A densification preprocessing plugin for LichtFeld Studio. It performs a densification pass on sparse reconstructions to generate dense point clouds using RoMa v2 matching.
Based on bounty #2.
- RoMa v2 Matching
- Live Reconstruction Updates
- Configurable Parameters via GUI
- Densify Regions from Camera Subsets
- Scene Integration
- Installable via LFS Plugin Marketplace
In LichtFeld Studio:
- Open the Plugins Marketplace panel under "View" menu bar.
- Click Install on the
Lichtfeld-Densification-Plugincard.
git clone https://github.com/shadygm/lichtfeld-densification-plugin.git ~/.lichtfeld/plugins/lichtfeld-densification-plugin- Open the Dense Initialization panel (side panel).
- Ensure a scene is loaded in LichtFeld-Studio.
- Optional: Select a subset of cameras in the scene graph and keep ROI: Selected Cameras Only enabled to densify only that region of views.
- Optional: Configure settings (e.g., RoMa quality, reference fraction, filtering thresholds) using the scrub controls.
- Click Start Densification.
- Monitor progress.
- To reuse the result, right-click the point cloud in the scene graph and select Save to Disk.
precise: Highest quality, slow and VRAM heavy (H_lr=800, bidirectional).high: High quality, (H_lr=64, bidirectional matching)base: Balanced (H_lr=640).fast: Default, fast (H_lr=512).turbo: Fastest (H_lr=320).
- Certainty Thresh: Min overlap certainty (0.0-1.0).
- Reproj Thresh: Max reprojection error (px).
- Sampson Thresh: Max Sampson error (px²).
- Min Parallax Deg: Min parallax angle (deg).
- No Filter: Disable all geometric checks for raw output.
Adjust these in the GUI's Advanced Settings.
- Selected Cameras Only: Use only the currently selected cameras for densification.
- Reference Fraction: Fraction of active cameras used as reference views.
- Neighbors per Ref: Number of nearest neighbor views matched per reference.
| Scene | Method | PSNR (dB) | SSIM | Num Gaussians |
|---|---|---|---|---|
| garden | DENSE | 28.0025 | 0.867532 | 1,000,000 |
| garden | SPARSE | 27.8082 | 0.857569 | 1,000,000 |
| bicycle | DENSE | 25.0693 | 0.812810 | 1,000,000 |
| bicycle | SPARSE | 24.9199 | 0.786084 | 1,000,000 |
| stump | DENSE | 27.5625 | 0.859106 | 1,000,000 |
| stump | SPARSE | 26.6290 | 0.807915 | 1,000,000 |
| bonsai | DENSE | 32.7332 | 0.951858 | 1,000,000 |
| bonsai | SPARSE | 32.9545 | 0.951101 | 1,000,000 |
| counter | DENSE | 30.5424 | 0.929821 | 1,000,000 |
| counter | SPARSE | 30.3430 | 0.924391 | 1,000,000 |
| kitchen | DENSE | 32.2915 | 0.938054 | 1,000,000 |
| kitchen | SPARSE | 32.3918 | 0.936084 | 1,000,000 |
| room | DENSE | 33.8272 | 0.938879 | 1,000,000 |
| room | SPARSE | 33.6627 | 0.936690 | 1,000,000 |
| Mean | DENSE | 30.0041 | 0.899723 | 1,000,000 |
| Mean | SPARSE | 29.8156 | 0.885691 | 1,000,000 |
*Max PSNR reached across all scenes with/without densification*
-
Quality Gain:
DENSE achieves +0.63% average PSNR (30.0041 dB vs 29.8156 dB) and +1.58% SSIM (0.8997 vs 0.8857) compared to MASTER at 30k iterations. -
Training Speed:
Densification reaches the same PSNR quality 13% faster (at ~26k iterations) than non-densified training (requires 30k iterations).
This plugin's code is released under GPL-3.0-or-later.
RoMa and DINOv3 have their own licenses—review them separately for redistribution or commercial use.