Instance-Shadow-Diffusion icon indicating copy to clipboard operation
Instance-Shadow-Diffusion copied to clipboard
Published 7 months ago • verified publisher icon MKFMIKU

Pre training

Open dezle13 opened this issue 1 year ago • 1 comments

How to implement the pre training process? The loss in the code seems to only be the diffusion loss, but as described in the article, there needs to be a feature encoder's feature loss.

dezle13 avatar Sep 29 '24 03:09 dezle13

Hi @dezle13 , the feature encode is trained with the diffusion loss directly and it is optimized in the first stage and frozen in the second stage.

Screenshot 2024-09-30 at 10 31 35 AM

MKFMIKU avatar Sep 30 '24 14:09 MKFMIKU