Research map / 3D vision / Gaussian splatting
Bézier Splatting for Fast and Differentiable Vector Graphics Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Bézier Splatting vectorizes images by sampling 2D Gaussians along Bézier curves and rasterizing them with a Gaussian splatting pipeline, plus an adaptive pruning and densification strategy, achieving much faster optimization than DiffVG.
Problem
Differentiable vector graphics rasterization (e.g., DiffVG) is slow to optimize and yields low-fidelity results for high-resolution images, and learning-based VG synthesis fails to generalize out of domain.
Approach
Replace DiffVG's boundary-sampling and Reynolds-transport gradient computation with splatting of 2D Gaussians sampled along Bézier curves (and interpolated curves inside closed regions), using alpha-blending and an adaptive curve pruning/densification scheme that reallocates curves to high-error regions.
Key results
- Open curves: forward rasterization 31.4x and backward 149.2x faster than DiffVG on a 2040x1344 image with 2048 curves (RTX 4090).
- On DIV2K with 1024 open curves, PSNR 25.45 vs DiffVG 22.62 with optimization 3.2 min vs 30.6 min.
- Closed curves: forward 6.0x and backward 18.2x faster than DiffVG; on DIV2K 1024 closed curves, PSNR 23.45 vs LIVE 23.11, optimization 8.6 min vs 5.1 h for LIVE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Bézier Splatting instead of DiffVG | Open-curve vectorization · DIV2K, 1024 curves | PSNR: 22.62 -> 25.45 | Helped |
| Use Bézier Splatting instead of DiffVG | Open-curve vectorization · DIV2K, 1024 curves | optimization time: 30.6min -> 3.2min | Helped |
| Use Bézier Splatting instead of DiffVG | Closed-curve vectorization · DIV2K, 1024 curves | PSNR: 22.95 -> 23.45 | Helped |
| Use Bézier Splatting instead of DiffVG | Closed-curve vectorization · Kodak, open curves, 1024 curves | PSNR: 26.04 -> 29.24 | Helped |
| Use Bézier Splatting instead of DiffVG | Closed-curve vectorization · Kodak, closed curves, 1024 curves | SSIM: 0.719 -> 0.708 | Hurt |
| Use Bézier Splatting instead of DiffVG | Closed-curve vectorization · Kodak, closed curves, 1024 curves | PSNR: 26.66 -> 26.91 | Helped |
| Use Bézier Splatting instead of DiffVG | Open-curve rasterization · 2,040×1,344 image, 2,048 curves | forward rasterization time: 141.3ms -> 4.5ms | Helped |
| Use Bézier Splatting instead of DiffVG | Open-curve rasterization · 2,040×1,344 image, 2,048 curves | backward rasterization time: 701.3ms -> 4.7ms | Helped |
| Add adaptive pruning and densification (ablation) | Bézier Splatting · DIV2K, closed curves, 512 curves | PSNR: 21.10 -> 22.11 | Helped |
| Add adaptive pruning and densification (ablation) | Bézier Splatting · DIV2K, closed curves, 512 curves | SSIM: 0.590 -> 0.607 | Helped |
| Add layer-wise curve addition during training (ablation) | Bézier Splatting · DIV2K, closed curves, 512 curves | SSIM: 0.607 -> 0.613 | Helped |
| Add layer-wise curve addition during training (ablation) | Bézier Splatting · DIV2K, closed curves, 512 curves | optimization time: 8.3 min -> 16.2 min | Hurt |
Surprising findings
- For closed curves on DIV2K, LIVSS achieves higher SSIM and LPIPS than Bézier Splatting at some curve counts despite lower PSNR.
- Layer-wise vectorization gives slightly better metrics but doubles training time compared to the proposed pruning and densification.
Limitations
- Closed curves are constrained to be convex via Xing loss, which may reduce representational capacity.
- Closed curves need more sampled Gaussians (about 20 extra curves per closed curve), making them slower than open curves.
- Evaluated mainly on image vectorization; text-to-VG synthesis not explored.
Experimental setup
Datasets: DIV2K Kodak Clipart1K DanbooRegion
Benchmarks: DIV2K Kodak
Metrics: SSIM PSNR LPIPS optimization time forward/backward runtime
Baselines: DiffVG LIVE LIVSS Adobe Image Trace
Builds on: DiffVG 3D Gaussian Splatting LIVE GaussianImage
Closest papers by meaning
- BG-Triangle: Bezier Gaussian Triangle for 3D Vectorization and Rendering (CVPR 2025 · similarity 0.87)
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.85)
- Anti-Aliased 2D Gaussian Splatting (NeurIPS 2025 · similarity 0.85)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.84)
- Deformable Radial Kernel Splatting (CVPR 2025 · similarity 0.84)
- 3D Skew Normal Splatting (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.