bbsolver: A Unified Error-Bounded Spatiotemporal Optimization Solver for Key Timing and Topology-Consistent Vector Paths

bbsolver is trying to solve a practical pain point in animation pipelines: dense sampling is great for capture and validation, but terrible as a final editable representation. The paper frames the problem as tolerance-bounded spatiotemporal reduction, so timing and shape are optimized together instead of with separate reducers that can fight each other.

The workflow is meant to be host-agnostic. A DCC or host app such as After Effects or Blender exports sampled animation and path data into a JSON bundle, the standalone solver computes sparse keys plus interpolation metadata, and the result is only accepted if replayed samples stay within the requested worst-case error. That makes it easier to preserve point identity, vertex structure, and a single error budget across both motion and topology.

The reported numbers are strong for production-style data. In After Effects validation, a DUIK humanoid walk cycle went from 12,684 samples to 540 keys at epsilon=1, and an ant rig from 11,956 samples to 653 keys, with maximum errors below 1 px and 1 degree. For Blender-sampled FBX mocap retargeting, the solver reached 214 keys from 13,455 samples at epsilon=3, and the paper says tuned baselines needed 4.5x to 27.5x more scalar key entries to match accuracy.

For vector paths, the solver handles constant-vertex streams directly and also provides diagnostics for variable-vertex-count cases. The paper cites a 6.7x After Effects-compatible procedural-path compression and exact transition-timing recovery in one diagnostic case. For teams building animation tools, this looks less like a...