NURBS Splatting: A Unified Differentiable Rendering Framework for Vector Graphics

arXiv cs.GR details NURBS Splatting, a unified differentiable rendering framework for vector graphics that targets planar rational splines rather than just Bézier curves. The paper frames curves as continuous Gaussian fields, then samples along the parameter domain and inside filled regions to turn rendering into smooth accumulation with stable gradients.

That matters because differentiable vector renderers are increasingly useful for inverse graphics, asset reconstruction, and optimization-driven design workflows. The authors say the method handles long splines, rational weights, non-uniform knots, and closed-region filling, all of which are common pain points when trying to keep gradients usable across complex vector content.

The work is positioned for calligraphy reconstruction, vectorization pipelines, and long-spline image abstraction, where stability can be as important as raw image quality. For game teams, the practical angle is strongest on tools and content creation: anything that improves editable vector reconstruction or procedural shape fitting can reduce manual cleanup and make hybrid raster/vector workflows less brittle.

The paper was submitted on 30 Jun 2026 and revised on 29 Sep 2026, with an ECCV 2026 journal reference listed. Specific implementation details and runtime characteristics are not the headline here, but the broader direction is clear: more of vector graphics may become differentiable enough to optimize directly, rather than approximate with fragile rasterization tricks.