Some questions about the principle level of the article

[Roveer]

Hello and thank you for this excellent work !!!

I just started to get in touch with diffusion model, I would like to ask a few questions:
(1)The Loss used in the training process is the distance between ' x_ki-0 ' and ' x_0 '( || x_ki-0 - x_0 || ),
and ' x_ki-0 ' just like:

But the Loss during the real diffusion training process is || noise - ε_θ(x_t, t) ||^2 ,
It looks like your training process broke the Markov chain, isn't that a problem? (forgive my ignorance = _ =)

(2)If this method works (which is indeed the case from the effect point of view ^-^),
Can I understand like that : this training strategy is actually compressing the part of the inference process?

Grateful if you could answer my questions !!

[ZHAOZHIHAO]

(1). For 1, I think they didn't change anything in backward process. In the forward process, the predicted noise still follows Gaussian so all the equations still hold.
(2). Don't know what do you mean by 'compressing the part of the inference process'.