Output inconsistency with the same input

Hi @haotongl,

Thanks for your incredible work! The model's performance and architecture are truly impressive.

I am currently integrating your model as a preprocessing stage in my pipeline. However, I have encountered a minor technical consistency issue:
**when running inference on the exact same input image multiple times, I noticed that the estimated extrinsics and depth maps exhibit slight variations. The differences are approximately 1e-5.**

While the error is small, it impacts the stability of my downstream tasks. I have already ensured that the inputs are identical.

Could you please provide some guidance on how to achieve perfectly deterministic outputs? Thank you agian for your time and for sharing this great research!

No files, tests, or entry points are named. Start by reproducing inference multiple times with the identical image and inspect the inference configuration and execution path for sources of nondeterminism. Done means repeated runs produce stable extrinsics and depth maps, or the remaining variation is documented with guidance for users.

Written by the indexing model from the issue text.