About single-view 3DGS reconstruction

Hi,
I've been testing **single-view** 3DGS reconstruction using DA3-GIANT-LARGE, but the results are consistently poor across various images. This happens both in the Hugging Face demo and my local setup.
Given that I'm using the largest checkpoint, is this level of performance expected for monocular inputs? Or are there specific configurations/hyperparameters I might be missing to improve the reconstruction quality?
Any insights would be appreciated. Thanks!

https://github.com/user-attachments/assets/80ec76e9-817f-4cff-aed9-cfc46165dd5a

Reproduce the reported single-view case with DA3-GIANT-LARGE in both the Hugging Face demo and the local setup, starting by comparing their inputs and configuration. Check the expected behavior for monocular 3DGS reconstruction and identify whether the result is expected or configuration-dependent. Done means documenting a clear explanation or a reproducible configuration in the issue.

Written by the indexing model from the issue text.