|
I am curious is this a limitation of pytorch or a problem within diffusers that once you switch models to BFloat16 it is no longer possible to combine it with FreeU. |
Answered by
sayakpaul
Feb 6, 2024
Replies: 2 comments 2 replies
|
Do you have a reproducible code snippet? |
0 replies
|
Basically just using FreeU as proposed here https://huggingface.co/docs/diffusers/using-diffusers/freeu but switching torch dtype to bfloat16 |
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Could be because internally FreeU uses Fourier transforms and it might not have sound behaviour in BFloat16.