Hi, thank you for sharing the great work.
I have a question about t~p(t), r~p(r | t, Iters).
- Are these t and r timesteps between 0 and 1000, which is the same as the traditional DDPM timestep range?
- In the loss code, you sampled the timestep with normal distribution with P_mean and P_std. When I checked the values of t, it is normally estimated to be close to 1. If our maximum timestep T=1000, is it the correct value? Or did you use any other timestep range?
# t ~ p(t) and r ~ p(r|t, iters) (Mapping fn)
rnd_normal = torch.randn([images.shape[0], 1, 1, 1], device=images.device)
t = (rnd_normal * self.P_std + self.P_mean).exp()
r = self.t_to_r(t)
Hi, thank you for sharing the great work.
I have a question about t~p(t), r~p(r | t, Iters).