Rigging — building the internal skeleton that lets a 3D character or object deform and animate — is one of the more stubbornly manual steps in character pipelines, and most attempts to automate it split into two separate problems: generating a skeleton for a single static mesh, versus generating temporally consistent skeletons across an entire animated sequence. FlowRig, a new framework from Texas A&M researchers Fengzhi Guo, Nate Morrill, Shiqi Li, and Cheng Zhang, treats both as the same problem, solved by the same model.
One feedforward pass, two very different inputs
FlowRig is built on flow matching, a generative modeling technique, and is explicitly template-free — it isn’t matching an input mesh against a fixed, predefined bone hierarchy the way a lot of production rigging tools do, which makes it more adaptable to unusual or non-humanoid shapes. Given a single static mesh, it produces a rig in one pass; given a full 4D mesh sequence (a mesh that changes shape over time, like an animated character), it produces temporally aligned skeletons across the whole sequence, also in one feedforward pass rather than frame-by-frame reconstruction.
The system also supports explicit skeleton-count control — letting a user specify how many bones they want rather than accepting whatever the model decides — plus uncertainty estimation for picking among multiple candidate rigs, and masking-based interpolation, forecasting, and backcasting. That last capability is the more unusual one: given a partial sequence, FlowRig can predict what the skeleton’s motion looks like both in the gaps and beyond the frames it was actually shown, rather than only reconstructing motion it has direct data for.
How it was trained to generalize
The authors used a curriculum training strategy: starting from large-scale static 3D rigging data, where the volume of examples is high, then adapting the model to the comparatively scarcer mid-scale 4D mesh-skeleton data, where animated, temporally-labeled training examples are far harder to come by. That staged approach is a practical way to get a model good at the harder, data-poor 4D problem without needing an equally large 4D dataset to train it directly.
Open and checkable, which is the rarer part
Unlike a lot of fresh research still awaiting peer review, FlowRig ships with a public project page and GitHub repository (TAMU-Visual-AI/FlowRig) alongside the arXiv paper, including interactive 3D and 4D skeleton viewers that let you inspect results directly against ground truth and against Puppeteer, an existing rigging baseline the authors compare against. For riggers and technical animators, that combination — genuinely novel unification of static and 4D rigging, plus code you can actually run rather than just read about — is the more immediately useful signal than the benchmark numbers alone.