About the training process
Thanks for the release of this amazing model.
I have tried CosXL_edit in some editing tasks like change color, materials, etc. However, I found that CosXL_edit is not good at removing and adding items from the given images.
So I try to train CosXL_edit use a script [train_instruct_pix2pix.py] from stabilityai. I found that the noise scheduler of CosXL_edit is EDMEulerScheduler, so I guess the training process following the setting of "do_edm_style_training". So I reference the do_edm_style_training code in [train_dreambooth_lora_sdxl_advanced.py] to conduct edm_style training with my dataset.
However, this training will destory the editing capacity of CosXL even if I use the InstructP2P dataset. I guess I miss some key points for training CosXL_edit model.
Is there any plan for the training script? Or any advices for fineturning the CosXL model for new editing tasks?
Thanks.
I have put my attempt of finetuning CosXL on the [github]. It would be very grateful if you could point my mistake output or give any hints.
I'm trying to train with onetrainer, but noticed the diffusers model wasn't released? It's asking for the vae and I have no idea how to get it.