The loss does not decrease and the training result is an all-white point cloud

[HITZY2002]

trafficstars

The latest [gaussian-splatting] repository code I use, and then diff-gaussian-rasterization uses a 3dgs_accel branch, which doesn't report errors during training, but the loss value doesn't decrease, and the training results are weird, all like overexposed white. The server system used for training is ubuntu20.04, there are 8 4090s, the cuda version is 11.8, and there is no problem with using the old version of the previous configuration for training on your own windows computer, but there is a problem with gaussian-splatting and submodules pulling the latest version. Can someone tell me why, thank you very much