Reproduce “CubicSplat: Differentiable Vector Graphics via Error-Bounded Forward Relaxation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenglong Liu, Xin Zhang, Yimeng Zhu, Liyang He, Yixiao Ma, Yu Su, Zhenya Huang, Qi Liu
CubicSplat addresses image vectorization by fitting editable Bézier curves to raster images. It replaces closest-point solving with uniformly sampled polylines and computes pixel distances in closed form, using soft coverage for gradient-based optimization. On 200-image DIV2K, 1,024 closed-fill curves achieve 25.78 dB PSNR versus 23.45 for Bézier Splatting, in 2.1 versus 8.6 minutes per image; open-curve LPIPS favors Bézier Splatting.
Key insights
- The "gradient seesaw" problem is a fundamental challenge in differentiable vector graphics, where design choices for forward accuracy often compromise backward gradient quality and stability.
- Replacing iterative Bézier distance solvers with error-bounded polyline approximations provides a stable and well-conditioned gradient oracle, mitigating ill-conditioning and aliasing without sacrificing geometric fidelity after optimization.
- The polyline sampling density (S) primarily functions as a "forward-relaxation knob" for oracle conditioning and runtime, with reconstruction quality showing diminishing returns beyond a modest threshold, allowing for efficient training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.