My childhood home, reconstructed as a Gaussian Splat scene.
I missed the days when I wore a tiny backpack as a toddler and walked home every afternoon along this narrow path under the trees. You can now follow the same path virtually and walk all the way up to the hill behind the village.
This became the most complex scene I have ever trained.
The environment contains:
* dense tree canopy
* narrow curved paths
* strong lighting contrast
* blue sky visible through foliage
* reflective surfaces
* houses tightly packed on both sides
The combination created all kinds of Gaussian Splatting challenges:
* floaters
* elongated spikes
* covariance explosions
* unstable reflections
* sky overgrowth
* canopy tearing
* equirectangular distortion issues from 360 footage
The scene was captured using an Insta360 X5 and trained in LichtFeld Studio.
After trying many different training configurations, I eventually settled on:
Strategy: MNRF + Mip Filter
Max Gaussians: 4M
Iterations: 30,000
SH Degree: 3
Mask Mode: Segment
Threshold: 0.55
Penalty Weight: 1.6
Penalty Power: 2
I also generated voxel collision using splat-transform for scene interaction. Unfortunately Supersplat currently does not support uploading the edited GLB collision version, so for now if you get stuck somewhere in the scene, press Space to jump out 🙂