search

lucidrains / recurrent-interface-network-pytorch

Implementation of Recurrent Interface Network (RIN), for highly efficient generation of images and video without cascading networks, in Pytorch
MIT License
194 stars 14 forks source link

scale during sampling #19

Closed zhangdan8962 closed 9 months ago

zhangdan8962 commented 9 months ago

Hi,

Thank you so much for bringing the idea from Chen's paper into RIN.

In Algo 3 from Chen's paper, they did not mention 'scale' back the image during sampling. So I am wondering if we really need to do that in DDPM/DDIM sampler?

lucidrains commented 9 months ago

@zhangdan8962 oh! yes, i think you are correct, thank you

made the change

have you tried training RIN with Chen's scaling?

zhangdan8962 commented 9 months ago

@lucidrains No, I did not have a chance yet.

I was trying to confirm the scaling operation during inference with Chen, but it says his email address not exists. I may have to check the pix2seq code for the answer.

lucidrains commented 9 months ago

@zhangdan8962 sg, you should just run it and share your results here. that would benefit everyone