Forward Problem

[FeiiYin]

Hi, thanks for your re-implementation.
There exist one question why do you use the U-Net to estimate the x_0 in the forward process? Why not directly add the fixed or sampled gaussian noise onto the given input and only use the U-Net in the reverse process?

Thanks in advance!

[wenqsun]

Hi, thanks for your re-implementation. There exist one question why do you use the U-Net to estimate the x_0 in the forward process? Why not directly add the fixed or sampled gaussian noise onto the given input and only use the U-Net in the reverse process?

Thanks in advance!

I think that he uses the ODE to formulate the adding noise process, and for the reverse ODE process, we need to compute the predicted x_0, for the reverse process, you can refer to the codes in improved_DDPM: https://github.com/openai/improved-diffusion/blob/783b6740edb79fdb7d063250db2c51cc9545dcd1/improved_diffusion/gaussian_diffusion.py#L524C5-L561C1