Overfit on lego

[StarsTesla]

Hi，there！ I tried the code on lego dataset with the recommend config in the readme and keep other config as default, and I found that the situation of training progress:

train psnr increasing to 20 fastly and then keep growing though,
val psnr around 10, and floating.

why is that happend?

Hardware:
GPU 3090

Software:
CUDA 11.6
pytorch 12.1
python 3.7.7
pytorch_lighting 1.7.7

I modified the structure of NeRF system with the new pytorch_lighting version(mostly the logger)
I have tired change the softplus for the density activation, increased the batch size, still got not change of the result.

[StarsTesla]

It seems I set the white_back = True, set it False make result become normal.

[YZsZY]

You know why need to set white_back=True? Thanks a lot!