Thank you for sharing this codebase. I am currently experimenting with it, but I am facing reproducibility issues. When I run the same experiment twice — using the same scene, the same set of context views, and with data augmentation and voxelization explicitly disabled — the loss curves still do not match between runs.

I would like to confirm whether there are other sources of randomness in the training pipeline that should be controlled to ensure reproducible results.
I am using these flags in my command
dataset.dl3dv.augment=false
dataset.dl3dv.intr_augment=false
dataset.dl3dv.input_image_shape=[448,448]
model.encoder.voxelize=false