I mean the dataset is extremely diverse. like millions clusters and every entry is noised when loaded on GPUs

But DO will prevent overfitting, I don't have any overfitting it's not the relevant tool

The split has been done such as the train and test are highly different. the loss are almost equal on both datasets.

The model is equivariant no dataset augmentation, no DO as well. The model doesn't overfit as I said

The loss here is for a denoiser, it can be seen as the variance between the noise and the noise predicted. So it's in this case a good metric