SCAMP: Sparse-anchor Control is One Small Projection

arXiv cs.GR details SCAMP, a sparse-anchor control method for text-to-motion generation that works through a frozen differentiable decoder instead of a separate conditioning branch. The core idea is to satisfy a few joint-and-frame constraints directly in anchor space, using damped Gauss-Newton with one dimensionless damping constant.

That matters because the control problem is tiny compared with the motion state: the paper frames it as a few hundred requested numbers against tens of thousands of latent values. By solving only in the anchor rows of the Jacobian, SCAMP keeps corrections orthogonal to the parts of motion the anchors do not constrain, which makes the solver cost scale with the request rather than the full clip.

The method is evaluated unchanged on seven published generators spanning diffusion, token, and latent designs. In those tests, it matches or exceeds released control methods on anchor error, and it can also close anchors on hosts that ship no dedicated control path. The authors also argue that decoder temporal support is a practical design choice for controllable motion systems, since windowed decoders and analytic recovery approaches show very different cost profiles.

The paper reports the authors’ own generator at 0.083 m anchor error with FID 0.102 in 0.50 s per clip. For game teams building animation tools or runtime motion systems, the takeaway is that sparse, editable motion constraints may be cheaper to enforce than many current pipelines assume, especially if the decoder architecture is chosen with control in mind.