NumMotion: Precise Numerical Control for 3D Human Motion Generation

Existing generative 3D human motion models can generate realistic movements but struggle to meet numerical targets. This limits their use for controlled biomechanical comparisons, which require varying specific movement characteristics to prescribed values while minimizing unintended changes to other properties. We propose NumMotion, a lightweight adaptation framework for precise numerical control of movement characteristics in text-conditioned 3D human motion generation. NumMotion conditions a frozen generator on explicit numerical inputs through a compact adapter. Target fitting matches motion measurements to requested values, while Jacobian regularization encourages proportional responses to changed requests and penalizes unintended changes in other controlled properties. We introduce a benchmark based on NimblePhysics covering seven action categories, with two numerical controls per category, to evaluate simultaneous target success, cross-control interference, and motion quality. We demonstrate that NumMotion meets the requested numerical target in 94% of trials versus 8.8% for baseline MDM.