w86763777 / pytorch-ddpm

Unofficial PyTorch implementation of Denoising Diffusion Probabilistic Models
Do What The F*ck You Want To Public License
506 stars 62 forks source link

Question about denoising sampling process #20

Open seung-hoon-lee opened 1 year ago

seung-hoon-lee commented 1 year ago

First of all thank you for a great work I have a question about your denoising sampling process image image

In your code, it looks like after predicting epsilon from noise prediction model, it compute X0, and computes the mean and variance of the diffusion posterior q(x{t-1} | x_t, x0), and use this diffusion posterior mean and variance to sample x{t-1}.

Can I know the reason why not directly sample x_{t-1} using the below equation? image

Thanks.