sndnyang / iddpm Goto Github PK

Thank you very much for your work. I have two questions; please help me.
(i) Is there any plan to separate the sampling work? I found that you are now implementing sampling in each iteration during the training process.
(ii) How to use open ai's pre-trained models to sample in your code.
The trouble you again.

Hi,

First of all thanks for the code! I was wondering if you were able to make it converge on 256x256 images. Specifically I am using ImageNet (a smaller version of it with 9469 samples). I cannot quite make it generate plausible samples. I am using T=1000, cosine scheduler and LR=1e-4 with warm up. I tried both predicting epsilon and xstart but both give weird samples as I attach here. Do you have any tip/suggestion on what could be going wrong or how to improve sample quality? Maybe 1000 are not enough timesteps but then sampling would be much slower :S

Epsilon prediction	Xstart prediciton

Thanks!