Thanks for your great work!
I am confused with the ism implementation:
|
prev_noisy_lat = self.scheduler.add_noise(latents, noise, self.timesteps[ind_prev_t]) |
This line appears to first disturb the rendering to t=1 before the ddim inversion. I wonder the motivation and explanation. If i miss something in the paper, please tell me.
I am looking forward to your reply.