Research map / 3D vision / Gaussian splatting
SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction
Zicheng Zhang, Xiangting Meng, Ke Wu, Wenchao Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace pixel-aligned prediction with entropy-based adaptive sampling (150k Gaussians vs 688k) | SparseSplat vs DepthSplat · DL3DV test | PSNR: 24.17 -> 24.20 | No clear effect |
| replace pixel-aligned prediction with adaptive sampling | SparseSplat vs DepthSplat · DL3DV test | LPIPS: 0.152 -> 0.168 | Hurt |
| replace pixel-aligned prediction with adaptive sampling (150k) | SparseSplat vs DepthSplat · DL3DV test | inference time: 0.128 s -> 0.398 s | Hurt |
| reduce Gaussians to 10k via temperature | SparseSplat · DL3DV test | PSNR: 24.20 -> 21.29 | Hurt |
| use SparseSplat over AnySplat at sparse budgets | SparseSplat vs AnySplat · DL3DV test | PSNR: 8.91 -> 21.29 | Improved |
| add entropy-based sampling instead of random (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 21.50 -> 22.36 | Improved |
| add entropy-based sampling instead of Laplacian (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 21.95 -> 22.36 | Improved |
| add KNN neighborhood (K=20) instead of single point (K=0) (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 21.53 -> 22.36 | Improved |
| increase K from 20 to 30 (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 22.36 -> 22.38 | No clear effect |
| use geo-aware attention head instead of MLP (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 21.47 -> 22.36 | Improved |
| use geo-aware attention head instead of graph convolution (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: 20.78 -> 22.36 | Improved |
| use max pooling (PointNet-style) head instead of attention (ablation) | SparseSplat · DL3DV ablation 40k | PSNR: Fail | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- DepthSplat — DepthSplat: Connecting Gaussian Splatting and Depth (CVPR 2025) · better on 1 of 2 shared results
- MVSplat — MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images (ECCV 2024) · better on 2 of 2 shared results
- Long-LRM (200k) — Long-LRM: Long-sequence Large Reconstruction Model for Wide-coverage Gaussian Splats (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.94)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.94)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.93)
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.93)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.93)
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.
