Possible errors about using q_sample(x, t)
ZGCTroy opened this issue · comments
Zheng Guang Cong commented
Jooyoung Choi commented
Yes, t should be modified to t-1. Thank you for letting me know the error.
While the experiments in the paper were done correctly with t-1, I made a mistake while moving to the guided-diffusion codebase.
In fact, the mistakenly added noise (variance \beta_t) is small enough to be eliminated by downsampling operation, thus the samples are fine even with t.
Zheng Guang Cong commented
Thanks for your reply. I agree with you that "the mistakenly added noise (variance \beta_t) is small enough to be eliminated by downsampling operation, thus the samples are fine even with t".