scale_reg of MCMC strategy when training with with_ut #823
Description
It seems that when training with 3dgut, the gradient cannot flow to splats' scales. If we add scale_reg loss to splats, those splats will gradually shrink to a tiny scale, is this a problem?