Open GYDDHPY opened 1 year ago
In my opinion, when you predict the x_start in t \approx T, with a cosine noise schedule, bar alphas (cumprod alphas) have very small values compared to linear noise schedules. So, in this case, the predicted x_start has tremendous large values.
In sampling functions, the value is clamped. Why the value range of clamp is set to (-1,1) ? Is it related to normalized image intensity value or added noise value or something else?
https://github.com/openai/improved-diffusion/blob/783b6740edb79fdb7d063250db2c51cc9545dcd1/improved_diffusion/gaussian_diffusion.py#L293-L298