Hello author. Thank you for your work, it's proving to be very helpful to me. Considering my limited server resources, I wonder if it's necessary to train the entire diffusion model or just fine tune and train the CFW and time aware encoder to achieve good results on my own dataset.