PlantRig - From Bones to Branches: Adaptation of Autoregressive Rigging Models for Plant Skeletal Reconstruction

Autoregressive rigging systems built for articulated characters are being tested on something much messier: plants. The work evaluates UniRig and SkinTokens on synthetic L-system trees and scanned specimens covering monopodial, sympodial, whorled, and vine-like forms, then uses that baseline to probe where the models break down.

The first pass exposed a practical failure mode. UniRig tended to collapse complex branching into near-linear chains, while SkinTokens preserved more topology but produced unstable outputs and over-segmented branches. The team traced UniRig’s collapse to sampling-level suppression of branch tokens, and also found the frozen mesh encoder was not sensitive enough to structural variation, suggesting the bottleneck sits in tokenization and geometry encoding rather than only in learned priors.

From there, the model was fine-tuned over multiple rounds on procedurally generated synthetic datasets. That iterative training progressively restored branching structure and improved generalization to plants with foliage, which is a tougher case because leaves are effectively zero-thickness geometry and depend heavily on mesh normals. The final model handled a wider range of plant forms without any leaf-specific architectural changes.

For game teams, the practical takeaway is that character-rigging priors may be more reusable than expected if the training pipeline is adapted carefully. That matters for procedural vegetation, scan cleanup, and any workflow that needs skeletal structure inferred from messy organic meshes rather than hand-authored branch...