Research map / 3D vision / Gaussian splatting
ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views
Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski, Stefan Roth
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 |
|---|---|---|---|
| add 3DGS-rasterized latent conditioning to the diffusion model (ablation) | ReconSplat · RE10K extrapolation (before fine-tuning) | image metrics | Improved |
| model joint appearance-depth distribution instead of appearance only (ablation) | ReconSplat · RE10K extrapolation (before fine-tuning) | image-wise metrics | Mixed |
| add annealed multi-resolution noise (ablation) | ReconSplat · RE10K extrapolation (before fine-tuning) | FID | Improved |
| add 2x upsampled-input fine-tuning (f=8 to f=4) (ablation) | ReconSplat · RE10K | image-quality metrics | Improved |
| train only 3D convolution blocks versus full fine-tuning (ablation) | ReconSplat · RE10K video (MEt3R) | multi-view consistency: 77M vs ~0.9B trainable parameters | No clear effect |
| use classifier-free guidance scale 3 instead of ~1 or >=5 (ablation) | ReconSplat · RE10K and DL3DV-10K | perceptual metrics | Improved |
| increase context views from 6 to 8 (ablation) | ReconSplat · DL3DV-10K | FID: 30.13 -> 27.40 | Improved |
| compare against LVSM on extrapolation | ReconSplat vs LVSM · RE10K extrapolation | LPIPS: 0.172 vs 0.222 | Hurt |
| compare against LVSM on extrapolation | ReconSplat vs LVSM · RE10K extrapolation | FID: 5.20 -> 4.89 | Improved |
| add generative refinement on regression 3DGS (vs DepthSplat) for extrapolation | ReconSplat vs DepthSplat · RE10K extrapolation | FID: 11.81 -> 4.89 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 5 of 5 on most shared results.
- MVSplat360 — MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024) · better on 7 of 7 shared results
- DepthSplat — DepthSplat: Connecting Gaussian Splatting and Depth (CVPR 2025) · better on 4 of 6 shared results
- pixelSplat — pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024) · better on 6 of 6 shared results
- MVSplat — MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images (ECCV 2024) · better on 5 of 6 shared results
- latentSplat — latentSplat: Autoencoding Variational Gaussians for Fast Generalizable 3D Reconstruction (ECCV 2024) · better on 5 of 6 shared results
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
- G4Splat: Geometry-Guided Gaussian Splatting with Generative Prior (ICLR 2026 · similarity 0.91)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.90)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.90)
- Decompositional Neural Scene Reconstruction with Generative Diffusion Prior (CVPR 2025 · similarity 0.90)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.89)
- HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction (ICLR 2025 · similarity 0.89)
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.
