Closed Luciennnnnnn closed 1 month ago
Thank you for your comment. As you mentioned, we require multiple model forward processes per each training step for making intermediate timestep samples . If the intermediate samples can be pre-calculated and saved in disk memory before training, we can definitely reduce the training time. However, since the intermediate sampling incorporate stochastic process, pre-calculated samples might be unable to contain all the diversity of distribution.
Hi, very interesting work! After careful reading through your paper, I guess this method requires simulation in training, which would results slow training and do not support fine-grained time partition (inference with more timestep). Does it right? If it is, how can we improve it?