About Diffusion timestep annealing

[Kirito-Ausna]

Hi! Thanks for opening source DreamCraft3D. I am learning your paper and codes. When I tried to find the corresponding code with the part Diffusion timestep annealing, I find your implementation is actually randomly sampling timestep. I got confused. Have I missed something important? I have noted the time prior branch but the prior is set to None and I was not very clear about this part of codes:

if self.cfg.time_prior is not None:
            time_index = torch.where(
                (self.time_prior_acc_weights - current_step_ratio) > 0
            )[0][0]
            if time_index == 0 or torch.abs(
                self.time_prior_acc_weights[time_index] - current_step_ratio
            ) < torch.abs(
                self.time_prior_acc_weights[time_index - 1] - current_step_ratio
            ):
                t = self.num_train_timesteps - time_index
            else:
                t = self.num_train_timesteps - time_index + 1
            t = torch.clip(t, self.min_step, self.max_step + 1)
            t = torch.full((batch_size,), t, dtype=torch.long, device=self.device)

Is this the implementation of dreamtime? Could you share some insights with me? Many thanks!

[Kirito-Ausna]

Now I understand this part is the implementation of dreamtime, but could you share with us the default hyper-parameter setting?