Use of DCT in loss function?

[jasonhu4]

I am wondering why in layers.py, the DCT is used like so: sq_error = dct(model_output - target) ** 2
In contrast to many other diffusion model papers where the loss is simply something like the MSE with a weighting factor? How does this impact the training process?