A question about your paper

[ArchMelow]

Hi, thanks for your great research.

This might be a dumb question, but I wonder why you are training a diffusion model to predict per-pixel gaussian parameters.
Why would it work better than other models such as convolutional neural networks without diffusion process?

I look forward for your reply.