Research map / 3D vision / Gaussian splatting
Generative Densification: Learning to Densify Gaussians for High-Fidelity Generalizable 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.
Proposes Generative Densification, a learned module that up-samples features of top-gradient Gaussians from feed-forward Gaussian models (LaRa, MVSplat) to generate fine Gaussians in a single forward pass.
Problem
Feed-forward generalizable Gaussian models struggle to reconstruct high-frequency details due to a limited number of Gaussians, and per-scene adaptive densification is too slow and overfits in sparse-view settings.
Approach
Instead of splitting and cloning raw Gaussian parameters, densify the penultimate feature representations of selected high-gradient Gaussians with a serialized-attention point transformer and up-sampling blocks, using learnable masking to filter which Gaussians need further densification, all in one forward pass.
Key results
- On Gobjaverse, object-level model with residual learning reaches 28.75 PSNR vs LaRa 27.49, and 31.23 vs 29.70 on GSO, with 134M params.
- Scene-level model reaches 27.08 PSNR on RE10k vs MVSplat 26.39 and DepthSplat 26.76, with 27.8M params reported in Table 2.
- Cross-dataset: on ACID 28.61 PSNR vs MVSplat 28.15; on DTU 14.05 vs 13.94.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add generative densification to LaRa | LaRa · Gobjaverse | PSNR: 27.49 -> 28.58 | Helped |
| Add generative densification to LaRa | LaRa · GSO | PSNR: 29.70 -> 31.06 | Helped |
| Add generative densification to LaRa | LaRa · Co3D | PSNR: 21.18 -> 21.72 | Helped |
| Add generative densification to MVSplat | MVSplat · RE10k | PSNR: 26.39 -> 27.08 | Helped |
| Add generative densification to MVSplat | MVSplat · ACID | PSNR: 28.15 -> 28.61 | Helped |
| Add generative densification to MVSplat | MVSplat · DTU | PSNR: 13.94 -> 14.05 | Helped |
| Add residual learning (ablation) | Ours object-level model · Gobjaverse | PSNR: 28.58 -> 28.75 | Helped |
| Increase LaRa from 2 to 3 Gaussians per voxel (ablation) | LaRa · GSO | PSNR: 29.70 -> 29.77 | No clear effect |
| Replace naive LaRa densification with generative densification at comparable Gaussian count (ablation) | LaRa · GSO | PSNR: 29.77 -> 30.95 | Helped |
| Increase MVSplat from 1 to 3 Gaussians per pixel (ablation) | MVSplat · RE10k | PSNR: 26.39 -> 26.54 | No clear effect |
| Use generative rather than naive densification (ablation) | MVSplat · RE10k | PSNR: 26.54 -> 27.08 | Helped |
| Add gradient masking to select Gaussians for densification (ablation) | Ours-fast · GSO | PSNR: 30.21 -> 30.73 | Helped |
Surprising findings
- Naive increase of Gaussians per voxel in LaRa (68% more Gaussians) gives only a slight PSNR gain, while the proposed densification gives a larger gain at comparable count.
- Replacing gradient masking with learnable masking causes a significant PSNR drop (30.63 to 29.31), indicating gradient masking is better for selecting Gaussians to densify.
Experimental setup
Datasets: Gobjaverse RE10k GSO Co3D ACID DTU
Benchmarks: Gobjaverse GSO Co3D RE10k ACID DTU
Metrics: PSNR SSIM LPIPS number of Gaussians inference time
Baselines: LaRa LGM GS-LRM MVSNeRF MuRF pixelNeRF GPNR pixelSplat MVSplat DepthSplat
Builds on: 3D Gaussian splatting LaRa MVSplat serialized attention (Point Transformer V3) adaptive density control
Closest papers by meaning
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.91)
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.90)
- Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting (ICML 2025 · similarity 0.89)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.89)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 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.