Question about z_T and L_round

[giuseppecartella]

Hi @lenabolliger
I have another question regarding L_round.

In particular, I cannot completely understand the role of tT_loss (line 856-857 of sp_gaussian_diffusion.py). It seems to be the same term as in the DIffuSeq paper (equation 17 of the supp material)---> || \mu(z_T) || ^2.

My question is: z_T should approach an isotropic Gaussian process by definition after the entire noising process. So what is the purpose of adding this term to the loss computation?

Thank you
Best,
Giuseppe