search

lucidrains / imagen-pytorch

Implementation of Imagen, Google's Text-to-Image Neural Network, in Pytorch
MIT License
8.11k stars 768 forks source link

Training timestep thresholding #265

Closed jtawade closed 2 years ago

jtawade commented 2 years ago

Hi @lucidrains. I was curious why the randomly sampled timestep is capped at 0.999 during training:

https://github.com/lucidrains/imagen-pytorch/blob/0a3afb1a4577831c199661758fbac388769af0b9/imagen_pytorch/imagen_pytorch.py#L210

but during sampling, the reverse diffusion process starts are timestep 1:

https://github.com/lucidrains/imagen-pytorch/blob/0a3afb1a4577831c199661758fbac388769af0b9/imagen_pytorch/imagen_pytorch.py#L217

lucidrains commented 2 years ago

@jtawade Hi Jessica,

Yes indeed that does not make any sense, I've removed it :pray:

jtawade commented 2 years ago

Thank you!