a question

[superY688]

The forward diffusion process is still to gradually add noise at 1000 time steps, but the training and sampling process is only carried out at 10 time steps. Is this correct?

I would appreciate it if you could tell me.Thanks

[Sebastianjhx]

Hi there,

We set up the noise scheduler over 1000 timesteps, where the image becomes progressively noisier from timestep 0 to 1000, ending as pure noise. Under this circumstance, Fast-DDPM uses only 10 timesteps out of 1000 during both the training and sampling processes.