Skin Tokens: A Learned Compact Representation for Unified Autoregressive Rigging

How to read this

Category

Discrete tokenized representation for skinning weights feeding a unified autoregressive rig generator

Contributions

• Introduces SkinTokens, a discrete compact representation for per-bone skinning weights learned via an FSQ-CVAE, reframing skinning weight prediction from high-dimensional sparse regression into token sequence prediction
• Builds TokenRig, a single autoregressive transformer that interleaves skeletal parameters and SkinTokens in one sequence, jointly modeling skeleton and skin instead of the decoupled two-stage pipelines common in prior autoregressive rigging work
• Adds a reinforcement learning fine-tuning stage with geometric and semantic reward functions to improve generalization to out-of-distribution assets, reporting 98 to 133 percent skinning accuracy improvement and 17 to 22 percent bone prediction improvement over the compared prior methods

Context

The paper is built directly against the decoupled autoregressive-skeleton-plus-separate-skinning-regression paradigm (RigNet, MoRig, and recent autoregressive skeleton generators), and against prior geodesic or weight-field skinning methods such as SkinCells, positioning joint modeling and RL refinement as its central departure from that lineage.

Correctness

The large percentage improvements are the paper's own benchmark comparisons against specific baselines and metrics, so they should be read in that context rather than as universal gains. The version read carries placeholder ACM DOI and article-number fields, indicating it had not yet finished the peer-review or camera-ready process at the time of this note.

Clarity

Well organized and readable, with an unusually thorough related-work taxonomy of traditional, learning-based, and autoregressive rigging methods that is useful as a standalone survey of the auto-rigging landscape.

How to read it

First pass: abstract, Figure 1 teaser, and the three bullet contributions. Second pass: Sections 2.1 and 2.2 to map the auto-rigging and skinning-weight landscape, then the Section 3.1 overview. Third pass: Section 3.2's FSQ-CVAE formulation and the experimental tables for the accuracy numbers.