SPRig: Self-Supervised Pose-Invariant Rigging from Mesh Sequences

How to read this

Category

Self-supervised fine-tuning framework adding temporal and pose invariance to existing static autoregressive rigging models

Contributions

• Identifies that state-of-the-art autoregressive rigging models such as Puppeteer assume a fixed canonical T-pose and produce temporally flickering, inconsistent skeletons and skinning when applied frame by frame to dynamic sequences that lack a canonical rest pose
• Proposes a frozen-teacher, trainable-student fine-tuning scheme where a pretrained model infers a high-quality rig on one canonical anchor frame, which then supervises the same model's predictions on the sequence's other frames, using autoregressive token-space constraints plus a Procrustes-aligned geometry-space loss for skeletons
• Introduces an articulation-invariant consistency loss with barycentric-tracked correspondences, consistency distillation, and entropy or geometric-proximity regularization for skinning, reporting improved temporal coherence without sacrificing, and sometimes improving, per-frame static rigging quality on out-of-domain datasets

Context

SPRig builds explicitly on top of static autoregressive rigging models such as UniRig and Puppeteer, extending them into the dynamic-sequence setting alongside other rigging-from-motion work like MoRig and LASR, and distinguishes itself as a fine-tuning wrapper rather than a rigging model trained from scratch.

Correctness

Improvements in temporal coherence and preserved static quality rest on experiments described as extensive over dynamic sequence datasets and out-of-domain static benchmarks, but this is an arXiv preprint, and since it fine-tunes on top of a frozen teacher, its ceiling is bounded by the quality of that base model's own per-frame predictions.

Clarity

The motivation, flickering rigs caused by the lack of a canonical pose in dynamic sequences, is easy to follow even without prior background in autoregressive rigging. The loss formulations (Procrustes alignment, barycentric tracking, consistency distillation) are dense and assume familiarity with the underlying Puppeteer or UniRig token representation.

How to read it

First pass: abstract and Figure 1's comparison of consistent skeletons and smoother skinning against Puppeteer. Second pass: the introduction and related work for how it differs from static autoregressive rigging and prior rigging-from-motion approaches. Third pass: the skeleton and skinning consistency loss formulations and the experimental section's coherence and quality metrics.