Bug

visibility_filter in gaussian_renderer/__init__.py uses (radii > 0).nonzero(), which returns a 2D index tensor of shape (N, 1) instead of a boolean mask.

This 2D index tensor causes incorrect advanced indexing behavior when used downstream in train.py for operations like:

gaussians.max_radii2D[visibility_filter] = torch.max(gaussians.max_radii2D[visibility_filter], radii[visibility_filter])

and in add_densification_stats:

self.xyz_gradient_accum[update_filter] += torch.norm(viewspace_point_tensor.grad[update_filter,:2], dim=-1, keepdim=True)

These callsites expect a 1D boolean mask for proper element selection, but the 2D index tensor from .nonzero() triggers PyTorch advanced indexing, which can produce unexpected shapes and incorrect results.

Fix

Removed the .nonzero() call so visibility_filter returns radii > 0 directly -- a proper 1D boolean mask that works correctly with all downstream indexing operations.