Research map / 3D vision / Gaussian splatting
Splatter Image: Ultra-Fast Single-View 3D Reconstruction
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.
Splatter Image predicts one 3D Gaussian per pixel with a U-Net in a feed-forward pass, enabling single- and few-view 3D object reconstruction at 38 FPS with single-GPU training.
Problem
Monocular and few-view 3D object reconstruction is slow with implicit (NeRF-style) or voxel/triplane representations and needs heavy compute to train; the paper targets fast, cheap, high-quality feed-forward reconstruction.
Approach
Store a set of 3D Gaussians in an image-shaped tensor by predicting per-pixel Gaussian parameters (depth, offset, opacity, covariance, colour) with a standard image-to-image U-Net, so reconstruction reduces to 2D convolution and Gaussian Splatting renders fast enough to use image-level losses like LPIPS. Multi-view input is handled by warping per-view Gaussian sets into a common frame and adding c
Key results
- On ShapeNet-SRN Cars single-view, PSNR 24.00 / LPIPS 0.078 versus FE-NVS 22.83 / 0.099, with best LPIPS among the compared methods on Cars.
- On Google Scanned Objects, PSNR 21.06 and LPIPS 0.111 versus OpenLRM's 18.06 and 0.129, trained for 7 GPU days on Objaverse with 2 GPUs over 3.5 days.
- Test-time reconstruction plus 250-view rendering takes 0.451 s on a V100, more than 1000x faster than PixelNeRF (463.3 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Splatter Image instead of PixelNeRF | Single-view reconstruction · CO3D Hydrant and Teddybear | LPIPS: 0.203 -> 0.150; 0.290 -> 0.231 | Helped |
| use Splatter Image instead of OpenLRM | Single-view reconstruction · Google Scanned Objects | PSNR: 18.06 -> 21.06 | Helped |
| replace the Splatter Image with fully connected, unstructured output (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | PSNR: 22.25 -> 20.60 | Hurt |
| predict 3D coordinates directly instead of depth along the ray (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | PSNR: 22.25 -> 21.21 | Hurt |
| remove view-dependent colour (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | PSNR: 22.25 -> 21.77 | Hurt |
| constrain Gaussians to be isotropic (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | PSNR: 22.25 -> 22.01 | Hurt |
| remove 3D offset prediction (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | PSNR: 22.25 -> 22.06 | Hurt |
| remove LPIPS perceptual loss (ablation) | Splatter Image · ShapeNet-SRN single-view ablation | LPIPS: 0.115 -> 0.141 | Hurt |
Surprising findings
- A 2D image-to-image U-Net can output 360-degree reconstructions from a single view, with unseen parts encoded by allocating background pixels to 3D locations.
Limitations
- Focuses on object-centric reconstruction rather than generalising to unseen scenes.
- Deterministic method; does not model the ambiguity of single-view reconstruction via generative modelling.
Experimental setup
Datasets: ShapeNet-SRN CO3D ShapeNet (multi-category, NMR splits) Objaverse-LVIS Google Scanned Objects
Benchmarks: ShapeNet-SRN CO3D ShapeNet multi-category Google Scanned Objects
Metrics: PSNR SSIM LPIPS rendering and forward time
Baselines: PixelNeRF VisionNeRF FE-NVS ViewsetDiff NeRFDiff CodeNeRF SRN OpenLRM
Builds on: Gaussian Splatting SongUNet PixelNeRF Viewset Diffusion
Closest papers by meaning
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- MonoSplat: Generalizable 3D Gaussian Splatting from Monocular Depth Foundation Models (CVPR 2025 · similarity 0.91)
- Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- DepthSplat: Connecting Gaussian Splatting and Depth (CVPR 2025 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 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.