The model could not be fitted if not predict xstart

[JJLi0427]

Thank you so much for sharing this great work.
I try to use DiT to my dataset, but I meet some problem when I train model using predict_xstart = False. The loss down a little and seem to stuck at a hight level and won't be down, when I use predict_xstart = True the loss down quickly.

[santisy]

So can you confirm that when using predict_xtart=True, the result would be much better?

[JJLi0427]

So can you confirm that when using predict_xtart=True, the result would be much better?

According to my dataset experiment predict_xtart=True can converge quickly

[XijiaWang]

According to my dataset experiment predict_xtart=True can converge quickly

@JJLi0427
Hello, I encountered the same problem, have you found the reason?

[JJLi0427]

According to my dataset experiment predict_xtart=True can converge quickly

@JJLi0427 Hello, I encountered the same problem, have you found the reason?

This seems to be because there is a problem with the implementation of the denoising function, and I only used predict xstart to do my task later.