-
|
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. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 2 replies
-
|
Do you have a reproducible code snippet? |
Beta Was this translation helpful? Give feedback.
-
|
Basically just using FreeU as proposed here https://huggingface.co/docs/diffusers/using-diffusers/freeu but switching torch dtype to bfloat16 |
Beta Was this translation helpful? Give feedback.
-
|
Could be because internally FreeU uses Fourier transforms and it might not have sound behaviour in BFloat16. |
Beta Was this translation helpful? Give feedback.
-
|
That is what I thought, that it is not really a model nor diffusers issue. Maybe a better handling of such error would be handy. Anyway thanks for giving out the missing info. |
Beta Was this translation helpful? Give feedback.
Could be because internally FreeU uses Fourier transforms and it might not have sound behaviour in BFloat16.