Research map / 3D vision / Gaussian splatting
S2Gaussian: Sparse-View Super-Resolution 3D 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 S2Gaussian, a two-stage 3D Gaussian Splatting framework that reconstructs high-resolution scenes from sparse low-resolution views using Gaussian Shuffle Split densification, pseudo-view super-resolution supervision, blur-free inconsistency modeling, and gradient-flag 3D robust optimization.
Problem
Reconstructing high-quality 3D scenes when input views are both sparse (few viewpoints) and low-resolution, a combination prior methods handle only separately.
Approach
Initialize high-resolution Gaussians by splitting a depth-regularized low-resolution Gaussian field into six sub-Gaussians, then refine them with super-resolved sparse and pseudo views, using a learnable inconsistency module with blur-kernel modeling and a per-Gaussian gradient-flag optimizer to suppress erroneous updates.
Key results
- On Mip-NeRF 360 x4 (24 views), S2Gaussian reaches 22.05 PSNR, 2.22 dB above the FSGS+SRGS combination (19.83).
- On LLFF x4 (3 views), S2Gaussian achieves 20.45 PSNR and 0.139 LPIPS, versus 19.49 and 0.175 for FSGS+SRGS.
- Ablation on Mip-NeRF 360 shows each component adds PSNR, from a 19.79 baseline to 22.05 with 3D robust optimization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a two-stage optimization process (ablation) | S2Gaussian · Mip-NeRF 360 ×4 (24 views) | PSNR: 19.79 -> 20.06 | Helped |
| Add Gaussian Shuffle Split (ablation) | S2Gaussian · Mip-NeRF 360 ×4 (24 views) | PSNR: 20.06 -> 20.62 | Helped |
| Add blur-free inconsistency modeling (ablation) | S2Gaussian · Mip-NeRF 360 ×4 (24 views) | PSNR: 20.62 -> 21.12 | Helped |
| Add 3D robust optimization (ablation) | S2Gaussian · Mip-NeRF 360 ×4 (24 views) | PSNR: 21.12 -> 22.05 | Helped |
| Vary the Gaussian Shuffle Split sub-Gaussian count across 4, 5, 6, 7, 8, and 9 (ablation) | S2Gaussian · Mip-NeRF 360 | PSNR: 21.76; 21.95; 22.05; 22.03; 22.09; 22.08 | Mixed |
Limitations
- Relies on an off-the-shelf 2D super-resolution model that cannot guarantee multi-view consistency.
- Initialization uses SfM; the authors note that DUSt3R-style initialization is left for future work.
Experimental setup
Datasets: Blender LLFF Mip-NeRF 360
Benchmarks: Blender x4 LLFF x4 Mip-NeRF 360 x4
Metrics: PSNR SSIM LPIPS FID
Baselines: NeRF-SR RegNeRF SparseNeRF 3DGS SRGS Mip-Splatting DNGaussian FSGS FSGS+SRGS
Builds on: 3D Gaussian Splatting FSGS SRGS ResShift DPT
Closest papers by meaning
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction (ICLR 2025 · similarity 0.90)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.90)
- LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS (NeurIPS 2024 · similarity 0.90)
- Superpoint Gaussian Splatting for Real-Time High-Fidelity Dynamic Scene Reconstruction (ICML 2024 · 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.