Questions regarding the timestep rescaling

[WangYixuan12]

Hi, thank you for the awesome work! As I read through the codebase, I found that the timestep was not rescaled during training and inference. Specifically, training code here takes in t ranging from 0 to 999, while inference code here takes in t ranging from 0 to 50. Will this be a problem for reconstructing the image, or this is just a minor detail not affecting the final performance a lot?

[jsu27]

Hi Yixuan, the reason why the default inference procedure takes 50 steps is because it uses DDIM (denoising diffusion implicit models) instead of DDPM. While the training procedure is the same, DDIMs can generate comparably high quality images in 10-100x fewer timesteps.