CubicSplat: Differentiable Vector Graphics via Error-Bounded Forward Relaxation

CubicSplat is a new differentiable vector rasterizer built around error-bounded forward relaxation, and it goes after the usual tradeoff between exact-looking rendering and usable gradients. Instead of relying on Bézier closest-point solvers, it uses uniform polyline surrogates with geometric error bounded at O(S^-2), which keeps the computation graph static and the gradients better behaved.

That matters because differentiable vector graphics has often been fragile as scene complexity rises: the more exact the forward pass gets, the more the gradient signal can degrade, and vice versa. CubicSplat’s approach is meant to reduce that “gradient seesaw” by construction, while a compositing-based visibility mechanism removes degenerate primitives without extra regularization tricks.

The paper reports state-of-the-art reconstruction results on DIV2K and Kodak, including more than a 2 dB PSNR gain in the closed-fill setting, alongside training runs up to 4x faster than prior methods. For teams working on vector-based editors, stylized rendering, or optimization-driven asset pipelines, the practical appeal is obvious: fewer heuristics, more predictable convergence, and a path toward editable vector primitives that scale better in training.

The code is available now, and the work sits at the intersection of graphics, vision, and ML—exactly where differentiable rendering tools are starting to matter for production workflows as well as research prototypes.