Splatfacto produces very low Gaussian/splat counts on large datasets (3k+ images) and large scenes (~2000+ sq ft)

[jawadaligohar]

Hi Nerfstudio team! First of all, brilliant work.

I’m consistently seeing very low Gaussian/splat counts during training with splatfacto, leading to under-coverage and poor reconstruction quality.

1. Large datasets (3k+ images)

With datasets larger than ~3,000 images, the resulting splat/Gaussian count is extremely low compared to smaller datasets, and the model appears underfit / missing detail.

2. Large scenes (2000+ sq ft)

For large physical areas (roughly >2,000 sq ft), the final splat count is too small to represent the full space. The model doesn’t have enough Gaussians to cover the scene, so the output quality is not acceptable.