How to determine the max_sigma

[HyeonA Kim]

Hi Algolzw,
I wonder you how to determine the max_sigma and surrogate differentiable function (20).
Could you explain in detail if it's decided by something?

[balabala]

Hi, we just manually chose the max_sigma. And in our Refusion paper, experiments show that setting the max_sigma to be 30 or 50 would good to most tasks.

[HyeonA Kim]

Alright, Thank you so much!
I have one more question. How to determine the surrogate differentiable function (20)?

[balabala]

Hi, the surrogate differentiable function is a tricky choice. I just tried many times on my draft and found it is the best function for solving the SDE.

[HyeonA Kim]

Thanks for the details. The more I read the paper, the more questions I have. One more thing I'm wondering:

When inferring, I understand that the Markov chain iterates the desired output for a number of timesteps, but when training, I wonder if you are training to generate x(t) as x(0) after degrading x(0) to that timestep and adding noise, or x(t) as x(t-1)?

[balabala]

In training, we generate x(t) from x(0) using this function, which is based on the forward-SDE solution (Equation (6) in the paper).

[HyeonA Kim]

Thank you! But, I mean I wonder the backward process. Is it right that the score network train the x(t) to x(t-1)?

[balabala]

Thank you! But, I mean I wonder the backward process. Is it right that the score network train the x(t) to x(t-1)?

Yes if you use maximum likelihood loss.