about how to train the diffusion model?

[xiaoc57]

I think this is a good job. I'm a little confused about training the diffusion model. Is your training method using direct prediction x_0 to calculate the loss with the real camera or calculating epsilon and epsilon_pred to calculate the loss? I used the second one, but it was all garbage when predicting. Is there any skill?

[jasonyzhang]

We tried and got reasonable results on both. I think our performance and training stability was better when doing X_0 prediction. I would suggest trying that first.

When doing epsilon prediction, are you using our ddpm inference code? You may need to make sure that the flag for epsilon vs X0 prediction is set properly.

[Youngju-Na]

@jasonyzhang Hi! sorry for commenting here if it's not appropriate. I'd like to politely ask when are you planning to release the training code :)

[ThisIsIsaac]

Hi! Is there any update to the release of the training code?

[jasonyzhang]

10

[jasonyzhang]

The dev branch now has a V1 of the training code.