# Initial sample
x = data * mask + sde.prior_sampling(data.shape).to(data.device) * (1. - mask)
Let's assume data is normalised to have approx std=1. In this case, we're initialising x as a tensor that has some parts with std=1 and some parts with std=prior_std, which is certainly out of distribution for the score network. Wouldn't it make more sense to initialise it similarly to the body of inpaint_update_fn?
I have tried the modification and visually I can't tell if one is significantly better than the other, but I imagine a more thorough benchmarking could reveal differences in FID.
My question concerns this line:
Let's assume
data
is normalised to have approxstd=1
. In this case, we're initialisingx
as a tensor that has some parts withstd=1
and some parts withstd=prior_std
, which is certainly out of distribution for the score network. Wouldn't it make more sense to initialise it similarly to the body ofinpaint_update_fn
?I have tried the modification and visually I can't tell if one is significantly better than the other, but I imagine a more thorough benchmarking could reveal differences in FID.
Original algortithm:
My modification: