stray points in the trained point cloud.

[nisarg-ujjainkar]

I am training the model for lego scene from nerf-sythetic dataset. Here is ow the result looks like.

Render

GT

My setup:

Code version: most recent comment.
Cuda: 12.2
Pytorch: 2.5.1

I have docker setup with an older commit , which does not have this issue. How do I fix this?

[panxkun]

Same result in Blender datasets. Maybe it happened with lasted update. In more early version everything works well.

[panxkun]

#1193

maybe this pr can resolve this problem