Additional Loss term compared with paper
zcai0612 opened this issue · comments
CAI Zeyu commented
May I ask about the first term of the loss function? (sigma_t / (sigma_t_next + 1e-9)) * (pred_epsilon - noise_pred)
It seems that it always equals zero.
Zike Wu commented
It is borrowed from Common Diffusion Noise Schedules and Sample Steps Are Flawed. You can optionally activate it by setting "recon_std_rescale" to non-zero integers. However, we found it not suitable in our work, so we set this rescale to zero as default.