hmdolatabadi/denoising_diffusion

Clipping

ariel415el opened this issue · 0 comments

Hi Your implementation as all others I saw clips the model output:

_maybe_clip = lambda x_: (x_.clamp(min=-1, max=1) if clip_denoised else x_)

This is does not show in the paper but seem crucial for image quality..
Can you explain why it's there?

I posted the same question at the official repo:
hojonathanho/diffusion#5 (comment)