`ref_unet` is called and loaded [here](https://github.com/zjp-shadow/CharacterGen/blob/02453103ecee16a8fd367c193e23bec1c163dfc7/2D_Stage/tuneavideo/pipelines/pipeline_tuneavideo.py#L499C1-L506C53) but the outputs `noise_pred_cond` is not used anywhere in the prediction What is the point of loading `ref_unet`, and is it trained?
