Hi expert,
Nice work and get key point about inversion with independent random noises.
But I can't understand which case of error accumulation could be affect zt and xt-1 in algorithm 1.
May you give me some examples about this?
Hi @inbarhub
First of all thanks for a very good and interesting paper, really enjoyed reading.
I wonder if it's possibly to apply the derived noise maps to schedulers other than DDPM/DDIM? For example have you tried substituting the noise maps in Euler Ancestral sampler? Since ddim/ddpm in general seems to produce lower quality results/requires larger number of steps
you mention Asyrp which is probably a reference to https://github.com/kwonminki/Asyrp_official - why is Asyrp not mentioned in your paper and why is the flag not used? I would love to integrate Asyrp with your work, but I have a hard time making Asyrp work.
Hi authors,
I fail to understand why the randomly sampled epsilons are dependent in respect to the timestep axis. And likewise, I fail to understand why does epsilon_hat is independent. Thank you for the great work and eager for your reply!