Open yuhaoliu7456 opened 1 year ago
Hi @yuhaoliu7456 ! sorry for the delay in reply (almost a year, wow, I wasn't maintaining this repo properly) In the paper it is shown this way for simplicity. However, when working with real images, it is both faster and more accurate to work with the features from the forward noising DDIM process; it accumulates less error and the time-consuming backward sampling is not needed.
Great work. I see that the latent embedding is extracted from the denoising process in your paper, but you set the 'latents_forward' as the default way in this code. Can you explain this? Is there any difference between these two settings? And Do you think which one is better?