Splatfacto + VGGT: large scene reconstruction issues

[W-OK-E]

Hi there! I’ve been experimenting with splatting recently and tried using VGGT outputs to train Splatfacto.

• Splatfacto (vanilla): Works okay for small scenes with VGGT poses, but struggles with high-frequency details. On large scenes, the reconstruction often collapses or breaks from certain views. VGGT’s camera poses seem roughly accurate (focal length error ~10px for fx, but larger ~100px for cx/cy). Could the noisy pointcloud or pose inaccuracies be causing this?

• Splatfacto-W: Performs much better for small scenes, preserving high frequencies. However, large scenes still show holes/shimmering. Also, ns-export doesn’t seem to work with Splatfacto-W — is there any workaround or known solution for exporting splats/meshes?

Any insights would be greatly appreciated!