Research map / 3D vision / Gaussian splatting
DisC-GS: Discontinuity-aware Gaussian Splatting
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.
Proposes DisC-GS, which pre-scissors projected 2D/3D Gaussians along cubic Bézier boundaries (with a learned per-Gaussian control-point attribute) and uses a Bézier-boundary gradient approximation to keep rendering differentiable, improving novel view synthesis on Mip-NeRF360, Tanks&Temples and Deep Blending.
Problem
Gaussian Splatting renders sharp discontinuities and boundaries poorly because continuous Gaussians spill over edges, and the resulting discontinuous indicator makes rendering non-differentiable for novel view synthesis.
Approach
Attach per-Gaussian cubic Bézier control points that are projected to the image plane to define boundary curves; each Gaussian is scissored by an indicator in the alpha-blending, and a gradient approximation based on solving for the control-point value that flips the indicator restores gradients.
Key results
- On Tanks&Temples, the 2D GS + DisC-GS variant reaches 0.866 SSIM, 24.96 PSNR and 0.120 LPIPS, versus 23.96 PSNR for FreGS (best prior in the table).
- On Mip-NeRF360, the method reports 28.01 PSNR and 0.189 LPIPS, above Mip-Splatting's 27.79 PSNR and 0.203 LPIPS.
- Applying the framework to 3D Gaussian Splatting raises Tanks&Temples PSNR from 23.14 to 24.67 and lowers LPIPS from 0.183 to 0.123.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DisC-GS to 2D Gaussian Splatting (ablation) | 2D Gaussian Splatting · Tanks&Temples | PSNR: 23.30 -> 24.96 | Helped |
| apply DisC-GS to 3D Gaussian Splatting (ablation) | 3D Gaussian Splatting · Tanks&Temples | PSNR: 23.14 -> 24.67 | Helped |
| use 2 control points per B e9zier curve instead of 4 (ablation) | DisC-GS · Tanks&Temples | PSNR: 24.96 -> 24.14 | Hurt |
| use 3 control points per B e9zier curve instead of 4 (ablation) | DisC-GS · Tanks&Temples | PSNR: 24.96 -> 24.58 | Hurt |
| use 5 control points per B e9zier curve instead of 4 (ablation) | DisC-GS · Tanks&Temples | PSNR: 24.96 -> 24.68 | Hurt |
| use discontinuity-aware rendering with B e9zier boundaries | 2D Gaussian Splatting · Tanks&Temples | PSNR: 23.14 -> 24.96 | Helped |
| use discontinuity-aware rendering with B e9zier boundaries | 2D Gaussian Splatting · Mip-NeRF360 | PSNR: 27.79 -> 28.01 | Helped |
| use discontinuity-aware rendering with B e9zier boundaries | 2D Gaussian Splatting · Deep Blending | PSNR: 29.93 -> 30.42 | Helped |
Surprising findings
- Using 4 control points per Bézier curve is best; 5 control points performs slightly worse than 4.
Limitations
- Challenges in rendering large scenes, shared with existing Gaussian Splatting methods.
- Gradient approximation is heuristic, relying on linear interpolation toward the nearest root of a cubic equation.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS
Baselines: Plenoxels INGP-Base INGP-Big Mip-NeRF360 3D-GS Surfsplatting FreGS GES Mip-Splatting
Builds on: 3D Gaussian Splatting 2D Gaussian Splatting Mip-Splatting
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- Sort-free Gaussian Splatting via Weighted Sum Rendering (ICLR 2025 · similarity 0.90)
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.