TianyiLi-92
commented
1 year ago Hi Kouen, I agree with you. In their code, they are using t-1 (with variable t_last__t), so it should be a typo.
Open Koueen opened 1 year ago
TianyiLi-92
commented
1 year ago Hi Kouen, I agree with you. In their code, they are using t-1 (with variable t_last__t), so it should be a typo.
Yoy specifiy in Algorithm 1, line 5, that the alpha prod is based on t, which is the current timestep of the flow. Since you are joining x_unknown and x_known, should be in the same timestep doesn't it? Because when you apply noise to the original image based on a given timestep (Eq. 7), you obtain the original sampled with noise for t, and you need t-1, because you are then joining them as I mentioned, with x_unknown, which is the estimation of previous x given noise predicted. Then the equation should be alpha prod with t-1, in order to obtain the sample with the noise in that timestep. The same for 8a, t, denotes the current not the previous.
In addition, 8a, should be the variance or std? In all papers appears with sqrt.
Regards
Kouen