TopoCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Animation

TopoCap proposes a unified video-to-animation pipeline that doesn’t assume a fixed body template like SMPL. Instead of treating each creature or prop as a special case, it learns a shared motion latent that can be decoded for different skeletal structures. The authors frame this as topology-agnostic motion priors: motion is continuous even if the rig graph is discrete and messy.

The system has two stages. First, a Graph CVAE learns a Universal Motion Manifold, compressing heterogeneous kinematic chains into a fixed-length code while conditioning the decoder on a structural embedding of the target rig. Second, a conditional flow-matching model predicts those codes from monocular video features, so the animation transfer happens without test-time optimization. That matters because optimization-heavy retargeting often becomes the bottleneck in practical pipelines.

The dataset angle is also important. The team introduces Mobjaverse, curated from Objaverse-XL, with more than 5,000 unique skeletal topologies and 2 million frames. The paper says that is roughly two orders of magnitude more structurally diverse than existing datasets, which is the kind of scale jump you usually need before a generalized prior starts to hold up outside human-only demos.

For game developers, the obvious upside is faster motion capture and retargeting for long-tail creatures, props, and hybrid characters where manual rigging or template fitting is expensive. The paper reports better results than specialist models on human and quadruped benchmarks, plus zero-shot retargeting on unseen topologies....