yuanzhi-zhu
commented
1 year ago Hey, you should notice that $\tilde{x}=x+\sigma z$ in Eq. 5, where $z$ is the standard gaussian distribution (hence it should be scaled by $\sigma$)
Open lihenryhfl opened 1 year ago
yuanzhi-zhu
commented
1 year ago Hey, you should notice that $\tilde{x}=x+\sigma z$ in Eq. 5, where $z$ is the standard gaussian distribution (hence it should be scaled by $\sigma$)
I may be wrong, but it appears to me that there is an error in the denoising score matching loss:
https://github.com/yang-song/score_sde_pytorch/blob/1618ddea340f3e4a2ed7852a0694a809775cf8d0/losses.py#L95
In particular,
std, the denominator forz, should be squared, shouldn't it? This is what is prescribed in Eq. 5 in the earlier work [1]. Does this seem to be the case?[1] Y. Song, S. Ermon, Generative Modeling by Estimating Gradients of the Data Distribution, 2019.