quantitative comparison across multiple metrics on the Anymate and ModelsResource datasets, per the abstract, though the specific metrics and margins sit in sections not read in this pass. Because the method is generative and image-conditioned, rig quality is bounded by both datasets' coverage of rig topologies and by how well the open-vocabulary joint labeling generalizes to skeleton structures outside those training sets, which the paper does not claim to solve universally.

Clarity

The abstract and introduction are clearly structured, laying out the problem of post-hoc versus joint rigging, the technical mechanism of coupled latents decoded by three specialized decoders, and the practical payoff of open-vocabulary joint labels for retargeting in a logical progression a rigger can follow without a generative-modeling background, though the underlying Structured Latent machinery requires the TRELLIS paper as prerequisite reading.

How to read it

First pass, read the abstract, Figure 1, and the contribution bullets to see what jointly generated means concretely: mesh, skeleton, and skin weights from one model. Second pass, read the introduction's framing of why post-hoc rigging breaks down for generative outputs, and skim TRELLIS background if unfamiliar with Structured Latents, since the whole method depends on that representation. Third pass, read the results section's comparison tables against Pinocchio, RigNet, and TARig-style baselines on Anymate and ModelsResource, and check the open-vocabulary joint labeling section to see whether it could plug into an existing retargeting template library.