The `voxel_skin` function propagates skinning weights from mesh vertices to a voxel grid using a graph-based shortest-path approach. It constructs a combined graph of mesh vertices, faces, grid points, and bone joints, then computes geodesic distances to assign weights. This is a reusable, self-contained algorithm for any mesh-to-voxel skinning transfer, demonstrated by its use in post-processing and vertex group extraction.

Provides a comprehensive set of augmentation classes for 3D character data, including random joint deletion, subtree dropping, skeleton collapse, linear/affine transformations, LBS pose perturbation, joint discretization, and vertex/normal jitter. Each augment is a dataclass with a `transform(asset)` method, making them composable and configurable via YAML/JSON. This pattern is reusable for any 3D mesh+skeleton data pipeline.

Implements a tokenizer for skeleton hierarchies that encodes joint positions as discrete coordinates, uses branch tokens for non-sequential joints, and inserts part tokens (e.g., 'body', 'hand') to denote anatomical regions. The tokenizer supports multiple skeleton templates (Mixamo, VRoid) and produces a reversible token sequence. It is a self-contained module that can be adapted to other skeleton representation tasks.

Implements Finite Scalar Quantization (FSQ), a learned discrete representation method that quantizes latent vectors to a fixed codebook via rounding in a bounded space. It supports multi-codebooks, projection layers, and straight-through gradient estimation. The module is extracted from the original FSQ paper and is reusable for any discrete representation learning task (e.g., VQ-VAE alternatives).