This arXiv preprint proposes one feedforward model, built on flow matching, that generates skeletons for both static 3D meshes and animated 4D mesh sequences.
What it does: Produces temporally aligned skeletons for a whole sequence in a single pass, and also handles single-mesh 3D rigging.
Controls: Supports explicit skeleton-count control, uncertainty estimates for picking candidates, and masking-based interpolation, forecasting and backcasting.
Training: Starts from large-scale 3D rigging data, then adapts to mid-scale 4D mesh-skeleton data, since paired 4D data is limited.
Results: The abstract claims strong static 3D rigging with controllable bone complexity, plus temporally consistent 4D results. It names the method MoTWorld, not FlowRig as in the title.
Sign in to comment.
TL;DR of the paper, generated automatically.
This arXiv preprint proposes one feedforward model, built on flow matching, that generates skeletons for both static 3D meshes and animated 4D mesh sequences.