Researchers release TorchCurves for differentiable parametric curves

TorchCurves, a PyTorch library for learnable B-spline and Legendre curves, is available on GitHub.

Researchers present TorchCurves, a PyTorch library that treats vector-valued parametric curves as learnable primitives. Its core operation maps batches of scalar inputs to points on B-spline or Legendre polynomial curves and differentiates through both inputs and coefficients. Coefficients can be optimized with a surrounding model or produced by another network. The library offers trainable modules, reusable basis modules, and a functional interface for continuous embeddings, Kolmogorov-Arnold networks, and functional regression. Timing tests compare curve evaluation with NumPy and SciPy and KAN training with EfficientKAN; the authors say overhead is practical for some sizes and substantial for others. Code is on GitHub.