Training Reproduce Suggestion

[ChunChenLin]

Hi, thanks for your great work!

I've successfully reproduced the training based on sd1.5. However, I'm trying to migrate the SuperEdit into Sana (https://github.com/NVlabs/Sana) but the results look really bad (seems the edited image cannot preserve input image).

Wondering if you have investigated or studied on image editing based on Sana. Appreciate if there's any suggestion :)

[liming-ai]

Hi @ChunChenLin,

Thanks for your question and I am glad that you already reproduced our training based SD1.5.

Unfortunately we did not conduct image editing experiments based on SANA model. I can offer you some experiences that training with larger models:

1. The training images is crucial, you cann not simply resize or interpolate the 512x512 images to 1024x1024, otherwise the training will crused
2. Instead of full-parameter fine-tuning, starting from LoRA fine-tuning is a better choice

Hope these suggestions would be helpful for you. Thanks again for your interest and question.