Alpha loss integration leads to low PSNR and high mask error during training

[sanscosk]

Hi, I integrated an alpha loss into the original 2DGS code, but the PSNR value during training is significantly lower than expected. Additionally, the mask_error remains high and doesn't decrease as training progresses.
Here is the part of the code I added:
in readCamerasFromTransforms assign gt_alpha to alpha_channel

in train.py add alpha loss to loss

During the training process, the changes of mask_error are as follows:

PSNR:

The command I used for training is:
OMP_NUM_THREADS=4 CUDA_VISIBLE_DEVICES=0 python train.py -s data/nerf_synthetic/chair/ -m output/exp_nerf_synthetic/chair --eval --white_background --lambda_normal 0.0 --lambda_mask 0.1

[hbb1]

I am not entirely sure what happens. I have experimented regressing RGBA with L1 loss, and it worked well for me.