Version 2 of my LEGO® Bonsai Tree (set 10281) splat, now complete: tree, pink blossoms,
the black pot, the pebbles inside it and the slatted wooden stand - with the room removed.
What's new since version 1:
- Stand and pebbles included: SAM 3 text prompts written from LEGO's own description of the
set ("pink cherry blossoms", "plant pot", "pebbles", "wooden stand"), with the cloth under
the stand excluded.
- No manual masking: frames where part of the object dropped out of SAM 3's mask were found
automatically (a sudden drop compared with neighbouring frames) and refilled by re-tracking
from the nearest good frames. A second pass asked SAM 3 for the pot, pebbles and stand in
every frame and added them wherever they were missing, including the inside of the pot.
- Sharper training: Nerfstudio's splatfacto-big (gsplat), 30,000 steps, tuned to keep adding
detail (about 3x more Gaussians than version 1).
- Clean cut-out: each Gaussian kept or removed by how much it actually paints inside vs
outside the masks across all 292 views (the FlashSplat method), then stray clusters removed,
centred and scaled for viewing.
Credits:
- Photos and camera positions: "bonsai" scene of the Mip-NeRF 360 dataset - Barron, Mildenhall,
Verbin, Srinivasan & Hedman, "Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields",
CVPR 2022 - used with the permission of the authors.
- Segmentation: SAM 3 (Meta AI).
- 3D cut-out method: Shen, Yang & Wang, "FlashSplat: 2D to 3D Gaussian Splatting Segmentation
Solved Optimally", ECCV 2024.
- Gaussian Splatting: Kerbl et al., SIGGRAPH 2023. Training: Nerfstudio + gsplat. Cameras
supplied with the dataset (computed with COLMAP).
Built step by step with help from Claude (Anthropic).
LEGO® is a trademark of the LEGO Group, which does not sponsor or endorse this model.