Research map / 3D vision / 3D generation
MVBoost: Boost 3D Reconstruction with Multi-View Refinement
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.
Builds a pseudo-GT multi-view dataset from single images (Era3D multi-view diffusion + LGM reconstruction + refinement) to train a LoRA-adapted feed-forward 3DGS reconstruction model, plus a per-input-view LPIPS pose/alignment optimization.
Problem
Feed-forward single-image 3D reconstruction is limited by scarce high-quality 3D training data, and multi-view diffusion outputs used at inference are often inconsistent across views.
Approach
Combine the accuracy of a multi-view diffusion model with the cross-view consistency of a large 3D reconstruction model: reconstruct 3DGS from diffusion views, render views from it, refine them with diffusion at a partial noise strength, and train a LoRA-adapted LGM on these 2D-derived pseudo-GT views, followed by input-view LPIPS optimization.
Key results
- On GSO orbiting-view evaluation, MVBoost reaches PSNR 18.561, SSIM 0.859, LPIPS 0.131, outperforming the best baseline VFusion3D (17.416 PSNR, LPIPS 0.155).
- Geometry on GSO: Chamfer distance 0.10110 and F-Score 0.7977, versus 0.12132 / 0.7598 for TriplaneGaussian.
- Scaling with data: LPIPS improves from 0.223 at 10K samples to 0.131 at 100K samples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MVBoost instead of VFusion3D | Single-view 3D reconstruction · GSO orbiting views | PSNR: 17.416 -> 18.561 | Helped |
| Apply MVBoost instead of VFusion3D | Single-view 3D reconstruction · GSO orbiting views | SSIM: 0.846 -> 0.859 | Helped |
| Apply MVBoost instead of VFusion3D | Single-view 3D reconstruction · GSO orbiting views | LPIPS: 0.155 -> 0.131 | Helped |
| Apply MVBoost instead of TriplaneGaussian | Single-view 3D reconstruction · GSO | CD: 0.12132 -> 0.10110 | Helped |
| Apply MVBoost instead of InstantMesh | Single-view 3D reconstruction · GSO | F-Score: 0.7369 -> 0.7977 | Helped |
| Set refinement strength to 0.95 rather than using original views (ablation) | MVBoost multi-view refinement · GSO | PSNR: 17.811 -> 19.132 | Helped |
| Vary refinement strength from 0.10 to 1.00 (ablation) | MVBoost multi-view refinement · GSO | PSNR: 17.798; 17.76; 17.80; 18.27; 19.132; 18.583 | Mixed |
| Train with refinement strength 0.95 rather than original supervision (ablation) | MVBoost reconstruction model · GSO | PSNR: 17.851 -> 18.093 | Helped |
| Vary refinement strength used for supervision from 0.10 to 1.00 (ablation) | MVBoost reconstruction model · GSO | PSNR: 17.624; 17.764; 17.904; 18.021; 18.093; 18.053 | Mixed |
| Increase training data from 10K to 100K (ablation) | MVBoost reconstruction model · GSO | PSNR: 17.212 -> 18.561 | Helped |
| Apply input view optimization (ablation) | MVBoost · GSO input view | LPIPS: 0.108 -> 0.002 | Helped |
Surprising findings
- Refinement at strength 0.95 works best; full-strength (1.00) refinement performs worse in the reconstruction ablation (PSNR 18.053 vs 18.093).
Limitations
- Input view optimization is a per-object post-processing step rather than part of the feed-forward model.
- Main comparisons exclude input view optimization to be fair, so the full pipeline's gain is shown only qualitatively.
Experimental setup
Datasets: GSO Omni3D (supplementary) synthetic 100k object images generated via LayerDiffuse
Benchmarks: Google Scanned Objects (GSO)
Metrics: PSNR SSIM LPIPS CD F-Score
Baselines: OpenLRM VFusion3D CRM InstantMesh TriplaneGaussian LGM
Builds on: LGM Era3D LoRA Zero123 LayerDiffuse
Closest papers by meaning
- Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (CVPR 2025 · similarity 0.88)
- MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model (CVPR 2025 · similarity 0.88)
- Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion (CVPR 2025 · similarity 0.87)
- MVD-Fusion: Single-view 3D via Depth-consistent Multi-view Generation (CVPR 2024 · similarity 0.87)
- View Transformation Robustness for Multi-View 3D Object Reconstruction with Reconstruction Error-Guided View Selection (AAAI 2025 · similarity 0.87)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.86)
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.