In train.py, what is the usage/meaning of 'prior_preservation'?

[HyeonHo99]

Thank you for the implementation.

There's one thing I don't clearly understand. What is the usage/effect of 'prior_preservation'?
Is it set to not None in training time and None in inference time?
Also, what is 'unet_2' doing in this code?

Thank you again.

[HyeonHo99]

This is not something included in the original paper right?
Can you explain about this modification, please?

Thank you.

[bryandlee]

Hi, prior_preservation was added as a regularization to prevent overfitting (originally from Dreambooth paper)

[HyeonHo99]

Thank you for clarifying it. I will check out the Dreambooth paper.
I have one more question. I think "train_temporal_conv" was not in the original Tune A Video paper.
Is this also one thing you added?
In default, it is False. Then what's the function of the conv_temporal (Conv1d) which is initialized as identity?
Do you recommend configuring 'train_temporal_conv' as True or False?

Thank you in advance.